review-runs-tracking: 2026-08

[dormouse-bot]

Monthly tracking issue for below-threshold findings. Each run appends findings as a comment. Future runs read these to build cumulative evidence.

Do not close manually — a new issue is created each month, and prior months are closed automatically.

[dormouse-bot]

Run 30694320339 — 2026-08-01T09:47:57Z (review-runs)

Clean, healthy window — 1 grounded nightly dead-code PR, 0 failures, 0 unexpected comments, 0 new defects

• Evidence level: N/A (negative evidence + positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 30694320339
• Workflow: https://github.com/diffplug/dormouse/actions/runs/30694320339
• Window: 2026-07-31 09:45 → 2026-08-01 09:45 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): 6 tend-mention (3 skipped-at-trigger; 3 verify-success + handle-skipped no-ops), 1 tend-nightly (success), ~14 tend-notifications (hourly 0–1 min silent no-ops), 1 tend-review (success, silent no-op), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 20 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the 4 runs with artifacts ≈ $8.05 list-price (tend-nightly $4.15 / 37 turns; yesterday's review-runs $2.44; tend-review $0.86; tend-notifications $0.60).

• tend-nightly (30691992241, $4.15, 20 min): opened 1 well-grounded PR — #335 (drop the dead PREFERRED_THEME_ORDER list + allThemes.sort(...) from lib/scripts/bundle-themes.mjs, −12/−0). Independently re-verified the dead-code claim against source: the single pinned id GitHub.github-vscode-theme.github-dark-default is absent from EXTENSIONS (no github-vscode-theme namespace remains after cd248fac "Only ship OpenVSX themes"), and absent from bundled.json (0 matches). So both indexOf calls always return -1, the comparator always returns 0, and the sort is a genuine order-preserving no-op — the removal is behavior-neutral, exactly as the PR describes. bundled.json correctly left un-regenerated (byte-identical output). All 7 checks green (Build & Test, Standalone Smoketest, Visual Regression, Cloudflare Pages, Storybook Publish, UI Tests "210 tests unchanged", review). Pending maintainer review (opened 08:49 today — too early to score acceptance).

• Outcomes cross-check (Step 4): No bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window. chore(themes): drop dead PREFERRED_THEME_ORDER sort from bundler #335 is the only bot PR opened, at the very end of the window, so there is no fresh acceptance to score yet. Prior bot batch (fix(host): block metadata range under equivalent address encodings #328/docs: sync WebGL-renderer and clipboard-async specs with code #329/test(tutorial-shell): cover OSC 633 shell-integration reporting #330 and earlier) remains merged; track record stays uniformly positive.

• Bot output discipline — 0 unexpected public comments. The only bot output in-window was yesterday's tracking-issue append on the (now-closed) review-runs-tracking: 2026-07 #199. tend-review 30692528628 posted 0 reviews on chore(themes): drop dead PREFERRED_THEME_ORDER sort from bundler #335 — the intended silent self-review-of-bot-PR no-op (tend#607, settled). The 3 handle-skipped tend-mention legs are the benign tend-mention: no-op session on bot's own APPROVED review of human PRs (upstream filing request) #168 bot-own-event pattern (fully upstream-tracked). The ~14 tend-notifications runs were silent no-ops.

• Not in scope: issue #301 [security-audit] FAIL updated in-window — authored/commented entirely by app/github-actions (the separate security-audit workflow), not a tend artifact.

• Housekeeping: first run of the month — created this 2026-08 tracking issue and auto-closed the 2026-07 issue (review-runs-tracking: 2026-07 #199).

• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: one behavior-neutral, source-verified nightly dead-code PR with correct restraint (no bundled.json churn, no speculative regression test), disciplined no-spam behavior, no failures/timeouts. Recorded as continued negative evidence.

---

Run 30742367890 — 2026-08-02T09:48:17Z (review-runs)

Clean, healthy window — 1 grounded nightly docs-correction PR, prior nightly PR #335 merged, 0 failures, 0 unexpected comments, 0 new defects

• Evidence level: N/A (negative evidence + positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 30742367890
• Workflow: https://github.com/diffplug/dormouse/actions/runs/30742367890
• Window: 2026-08-01 09:46 → 2026-08-02 09:46 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): 6 tend-mention (3 skipped-at-trigger; 3 verify-success + handle-skipped no-ops), 1 tend-nightly (success), ~14 tend-notifications (hourly 0–1 min silent no-ops), 1 tend-review (success, silent no-op), 1 tend-triage (success, no-op). No long-running (>30 min) or near-timeout jobs — longest was nightly at 12 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the 4 runs with artifacts ≈ $6.69 list-price (tend-nightly $4.65 / 34 turns / 1.5M cache-read; tend-review $0.90; tend-notifications $0.60; tend-triage $0.54).

• tend-nightly (30740097155, $4.65, 12 min): opened 1 well-grounded PR — #337 (fix an inaccurate doc-comment in lib/src/components/wall/agent-browser-screen.ts: the absent-from-params RenderMode default is iframe, not ab-screencast, +5/−2). Independently re-verified the claim against source: resolveRenderMode in lib/src/components/wall/browser-surface.ts (L19–22) returns p.renderMode === 'ab-screencast' || p.renderMode === 'ab-popout' ? p.renderMode : 'iframe' — so a surface with no explicit mode resolves to iframe, exactly as the corrected comment states, and matches docs/specs/dor-browser.md → "Canonical Params". The PR also correctly preserves the distinct ScreenSnapshot.renderMode glyph default (absent ⇒ ab-screencast), calling out the difference rather than conflating the two. All 7 checks green (Build & Test, Standalone Smoketest, Visual Regression, Cloudflare Pages, Storybook Publish, UI Tests "210 tests unchanged", review). Pending maintainer review (opened 08:36 today — too early to score acceptance).

• Outcomes cross-check (Step 4): Prior nightly PR #335 (drop dead PREFERRED_THEME_ORDER sort) merged 2026-08-02 08:10 — positive acceptance of last window's output. No bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window.

• Bot output discipline — 0 unexpected public comments. tend-review 30740372328 posted 0 reviews on docs(agent-browser): fix RenderMode absent-default (iframe, not ab-screencast) #337 — the intended silent self-review-of-bot-PR no-op (tend#607, settled). The tend-mention 30740389435 leg was triggered by the cloudflare-workers-and-pages[bot] deploy comment on docs(agent-browser): fix RenderMode absent-default (iframe, not ab-screencast) #337 → handle-skipped no-op (benign non-directed-bot-comment pattern, tend-mention: no-op session on bot's own APPROVED review of human PRs (upstream filing request) #168-tracked). tend-triage 30694441395 fired on the creation of this tracking issue review-runs-tracking: 2026-08 #336 (an issues event on a bot-authored issue) and correctly no-op'd (1 min, no comment posted). The ~14 tend-notifications runs were silent no-ops.

• Minor recurring observation (below threshold, no action): tend-triage fires once per month on the review-runs tracking-issue creation ($0.54, ~1 min, silent no-op). Correctly handled by the self-loop / bot-authored-issue guard; cost is negligible and behavior is correct. Recorded only for pattern-tracking — not a defect.

• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: one source-verified nightly docs-correction PR, disciplined no-spam behavior, prior PR merged, no failures/timeouts. Recorded as continued negative evidence.

---

Run 30807659399 — 2026-08-03T11:03:18Z (review-runs)

Clean, healthy window — nightly shipped a grounded workflow-update PR + config-drift issue with strong operational analysis; prior PR #337 merged; 0 failures, 0 timeouts, 0 spam

• Evidence level: N/A (negative evidence + positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 30807659399
• Workflow: https://github.com/diffplug/dormouse/actions/runs/30807659399
• Window: 2026-08-02 11:00 → 2026-08-03 11:00 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): 1 tend-nightly, 1 tend-triage, 4 tend-review (each a distinct PR — no double-posting), ~9 tend-mention (several skipped-at-trigger + handle-skipped no-ops), ~14 tend-notifications (hourly silent no-ops), plus 2 skipped tend-ci-fix. No long-running (>30 min) or near-timeout jobs — longest was nightly at 8 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml). Token spend across the 6 runs with artifacts ≈ $9.75 list-price (tend-review 4× = $5.18; tend-nightly $3.62 / 32 turns; tend-triage $0.95).

• tend-nightly (30804092701, $3.62, 8 min): ran tend check, opened 1 grounded PR — #340 (regenerate all 8 tend workflows 0.1.12 → 0.1.13, incl. tend-mention relay rework + environment: tend gating) — and filed #339 documenting three non-bot-fixable tend check config-drift FAILs (missing tend environment, repo-level operational secrets). All checks green on chore: update tend workflows (0.1.12 → 0.1.17) #340 (Build & Test, Standalone Smoketest, Cloudflare Pages, review). Pending maintainer review.

• tend-review self-review of chore: update tend workflows (0.1.12 → 0.1.17) #340 (30804455994): posted 1 substantive, non-spam review note — correctly flagged that merging won't break CI (repo-level secrets stay readable + GitHub auto-provisions the referenced environment), that the environment: tend hardening stays inert until the secrets are moved and a deployment-branch rule is added, and — the non-obvious catch — that a "Required reviewers" rule would stall every scheduled bot run since the bot can't approve its own deployment. Grounded and genuinely useful; not a self-conversation no-op because there was real substance to surface.

• tend-triage on tend check: configuration drift on diffplug/dormouse #339 (30804282066): correctly treated the bot-authored issue as a report to act on (per the bot-authored-issue exception), and added a valuable nuance — adding the two tokens to secrets.allowed only silences the check while leaving them PR-readable, whereas the protected-environment route actually closes the exposure. Complementary to the chore: update tend workflows (0.1.12 → 0.1.17) #340 review (PR-merge angle vs. issue-remediation angle), not duplicative.

• tend-review of fix(remote): make iOS push actually deliver, and let an expired Pocket session re-auth #338 (30773746710, $1.98 / 23 turns): thorough read of nedtwigg's human PR fix/push-vapid-and-session-recovery → APPROVED (empty body, no concerns). The other two 10:11 review runs targeted Renovate PRs (ascii-splash, hono) — separate PRs, not re-reviews of chore: update tend workflows (0.1.12 → 0.1.17) #340.

• Outcomes cross-check (Step 4): prior nightly PR #337 (RenderMode absent-default docs fix) merged 2026-08-03 00:26 — positive acceptance of last window's output. No bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window.

• Bot output discipline — 0 unexpected public comments. tend-mention legs at 00:08 (Cloudflare deploy comment on fix(remote): make iOS push actually deliver, and let an expired Pocket session re-auth #338) and 00:13 (bot's own APPROVED review on fix(remote): make iOS push actually deliver, and let an expired Pocket session re-auth #338) were benign handle-skipped / self-loop no-ops. The ~14 tend-notifications runs were silent no-ops.

• Not a tend artifact: [workflow-audit] 1 change(s) on 2026-08-03 #343 ([workflow-audit]) and [security-audit] FAIL on 2026-07-27 #301 ([security-audit]) are authored/updated by the separate app/github-actions workflows, not tend.

• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: grounded nightly PR + config-drift issue with strong, non-obvious operational analysis; correct approve-and-note discipline; prior PR merged; no failures/timeouts. Recorded as continued negative evidence.

---

Run 30900093727 — 2026-08-04T (review-runs)

Clean, healthy window — nightly shipped 2 grounded PRs (1 real bug fix + test, 1 verified docs fix), 11 grounded reviews incl. a standout xterm fork-drift catch; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + positive quality signal)
• Occurrences this run: 0 new defects
• Run ID: 30900093727
• Workflow: https://github.com/diffplug/dormouse/actions/runs/30900093727
• Window: 2026-08-03 10:20 → 2026-08-04 09:22 UTC (prior review-runs covered through ~11:00 08-03).

Runs (all success or skipped — no failed / cancelled / timed-out runs): 1 tend-nightly, 11 tend-review (each a distinct PR — no double-posting), ~25 tend-mention (several skipped-at-trigger + handle-skipped no-ops), ~12 tend-notifications (hourly silent no-ops). No long-running (>30 min) or near-timeout jobs — longest was nightly at 18 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the 17 runs with artifacts in the report window ≈ $24.05 list-price (tend-review 11× = $11.67; tend-nightly $7.91 / 111 turns / 6M cache-read; tend-notifications 3× = $2.36; yesterday's review-runs $1.61; tend-mention $0.50).

• tend-nightly (30894852046, $7.91, 18 min): the elevated spend (~2× the prior three nightlies at $3.62–$4.65) is explained — this run opened two grounded PRs from the code-quality survey, not one:

  • #351 fix(alert): emit onStateChange when WATCHING turns off on command finish (+29/−2, alert-manager.ts + test). Independently re-verified against source: finishCommandExitWatch calls applyWatchingRule (which returns whether the monitor changed) but on the disposal-without-arming path returns wasArmed (false), so both call sites (commandFinish, onExit) skip the trailing notify — the WATCHING-off transition is dropped while commandStart is unconditional, a genuine on/off asymmetry. getState returns the correct post-finish value, so push-based onStateChange consumers render stale watchingEnabled: true until the next commandStart. Fix captures the change signal on the non-armed path; armed paths still return true (ringing/arming unchanged). Regression test subscribes after command start and asserts the false transition arrives on finish — fails against old code, passes with fix. All 7 checks green.
  • #352 docs(agent-browser): repoint stale Source-of-truth pointers (+4/−4, dor-browser.md). Three Source of truth: parentheticals credited AgentBrowserPanel.tsx for logic that moved into agent-browser-surface-controller.ts when the panel became a thin view. Grep-confirmed each moved symbol (rememberRestorableUrl, screenActions/popOut/popIn, CDP observer/maybeAutoRevert) absent from the panel and present in the controller; pnpm lint:specs passes. Correctly left the still-accurate panel references (renderer view) untouched.

  Both PRs pending maintainer review (opened ~09:06 today — too early to score acceptance). Atomic split (bug fix vs docs) is correct per the atomic-PR rule.

• tend-review — 11 runs, 11 distinct PRs, all grounded, no double-posting. Nine Renovate dep bumps auto-APPROVED (github-actions chore(deps): update github-actions #344, react chore(deps): update react #345, storybook chore(deps): update storybook to v10.5.8 #346, vite chore(deps): update vite #347, tailwind-variants chore(deps): update dependency tailwind-variants to v3.3.1 #349, pnpm chore(deps): update pnpm to v11.21.0 #350, plus the two bot nightly PRs self-reviewed silently). Two carried substance:

  • chore(deps): update xterm #348 (xterm) — standout review (30870949215, $1.66 / 19 turns): posted a COMMENTED (not blocking) review that confirmed the bump is clean and flagged a non-obvious project-health issue — the beta.291 → beta.292 range is a single upstream fix ("Fix stale rendering after a shared texture atlas is cleared", xterm.js#6055) landing in addons/addon-webgl/src/TextureAtlas.ts, the exact file the SDF fork extends. Backed with gitHead SHAs, the compare-API file list, and the canopy/ pin (sdf-v0.20.0-sdf291.0, pre-fix), correctly citing webgl-text.md → "Following upstream". Honestly named the out-of-repo (diffplug/xterm.js) rebase as work it can't do from CI and left the merge decision to the maintainer. Exemplary grounded review — exactly the source-evidence bar the skill asks for.
  • chore(deps): update github-actions #344 double-check: two review runs targeted chore(deps): update github-actions #344 (renovate synchronize), but only one APPROVED review was posted — the re-review correctly no-op'd. No spam.

• Outcomes cross-check (Step 4): prior nightly PRs #335 and #337 remain merged; no bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window. Last window's #340 (workflow regen) and #339 (config-drift) are still open — maintainer's call, not a defect; nightly's tend check correctly re-refreshed tend check: configuration drift on diffplug/dormouse #339 rather than filing a duplicate.

• Bot output discipline — 0 unexpected public comments. No bot issue-comments posted in-window beyond the intended tend-review outputs above. tend-mention legs were the benign skipped-at-trigger / handle-skipped no-ops (Cloudflare deploy comments + bot's own review events, tend-mention: no-op session on bot's own APPROVED review of human PRs (upstream filing request) #168-tracked). tend-review self-reviews of the two bot nightly PRs (fix(alert): emit onStateChange when WATCHING turns off on command finish #351, docs(agent-browser): repoint stale Source-of-truth pointers to the surface controller #352) were the intended silent no-ops (tend#607, settled). The ~12 tend-notifications runs were silent no-ops.

• Not a tend artifact: [workflow-audit] 1 change(s) on 2026-08-04 #353 / [workflow-audit] 1 change(s) on 2026-08-03 #343 ([workflow-audit]) are authored by the separate app/github-actions workflow, not tend.

• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: two source-verified nightly PRs (a real push-based bug fix with a proper regression test, plus a grounded docs-pointer correction), eleven grounded distinct-PR reviews including a genuinely excellent xterm fork-drift catch, correct atomic-PR and self-review discipline, no failures/timeouts/spam. Recorded as continued negative evidence.

---

Run 30996725754 — 2026-08-05 (review-runs)

Clean, healthy window — nightly ran a thorough grounded 7-file quality sweep (0 bugs, no PR), reviews correct, 3 prior bot PRs still pending maintainer; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence)
• Occurrences this run: 0 new defects
• Run ID: 30996725754
• Workflow: https://github.com/diffplug/dormouse/actions/runs/30996725754
• Window: 2026-08-04 10:18 → 2026-08-05 10:18 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): 1 tend-nightly (success), 3 tend-review (all no-op/correct), ~11 tend-mention (skipped-at-trigger + handle-skipped no-ops), ~13 tend-notifications (hourly silent no-ops), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 7 min (09:04→09:11), far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the 7 runs with artifacts ≈ $10.85 list-price (tend-nightly $3.96 / 31 turns / 1.4M cache-read; tend-review 3× = $3.79; tend-notifications 2× = $1.70; yesterday's review-runs $1.40).

• tend-nightly (30991711666, $3.96, 31 turns, 7 min): ran a code-quality survey over 7 files (remote-host.ts, reconnect.ts, pairing-approval.ts, lath/persistence.ts, agent-browser-tab.ts, SelectionOverlay.tsx, LathHost.tsx) plus their specs (tiling-engine, transport, mouse-and-clipboard, dor-browser, remote-security-model, remote-api) and opened no PR — nothing actionable found. The sweep was genuinely grounded, not a shallow no-op: it chased and cleared real-looking risks with source evidence — e.g. confirmed void this.#onConnect2(...) (remote-host.ts:257) is safe because every verifier in authorizeConnection is internally guarded and never throws on hostile input (deviceKey.ts:104-152 catch; passkey.ts returns {ok:false}; key-hash .catch(()=>null)); confirmed the door-vs-pane field asymmetry in reconnect.ts (component vs surfaceType) is correct per session-types.ts; noted reconnect.test.ts/remote-host.test.ts/persistence.test.ts/LathHost.test.tsx cover the non-trivial modules so no coverage-gap findings. Correctly declined to manufacture a PR from a clean survey. This explains the $3.96 spend with no artifact — real review work, not a stuck run.

• tend-review — 3 runs, all correct, 0 new reviews posted this window:

  • Two runs on chore(deps): update github-actions #344 (renovate/github-actions, pull_request_target): 30967563901 (01:50, $1.03) and 30979512156 (05:53, $1.43). chore(deps): update github-actions #344 carries a single APPROVED review from 2026-08-03 16:39 — both re-review runs correctly posted nothing (already approved, no new commit). No double-posting.
  • chore(deps): update storybook to v10.5.8 #346 (renovate/storybook v10.5.6): 30940608117 (18:51, $1.33) re-approved after a Renovate rebase — the head commit e63c4884 was committed 18:51:49 and the synchronize event fired the run at 18:51:53; the prior 08-03 20:40 approval was on the pre-rebase commit. The fresh 18:55 APPROVED is the correct re-approval of a new commit, not a duplicate.

• Outcomes cross-check (Step 4): three bot PRs remain open pending maintainer — #351 (alert WATCHING-off onStateChange fix + test), #352 (agent-browser docs-pointer fix), #340 (workflow regen 0.1.12→0.1.13). No bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window. #339 (config-drift) stays open (maintainer's call); nightly correctly did not re-file a duplicate.

• Bot output discipline — 0 unexpected public comments. Both tend-notifications runs inspected (30972699943, 30905597151) were correct silent no-ops: every trigger was an automated cloudflare-workers-and-pages deploy-status comment on a Renovate/bot PR (bot-to-bot noise, nothing directed at the bot). The slightly-above-baseline cost ($0.79/$0.91 vs the usual ~$0.60) is from reading PR context to verify the trigger before deciding — proper diligence, not waste. tend-review self-reviews of the two bot nightly PRs stayed silent (tend#607, settled). tend-mention legs were the benign skipped-at-trigger / handle-skipped no-ops (tend-mention: no-op session on bot's own APPROVED review of human PRs (upstream filing request) #168-tracked).

• Not a tend artifact: [workflow-audit] 1 change(s) on 2026-08-05 #354 ([workflow-audit]) and [security-audit] FAIL on 2026-07-27 #301 ([security-audit]) are authored/updated by the separate app/github-actions workflows, not tend.

• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: a genuinely grounded nightly quality sweep that correctly produced no PR, correct re-review discipline (re-approve-on-rebase, no-op-on-already-approved), disciplined no-spam notifications, three prior grounded PRs still pending maintainer, no failures/timeouts/spam. Recorded as continued negative evidence.

---

Run 31092657711 — 2026-08-06T10:24:07Z (review-runs)

Clean, healthy window — nightly filed 1 grounded design-question issue + refreshed the workflow-regen PR; reviews/triage/mention discipline all correct; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + positive quality signal)
• Occurrences this run: 0 new defects
• Run ID: 31092657711
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31092657711
• Window: 2026-08-05 10:19 → 2026-08-06 10:19 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): 1 tend-nightly, 1 tend-triage, 3 tend-review, ~9 tend-mention (several skipped-at-trigger + handle-skipped no-ops), ~14 tend-notifications (hourly silent no-ops), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 11 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the 7 runs with artifacts ≈ $10.48 list-price (tend-nightly $4.06 / 32 turns / 1.9M cache-read; tend-review 3× = $3.89; tend-notifications 2× = $1.58; tend-triage $0.95).

• tend-nightly (31087623076, $4.06, 32 turns, 11 min): ran the code-quality survey and produced two grounded outputs, no stuck run:
  • Pushed a fresh commit to the existing workflow-regen PR #340 bumping the tend workflows 0.1.13 → 0.1.14 (commit at 09:10:08) rather than opening a duplicate PR — correct dedup against the still-open chore: update tend workflows (0.1.12 → 0.1.17) #340.
  • Filed #355 — a well-grounded design-question issue: dor send --stdin runs piped bytes through interpretTextEscapes (send.ts:227-228), so a shell script piped via the flagship cat script.sh | dor send --stdin example gets its literal \t/\n/\\ rewritten. Every claim is source-anchored with SHA-pinned send.ts links, includes a concrete repro, correctly notes the behavior is consistent with the documented text-input contract, lays out two options, and explicitly leaves the contract-change call to a maintainer (offering to open the PR + regression test once a direction is chosen). Exactly the "issue not drive-by PR" restraint the design-question path calls for.
• tend-review — 3 runs, all correct:
  • 31064433204 (02:01, $1.35, pull_request_target on chore(deps): update github-actions #344 renovate/github-actions): a Renovate force-push/rebase landed at 02:01:46; the run correctly posted a fresh APPROVED at 02:04:54. Legit re-approval of a rebased head, not a duplicate.
  • 31088005941 (09:10, pull_request_target on chore(deps): update github-actions #344) and 31087959613 (09:10, pull_request_target on chore: update tend workflows (0.1.12 → 0.1.17) #340): both correctly posted nothing — chore(deps): update github-actions #344's 09:10 force-push was a no-content rebase, and chore: update tend workflows (0.1.12 → 0.1.17) #340 is the bot's own PR (silent self-review-of-bot-PR no-op, tend#607, settled). No double-posting.
• tend-triage (31088427169, $0.95, 10 turns): fired on the issues creation event for bot-authored dor send --stdin interprets backslash escapes in piped bytes #355 and correctly posted nothing. dor send --stdin interprets backslash escapes in piped bytes #355 already carries a complete grounded analysis and a clear maintainer ask; a triage comment would be redundant self-conversation. Correct restraint under the bot-authored-issue exception.
• Outcomes cross-check (Step 4): three bot PRs remain OPEN pending maintainer — #340 (workflow regen 0.1.14), #351 (alert WATCHING-off onStateChange fix + test, opened 08-04), #352 (agent-browser docs-pointer fix, opened 08-04). No bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window. No new merges in-window (last merge docs(agent-browser): fix RenderMode absent-default (iframe, not ab-screencast) #337 on 08-03); PR ages are within normal maintainer-review latency — nedtwigg lands these at their own pace, not a defect.
• Bot output discipline — 0 unexpected public comments. A full window scan of active threads (remote: no account-wide device list — pairing, revocation, and push targets are all per-Host #331/tend check: configuration drift on diffplug/dormouse #339/chore: update tend workflows (0.1.12 → 0.1.17) #340/chore(deps): update github-actions #344–docs(agent-browser): repoint stale Source-of-truth pointers to the surface controller #352/dor send --stdin interprets backslash escapes in piped bytes #355) found no bot issue-comments and only the one intended chore(deps): update github-actions #344 re-approval review. tend-mention legs (issue_comment on tend check: configuration drift on diffplug/dormouse #339/chore: update tend workflows (0.1.12 → 0.1.17) #340/chore(deps): update github-actions #344 events + a pull_request_review on the bot's own chore(deps): update github-actions #344 approval) were the benign self-loop / handle-skipped no-ops (tend-mention: no-op session on bot's own APPROVED review of human PRs (upstream filing request) #168-tracked). tend-notifications runs were silent no-ops.
• Not a tend artifact: [workflow-audit] 2 change(s) on 2026-08-06 #356/[workflow-audit] 1 change(s) on 2026-08-05 #354/[workflow-audit] 1 change(s) on 2026-08-04 #353 ([workflow-audit]) and [security-audit] FAIL on 2026-07-27 #301 ([security-audit]) are authored/updated by the separate app/github-actions workflows, not tend.
• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: a source-verified nightly design-question issue with correct issue-not-PR restraint, correct workflow-PR dedup (refresh not duplicate), correct review discipline (re-approve-on-rebase, no-op-on-no-change and on-own-PR), correct triage restraint on a bot-authored issue, disciplined no-spam behavior, no failures/timeouts. Recorded as continued negative evidence.

Run 31162600138 — 2026-08-07T08:45:01Z (review-runs)

Clean, healthy window — 1 grounded nightly issue + 1 grounded nightly PR, exemplary design-gated triage no-op, 0 real failures, 0 gate-passing findings

• Evidence level: N/A (negative evidence + positive acceptance signals)
• Occurrences this run: 0 new defects
• Run ID: 31162600138
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31162600138
• Window: 2026-08-06 08:40 → 2026-08-07 08:40 UTC.

Runs. No long-running (>30 min) or near-timeout jobs — longest was tend-nightly at 11 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). The six failure runs (tend-mention 31122431347/31122427290, tend-review 31122392897/31122376687, tend-notifications 31125648747/31120900418) are all cancel-in-progress concurrency cancellations — verify/review/notifications jobs cancelled before any Claude session started, superseded by newer renovate-PR events (#349/#350). Not behavioral failures. All in-window tend-mention runs were verify=success + handle=skipped gate-only no-ops (no Claude session, ~$0). Token spend across the 10 runs with artifacts ≈ $17.16 list-price (tend-nightly $8.20 / 2 runs; tend-review $4.46 / 4 runs; yesterday's review-runs $1.99; tend-notifications $1.56 / 2 runs; tend-triage $0.95).

Cross-checked chain (all grounded, all correct):

• tend-nightly (31087623076, $4.06, 11 min, 32 turns): surveyed ~18 files, ranked 4 findings, and filed only the top one as #355 (dor send --stdin runs backslash-escape interpretation on piped bytes) — good restraint, the 3 low findings (dead default branch, latent focus-trap, door-nav asymmetry) were noted but not spammed as issues.
• tend-triage (31088427169, $0.95, 10 turns) on bot-authored dor send --stdin interprets backslash escapes in piped bytes #355: exemplary. Independently verified the bug against dor/src/commands/send.ts:185/228 and confirmed the escape-on-stdin behavior is not a pinned spec invariant in docs/specs/dor-cli.md. Correctly declined a PR (fix is design-gated — maintainer must pick raw-by-default vs. document-the-footgun; even a repro-test PR would prejudge the direction) and correctly declined a self-comment (issue body already complete; a "confirmed, awaiting direction" comment would be redundant self-conversation). Exited silently, as it should.
• tend-nightly (31158596178, $4.14): filed #357 removing a genuinely-unused type WallActions import and an unused server destructured binding (−2/+2, all checks green).
• tend-review / tend-notifications: clean no-ops (self-review of bot PR chore: remove unused test bindings #357 = intended silent no-op per running-tend; hourly notifications gate-only).

No action taken — nothing passed Gate 1. The window shows tend behaving well: grounded issue/PR generation, restraint on low-value findings, and a textbook design-gated triage decision that avoided both a premature PR and comment noise.

---

Run 31248223641 — 2026-08-08T (review-runs)

Clean, healthy window — 2 grounded atomic nightly PRs + correct issue-body refresh, disciplined review/mention/notification no-ops; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 31248223641
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31248223641
• Window: 2026-08-07 08:22 → 2026-08-08 08:22 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out tend runs): 12 tend-mention (2 skipped-at-trigger; rest verify-success + handle-skipped bot-own-event no-ops, #168 pattern), 1 tend-nightly (success), ~24 tend-notifications (hourly silent no-ops), 4 tend-review (success), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 13 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the runs with artifacts ≈ $12.01 list-price (tend-nightly $5.20 / 44 turns; 4× tend-review $4.36; yesterday's review-runs $1.81; sampled tend-notifications $0.64).

• tend-nightly (31245729189, $5.20, 13 min, 44 turns): opened 2 well-grounded atomic PRs and refreshed one issue body —

  • #359 (chore(dor): drop unused ppid from ambient process declaration, −1). Independently re-verified: grep -rn ppid dor/src/ returns only the declaration line itself (node-runtime.d.ts:27) — zero call sites — so the removal from the minimal ambient process type is behavior-neutral. All checks green.
  • #360 (test(pocket): stage push-disabled case in NeedsHomeScreenInstall story, +1/−1). Adds pushConfigStatus: 'disabled' to the NeedsHomeScreenInstall story args; coherent with the story's own comment (the home-screen notice must not disagree with the push row). All automated checks green; UI Tests (Chromatic) PENDING = human-gated, expected.
  • #339 config-drift issue — nightly updated the issue body to match tonight's tend check rather than posting a new comment (correct no-spam restraint; no new comment added).

• tend-review — 4 runs, all correct discipline: two were silent self-reviews of the bot's own new PRs chore(dor): drop unused ppid from ambient process declaration #359/test(pocket): stage push-disabled case in NeedsHomeScreenInstall story #360 (intended no-op, settled tend#607); two were re-reviews of Renovate PRs #344 (github-actions) and #346 (storybook v10.5.7), each re-approving a genuinely-new Renovate head SHA (correct re-approve-on-rebase). Both Renovate PRs green.

• Outcomes cross-check (Step 4): No bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window. chore(dor): drop unused ppid from ambient process declaration #359/test(pocket): stage push-disabled case in NeedsHomeScreenInstall story #360 opened at the very end of the window (07:22–07:23) — too early to score acceptance. Prior merged batch (chore(themes): drop dead PREFERRED_THEME_ORDER sort from bundler #335/docs(agent-browser): fix RenderMode absent-default (iframe, not ab-screencast) #337 and earlier) stays merged; track record remains uniformly positive. Open bot PRs chore: update tend workflows (0.1.12 → 0.1.17) #340/fix(alert): emit onStateChange when WATCHING turns off on command finish #351/docs(agent-browser): repoint stale Source-of-truth pointers to the surface controller #352/chore: remove unused test bindings #357 remain pending maintainer (not regressions).

• Bot output discipline — 0 unexpected public comments. In-window bot writes were: 2 nightly PRs, the tend check: configuration drift on diffplug/dormouse #339 body refresh, and yesterday's tracking-issue append. No stray comments on notifications/mentions.

• Not in scope: security-audit run 31241052312 failed in-window — this is the separate security-audit workflow (app/github-actions), not a tend-* artifact, and matches the recurring [security-audit] FAIL pattern (issues [security-audit] FAIL on 2026-07-13 #249/[security-audit] FAIL on 2026-07-20 #272). Out of review-runs scope.

• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: two source-verified atomic nightly PRs, correct issue-body-refresh-not-new-comment restraint, disciplined review/mention/notification no-ops, no failures/timeouts/spam. Recorded as continued negative evidence. This is the 8th consecutive clean window.

---

Run 31303414445 — 2026-08-09T08:27:23Z (review-runs)

Clean, healthy window — 2 grounded atomic nightly PRs (1 fix+test, 1 docs), both source-verified; correct #339 body-refresh restraint; disciplined review/mention/notification no-ops; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence)
• Occurrences this run: 0 new defects
• Run ID: 31303414445
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31303414445
• Window: 2026-08-08 08:24 → 2026-08-09 08:24 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out tend runs): 2 tend-mention (verify-success + handle-skipped bot-own-event no-ops, #168 pattern) + 2 skipped-at-trigger, 1 tend-nightly (success), ~24 tend-notifications (hourly silent no-ops), 2 tend-review (silent self-reviews of the two new bot PRs), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 26 min, under the 30-min flag threshold and far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the runs with artifacts ≈ $8.14 list-price (tend-nightly $6.00 / 58 turns / 4.6M cache-read; 2× tend-review $1.55; sampled tend-notifications $0.59).

• tend-nightly (31300736176, $6.00, 58 turns, 26 min): ran the code-quality survey and opened 2 well-grounded atomic PRs plus refreshed one issue body —

  • #362 (fix(fake-scenarios): preserve endsWithPrompt through flattenScenario, +16/−0, fake-scenarios.ts + test). Independently source-verified against main: flattenScenario (fake-scenarios.ts:56-60) copies name/chunks/exitCode but drops the sibling semantic field endsWithPrompt, exactly as claimed; and SCENARIO_FAST_OUTPUT (fake-scenarios.ts:143-164) ends with instant(PROMPT, 200) (line 162) yet — unlike its four prompt-ending siblings at lines 69/96/123/139 — lacks endsWithPrompt: true. The PR honestly frames the flatten gap as latent today (only Storybook consumes flattenScenario, and it never reads the flag) rather than overselling it, and adds two targeted regression cases. All 7 checks green.
  • #363 (docs(lath): correct replaceAtPath off-tree-path comment, +2/−1, lath/model.ts). Docs-only: the comment claimed an off-tree path returns replacement, but the code returns the original subtree reached so far and never inserts replacement. Correctly framed as behavior-neutral, no test added. All 7 checks green.
  • #339 config-drift issue — nightly updated the issue body (updatedAt 07:18, in-window) to match tonight's tend check rather than posting a new comment (last comment stays 2026-08-06). Correct no-spam restraint, consistent with prior windows.

  Atomic split (fix+test vs docs) is correct per the atomic-PR rule. Both PRs opened at the very end of the window (07:36–07:37) — too early to score acceptance.

• tend-review — 2 runs, both correct: silent self-reviews of the bot's own new PRs fix(fake-scenarios): preserve endsWithPrompt through flattenScenario #362/docs(lath): correct replaceAtPath off-tree-path comment #363 (only Cloudflare deploy comments present; no bot review posted) — the intended silent self-review-of-bot-PR no-op (tend#607, settled; running-tend "Settled upstream rulings").

• Outcomes cross-check (Step 4): No bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window (0 bot PRs closed-unmerged in-window). Prior merged batch (chore(themes): drop dead PREFERRED_THEME_ORDER sort from bundler #335/docs(agent-browser): fix RenderMode absent-default (iframe, not ab-screencast) #337 and earlier) stays merged; track record remains uniformly positive.

• Below-threshold observation (no action) — nightly spend/duration ~2× baseline. This nightly ran 26 min / 58 turns / $6.00 vs the recent baseline of 7–20 min / 31–44 turns / $3.6–5.2. It is explained, not anomalous: the run produced 2 atomic PRs (one with a test) plus a body refresh — roughly double the usual single-output nightly — and both PRs' claims independently source-verify, so the turns map to real iterative work, not a stuck tool call or push-wait-fix loop. Under the 30-min flag threshold. Recorded only for cost-trend tracking.

• Below-threshold observation (no action) — open bot-PR backlog growing. 8 bot PRs now open pending maintainer (chore: update tend workflows (0.1.12 → 0.1.17) #340, fix(alert): emit onStateChange when WATCHING turns off on command finish #351, docs(agent-browser): repoint stale Source-of-truth pointers to the surface controller #352, chore: remove unused test bindings #357, chore(dor): drop unused ppid from ambient process declaration #359, test(pocket): stage push-disabled case in NeedsHomeScreenInstall story #360, fix(fake-scenarios): preserve endsWithPrompt through flattenScenario #362, docs(lath): correct replaceAtPath off-tree-path comment #363); no bot PR has merged since docs(agent-browser): fix RenderMode absent-default (iframe, not ab-screencast) #337 on 2026-08-03. This is within normal maintainer-review latency (nedtwigg lands these at their own pace) and is the maintainer's call, not a tend defect — none are stale-as-in-broken, all carry green checks. Noted so a future window can distinguish "maintainer busy" from an emergent "PRs consistently rejected" signal if the pattern ever inverts.

• Bot output discipline — 0 unexpected public comments. In-window bot writes were: 2 nightly PRs, the tend check: configuration drift on diffplug/dormouse #339 body refresh, and yesterday's tracking-issue append. No stray comments on notifications/mentions; the 2 tend-mention successes were handle-skipped bot-own-event no-ops.

• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: two source-verified atomic nightly PRs (a latent-fidelity fix with a proper regression test + a docs-comment correction), correct issue-body-refresh-not-new-comment restraint, disciplined review/mention/notification no-ops, no failures/timeouts/spam. Recorded as continued negative evidence. This is the 9th consecutive clean window.

---

Run 31372635650 — 2026-08-10T09:05:19Z (review-runs)

Clean, healthy window — 2 grounded atomic nightly PRs (#364 real behavior fix + tests, #365 test-assertion tightening), all checks green; disciplined review/mention/notification no-ops; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 31372635650
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31372635650
• Window: 2026-08-09 09:02 → 2026-08-10 09:02 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): 1 tend-nightly, 1 tend-weekly, 3 tend-review, ~10 tend-mention (3 skipped-at-trigger + verify-success/handle-skipped no-ops; 3 ran handle sessions on renovate pull_request_review events), ~30 tend-notifications (hourly silent no-ops), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was tend-nightly at 14 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the 9 runs with artifacts ≈ $11.76 list-price (tend-nightly $5.02 / 49 turns / 3.1M cache-read; tend-review 3× = $3.07; tend-weekly $1.95; tend-mention 3× = $1.13; tend-notifications $0.59).

Cross-checked chain (all grounded, all correct):

• tend-nightly (31368399586, $5.02, 49 turns, 14 min): ran the code-quality survey and shipped two atomic PRs, both with all checks green (Build & Test, Standalone Smoketest, Visual Regression, UI Tests unchanged):
  • #364 (+21/−6) — genuine behavior fix: detectResumeCommand scanned the 50-line scrollback oldest-first and returned the first matching resume command, so a pane that resumed more than once would persist a stale session id. Rewrites the scan newest-line-first, adds two tests (repeated-command + cross-pattern-type recency), and documents the recency invariant. Source-grounded, minimal.
  • #365 (+5/−1) — test-strengthening: the trailing-call assertion was > duringBurst - 1, i.e. >= duringBurst, which holds even if no trailing call fired — a tautology defeating the check. Tightened to > duringBurst (requires exactly one post-settle fire) with an explanatory comment. Correct atomic split from feat(resume): offer to re-run an agent session in a cold-restored pane #364.
• tend-review — 3 runs, all correct:
  • 31369054093 & 31369101566 (pull_request_target self-reviews of the bot's own feat(resume): offer to re-run an agent session in a cold-restored pane #364/test(throttle): assert the trailing call actually fires #365): both ran the diff read and correctly posted nothing — silent self-review-of-bot-PR no-ops (tend#607, settled). review check passed green on both.
  • 31326484722 (pull_request_target on renovate chore(deps): update vite #347 update vite, $1.30): reviewed a rebased head and correctly posted nothing — chore(deps): update vite #347 already carries a valid 08-03 APPROVED and the rebase was a no-content dep bump (re-review discipline: no-op-on-no-change). No duplicate approval.
• tend-weekly (31307033443, $1.95, 23 turns, 8 min): dependency-PR maintenance sweep, success, no unexpected output.
• tend-mention: the 08-09 09:59–10:02 pull_request_review-triggered handle runs (on renovate chore(deps): update github-actions #344 xterm / tailwind-variants / pnpm review events) were the benign self-loop / gate no-ops — small ($0.19–0.50), no comments posted.
• Bot output discipline — 0 unexpected public comments. A window scan of [workflow-audit] 1 change(s) on 2026-08-07 #358–test(throttle): assert the trailing call actually fires #365 found no bot issue-comments and no reviews beyond the intended silent no-ops. tend-notifications runs were hourly silent no-ops.
• Outcomes cross-check (Step 4): four bot PRs remain OPEN pending maintainer — #362, #363 (both 08-09), #364, #365 (both 08-10). No bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window. PR ages are within normal maintainer-review latency.
• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: two source-grounded atomic nightly PRs (a real stale-session-id fix and a tautological-assertion fix, correctly split), correct review discipline (self-PR no-op, rebase no-op-on-no-change), disciplined no-spam mention/notification behavior, no failures/timeouts. Recorded as continued negative evidence.

Run 31474034438 — 2026-08-11T08:41:50Z (review-runs)

Clean, healthy window — 1 grounded atomic nightly PR (#367 test-coverage, green), disciplined review/mention/notification no-ops; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 31474034438
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31474034438
• Window: 2026-08-10 08:39 → 2026-08-11 08:39 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): 14 tend-mention (3 skipped-at-trigger; rest verify-success + handle-skipped no-ops — none consumed tokens, confirming every mention trigger was undirected), 1 tend-nightly (success), ~24 tend-notifications (hourly 0–1 min silent no-ops), 5 tend-review (success, silent no-ops on bot's own PRs — intended per running-tend), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 14 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the 8 runs with artifacts ≈ $14.16 list-price (tend-review 5×, $6.34; tend-nightly $4.79 / 48 turns; yesterday's review-runs $1.53; tend-notifications 2×, $1.50).

• tend-nightly (31469671153, $4.79, 48 turns, 14 min): opened 1 well-grounded atomic PR — #367 (test-only, +25/−1 to standalone/src/updater.test.ts): covers the auto-updater's download-failure recovery path — asserts hasPendingUpdate() stays false after a failed download (no phantom pending install survives to quit) and that a second approveUpdate() re-invokes download (retry path, availableUpdate survives). No production change; all checks green (Build & Test, review, Standalone Smoketest, Cloudflare Pages). Body accurately describes the downloadApprovedUpdate catch-branch invariants and links docs/specs/auto-update.md.
• Prior bot PRs still open, no human pushback: #364 (resume-patterns fix+test), #365 (throttle test-assertion) — both from the prior window, awaiting maintainer; only automated Cloudflare deploy comments on all three.
• tend-review (5 runs): all silent no-ops on the bot's own PRs (feat(resume): offer to re-run an agent session in a cold-restored pane #364/test(throttle): assert the trailing call actually fires #365/test(updater): cover download-failure recovery and retry #367) — the diff-read-is-the-review pattern ruled intended in max-sixty/tend#607; no reviews posted, no spam.
• tend-mention / tend-notifications: disciplined no-ops throughout; no unexpected comments anywhere.

Assessment: 11th consecutive clean window this month. Bot behavior is grounded and disciplined; no gate-passing findings, nothing to fix.

[dormouse-bot]

---

Run 31580273569 — 2026-08-12T08:55:40Z (review-runs)

Clean, healthy window — 1 grounded atomic nightly PR (#369 test-coverage, green + source-verified), disciplined review/mention/notification no-ops; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence)
• Occurrences this run: 0 new defects
• Run ID: 31580273569
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31580273569
• Window: 2026-08-11 08:52 → 2026-08-12 08:52 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): ~10 tend-mention (2 skipped-at-trigger; rest verify-success + handle-skipped no-ops — none consumed tokens, confirming every mention trigger was undirected), 1 tend-nightly (success), ~24 tend-notifications (hourly 0–1 min silent no-ops), 3 tend-review (silent self-review of the bot's own #369 + two Renovate re-approvals), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 16 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action; single attempt, no failed steps — not stuck). Token spend across the 6 runs with artifacts ≈ $11.26 list-price (tend-nightly $5.67 / 45 turns / 3.2M cache-read; tend-review 4× = $4.98; tend-notifications $0.61).

• tend-nightly (31575744621, $5.67, 45 turns, 16 min): ran the code-quality survey and opened 1 well-grounded atomic PR plus refreshed one issue body —

  • #369 (test(dor): cover kill's conflicting-confirmation rejection, +15/−0, dor/test/cli-output.test.mjs + new snapshot). Independently source-verified against main: dor/src/commands/kill.ts:93–99 builds confirmations = [confirmDangerously===true, confirmIfRead!==undefined].filter(Boolean).length and rejects when !== 1, so the single guard covers both zero-mode and two-conflicting-mode inputs; the two flags are declared independently (optional: true, no parser-level exclusivity, kill.ts:58–59) so --confirm-dangerously --confirm-if-read done reaches the branch at runtime. The existing suite only exercised the zero-mode case (kill missing confirmation output, cli-output.test.mjs:707) — the two-mode case was unpinned. The new snapshot matches the missing-confirmation output byte-for-byte (same error/exit), guarding the mutual-exclusivity contract against a future refactor. No behavior change. All 4 checks green (Build & Test, Standalone Smoketest, Cloudflare Pages, review).
  • #339 config-drift issue — nightly updated the issue body (updatedAt 07:53, in-window) to match tonight's tend check rather than posting a new comment (last comment stays 2026-08-06). Correct no-spam restraint, consistent with prior windows.

• tend-review — 3 runs, all correct:

  • 31576589816 (pull_request_target self-review of the bot's own test(dor): cover kill's conflicting-confirmation rejection #369): ran the diff read and correctly posted nothing — silent self-review-of-bot-PR no-op (tend#607, settled; running-tend "Settled upstream rulings"). review check passed green.
  • 31562661357 (chore(deps): update storybook to v10.5.8 #346 renovate/storybook v10.5.7) and 31562735797 (chore(deps): update github-actions #344 renovate/github-actions): each re-approved a genuinely-new Renovate head SHA after a rebase (correct re-approve-on-rebase; approval doesn't carry across SHAs). Both PRs green.

• tend-mention burst (04:15–04:18, 6 legs): all benign no-ops from the chore(deps): update github-actions #344/chore(deps): update storybook to v10.5.8 #346 Renovate rebases — issue_comment legs on Cloudflare deploy comments + pull_request_review legs on the bot's own re-approval events, all handle-skipped with $0 spend (tend-mention: no-op session on bot's own APPROVED review of human PRs (upstream filing request) #168 pattern). The 07:53 leg (issues event on the tend check: configuration drift on diffplug/dormouse #339 body refresh) was skipped-at-trigger; the 08:03 leg (issue_comment on test(dor): cover kill's conflicting-confirmation rejection #369) was a bot-own-event handle-skip.

• Bot output discipline — 0 unexpected public comments. A window scan of issues/comments (created > window start) returned no new bot comments anywhere; the only in-window bot writes were the test(dor): cover kill's conflicting-confirmation rejection #369 PR, the tend check: configuration drift on diffplug/dormouse #339 body refresh, the two Renovate re-approval reviews, and yesterday's tracking-issue append. tend-notifications runs were hourly silent no-ops.

• Outcomes cross-check (Step 4): No bot PR was closed-as-unhelpful, reverted, or flagged post-merge this window (0 bot PRs closed-unmerged in-window). Prior merged batch (chore(themes): drop dead PREFERRED_THEME_ORDER sort from bundler #335/docs(agent-browser): fix RenderMode absent-default (iframe, not ab-screencast) #337 and earlier) stays merged; track record remains uniformly positive.

• Below-threshold observation (no action) — open bot-PR backlog now 12, no merge since docs(agent-browser): fix RenderMode absent-default (iframe, not ab-screencast) #337 (2026-08-03, 9 days). Open pending maintainer: chore: update tend workflows (0.1.12 → 0.1.17) #340, fix(alert): emit onStateChange when WATCHING turns off on command finish #351, docs(agent-browser): repoint stale Source-of-truth pointers to the surface controller #352, chore: remove unused test bindings #357, chore(dor): drop unused ppid from ambient process declaration #359, test(pocket): stage push-disabled case in NeedsHomeScreenInstall story #360, fix(fake-scenarios): preserve endsWithPrompt through flattenScenario #362, docs(lath): correct replaceAtPath off-tree-path comment #363, feat(resume): offer to re-run an agent session in a cold-restored pane #364, test(throttle): assert the trailing call actually fires #365, test(updater): cover download-failure recovery and retry #367, test(dor): cover kill's conflicting-confirmation rejection #369. Growing ~1/night. This is within normal maintainer-review latency (nedtwigg lands these at their own pace) and is the maintainer's call, not a tend defect — none are stale-as-in-broken, all carry green checks, none rejected. Continues the trend first noted 2026-08-09. Recorded so a future window can distinguish "maintainer busy" from an emergent "PRs consistently rejected" signal if the pattern ever inverts. If it inverts (a batch closed-unmerged), that becomes an actionable High-confidence signal.

• Below-threshold observation (no action) — repeated Renovate re-approval cost. chore(deps): update github-actions #344 (github-actions) now carries 6 APPROVED reviews (08-03/06/07/08/11/12) and chore(deps): update storybook to v10.5.8 #346 (storybook) 4 — each Renovate rebase of an unmerged dep PR re-fires a tend-review run that re-approves at ~$1–1.5. Over the 9-day-open lifetime this is meaningful cumulative spend, but the re-review is correct behavior (new head SHA; approval doesn't carry) and the root cause is the PR not merging — a maintainer choice outside tend's control, not fixable via skill/config without suppressing legitimate re-review. Below threshold; recorded for cost-trend tracking only. Downstream of the same merge-latency backlog above.

• Not a tend artifact: any [workflow-audit] / [security-audit] issue activity is authored by the separate app/github-actions workflows, not tend.

• Action: None. No finding passes the Gate-1/Gate-2 thresholds. Clean, healthy window: one source-verified atomic nightly test-coverage PR, correct issue-body-refresh-not-new-comment restraint, correct review discipline (self-PR no-op, re-approve-on-rebase), disciplined no-spam mention/notification behavior, no failures/timeouts. Recorded as continued negative evidence. This is the 12th consecutive clean window this month.

Run 31684173139 — 2026-08-13T08:59:22Z (review-runs)

Clean, healthy window — nightly refreshed the 0.1.15 workflow PR + drift issue with a clean codebase survey (no new code PR piled on), both reviews correct restraint; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 31684173139
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31684173139
• Window: 2026-08-12 08:56 → 2026-08-13 08:56 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): ~18 tend-notifications (hourly, all 0–1 min; all silent no-ops except one substantive at 17:41), 7 tend-mention (2 skipped-at-trigger; the rest verify-success + silent no-op, no artifacts), 1 tend-nightly, 2 tend-review, plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 11 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the 4 runs with artifacts ≈ $6.20 list-price (tend-nightly $3.32 / 47 turns; tend-review $1.16 + $1.09; tend-notifications $0.63).

• tend-nightly (31679577861, $3.32, 11 min, 47 turns): rolling survey (bucket 14/28, ~15 substantive source/test files) found the codebase all clean — no new code PR opened (good discipline: only acts when it finds something). Its two outputs both grounded: (1) refreshed the open workflow-update PR #340 from 0.1.14 → 0.1.15 (204 non-stamp lines, retitled to 0.1.12 → 0.1.15), CI green (Build & Test, Standalone Smoketest, Cloudflare Pages, review all SUCCESS); (2) refreshed the config-drift issue #339, explaining why security-audit drift cleared upstream (max-sixty/tend#825, #900). Verified: chore: update tend workflows (0.1.12 → 0.1.17) #340 shows 4/4 SUCCESS checks + bot COMMENTED (not APPROVED — correct for a generated self-PR).
• tend-review chore(deps): update pnpm to v11.21.0 #350 (31620786317, $1.16): correct trivial-skip on a Renovate pnpm bump (verified 1 addition / 1 deletion, pnpm 11.20.0 → 11.21.0). Prior APPROVE still stands at HEAD (reviewDecision: APPROVED), no stale review, no unresolved bot threads → stayed silent. Correctly noted renovate/artifacts FAILURE is Renovate-side lockfile tooling, not a repo CI job, and not something to push a fix for on a version-string-only change.
• tend-review chore: update tend workflows (0.1.12 → 0.1.17) #340 (31680092245, $1.09): reviewed the nightly's own generated workflow-update PR. Correct self-authored + dedup restraint — no approval possible on a bot PR, no new diff-grounded finding (generated files), prior COMMENT remains the active verdict → didn't re-post. Verified against source (0.1.15 stamps, id-token: write removal, environment: gating).
• tend-notifications (31623859703, $0.63): 1 notification — a Renovate CONTRIBUTOR bot-to-bot status message on chore(deps): update pnpm to v11.21.0 #350 reporting a transient 503 Service Unavailable (node-prebuild install failing pnpm-lock regeneration). Not directed at bot, asks nothing, Renovate auto-retries → marked read without responding. Correct.

Non-actionable observation (below gate — recorded, not acted on): 12 bot-authored PRs remain open/unmerged (#340, #351, #352, #357, #359, #360, #362, #363, #364, #365, #367, #369, back to Aug 3). This is a maintainer merge-backlog, not a bot defect — each PR is individually grounded (prior runs verified them), and today's nightly correctly declined to pile on a new code PR when its survey found nothing. No bot-behavior decision point exists here, so it fails Gate 1 (not a bot problem to fix via skill change). Noted for cross-run awareness; the signal to watch is whether the nightly ever starts opening redundant PRs against already-covered ground — it did not this window.

Run 31785383895 — 2026-08-14T08:51:33Z (review-runs)

Clean, healthy window — nightly shipped 2 grounded atomic PRs (#372 real theme-crash fix + regression test, #373 test rename), both green + mergeable; disciplined review/mention/notification no-ops; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 31785383895
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31785383895
• Window: 2026-08-13 08:49 → 2026-08-14 08:49 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out): 1 tend-nightly (success), 4 tend-review (success), 6 tend-mention (1 skipped-at-trigger; 5 verify-success + handle-skipped self-loop no-ops on bot's own PR/comment events), ~19 tend-notifications (hourly 0–1 min silent no-ops), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 17 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action). Token spend across the 8 runs with artifacts ≈ $14.15 list-price (tend-nightly $6.48 / 57 turns — higher than the typical ~37 because it produced 2 PRs incl. a real fix; tend-review $0.63–$1.38; yesterday's review-runs $1.67; tend-notifications $0.61–$0.87). All in normal ranges.

Cross-checks:

• tend-nightly (#372, #373) — fix(themes): drop malformed installed-theme entries instead of crashing #372 is a genuinely grounded bug fix: getInstalledThemes()'s Array.isArray() guard rejects a non-array top level but passes malformed elements ([null], missing id), so downstream .id derefs throw TypeError; fix filters to well-formed elements + adds a failing-before regression test. test(fake-adapter): rename re-kill test to match its assertion #373 renames a fake-adapter re-kill test to match its assertion. Both atomic, both all-checks-green, both MERGEABLE.
• tend-review — approved renovate #344 (7th re-approval since 08-03; renovate re-pushes each update, re-triggering review — intended, and chore(deps): update github-actions #344 staying open is a maintainer merge decision, not a tend defect). Other 3 review runs were the intended silent full-review no-ops on the bot's own PRs fix(themes): drop malformed installed-theme entries instead of crashing #372/test(fake-adapter): rename re-kill test to match its assertion #373 (settled: tend#607).
• tend-mention / tend-notifications — self-loop-guard no-ops on the bot's own PR/comment events; correct restraint.

Below-threshold observation (not a defect, no action): renovate PR #344 has now been re-approved 7 times over 11 days without merging. This is expected re-review-on-update behavior, not tend waste — recorded only so a future run doesn't mistake the repetition for a loop.

Run 31873380032 — 2026-08-15T08:04:22Z (review-runs)

Clean, healthy window — nightly refreshed the 0.1.17 workflow PR #340 (green) + #339 drift-footer with a grounded 12-file survey (no new findings); all reviews/mentions/notifications correct no-ops; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 31873380032
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31873380032
• Window: 2026-08-14 08:01 → 2026-08-15 08:04 UTC.

Runs (all success or skipped — no failed / cancelled / timed-out runs): 11 tend-mention (2 skipped-at-trigger; rest verify-success + handle-skipped no-ops, incl. a 5-run burst at 08-14 20:38–20:39 fired by a batch of renovate opens), 1 tend-nightly (success), ~30 tend-notifications (hourly sub-minute silent no-ops), 4 tend-review (all silent no-ops — see below), plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was nightly at 8 min, far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml; no non-tend-* workflows use the action).

Token spend across the 8 runs with artifacts ≈ $10.79 list-price (a large multiple below the Claude Code subscription effective rate): tend-review 4×≈$5.13, tend-nightly $2.69/41 turns, yesterday's review-runs $1.49, tend-notifications 2×≈$1.48. Nothing anomalous.

Cross-checks:

• tend-nightly (31870720457, 41 turns): grounded. Regenerated tend workflows 0.1.16 → 0.1.17, pushed the incremental delta to open PR chore: update tend workflows (0.1.12 → 0.1.17) #340, updated its title/body to the full 0.1.12 → 0.1.17 span — all four checks green (Build & Test, Standalone Smoketest, Cloudflare Pages, review). tend check returned the same 3 FAILs (environment, credential-environments, repo-secret-allowlist) as the standing drift issue tend check: configuration drift on diffplug/dormouse #339, so it correctly refreshed only the footer date (no new comment), per the identical-failure-set rule. Rolling survey bucket 16/28: reviewed 12 meaningful source/test files, no new findings; correctly did not re-file the already-tracked shell-quote.ts cmd-quoting bug (quoteCmdArg corrupts arguments containing a literal " on cmd surfaces #162). No new PR piled on top of the workflow refresh.
• 4 tend-review runs (renovate/storybook chore(deps): update storybook to v10.5.8 #346, renovate/vite chore(deps): update vite #347, renovate/github-actions chore(deps): update github-actions #344, self-review of tend/update-workflows chore: update tend workflows (0.1.12 → 0.1.17) #340): all fired on synchronize/rebase and posted no new review — the latest standing approvals on each PR predate this window (chore(deps): update github-actions #344 08-14T02:53, chore(deps): update storybook to v10.5.8 #346 08-12, chore(deps): update vite #347 08-03). Silent no-ops on already-approved renovate rebases + the self-review of the bot's own workflow PR are intended behavior; nothing was missed.
• tend-mention / tend-notifications: all correct no-ops (handle skipped / sub-minute silent exits). No spurious comments.
• Bot output acceptance: 0 new PRs/issues created this window. Standing backlog of open bot PRs (14, incl. chore: update tend workflows (0.1.12 → 0.1.17) #340 open since 08-03) and issues (tend check: configuration drift on diffplug/dormouse #339, dor send --stdin interprets backslash escapes in piped bytes #355, quoteCmdArg corrupts arguments containing a literal " on cmd surfaces #162) remains blocked on maintainer bandwidth / admin action — not a bot defect.

Assessment: clean. The only actionable output was the routine daily tend-workflow refresh (#340, green). No behavioral problems, no skill/workflow gaps, no gate-passing findings. Continues the multi-day healthy streak.

---

Run 31935382302 — 2026-08-16T08:11:08Z (review-runs)

Clean, healthy window — busiest day on record ($79.66, 44 sessions): 4 maintainer PRs reviewed across 25 review sessions with every finding accepted, 13 bot PRs merged, 3 new bot PRs all green; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + strong positive acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 31935382302
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31935382302
• Window: 2026-08-15 08:00 → 2026-08-16 08:02 UTC (anchored on predecessor run 31873380032).

Runs (234 total; 0 failure): 123 tend-mention (14 skipped-at-trigger, 109 handle — 98 of them sub-60s no-ops with no session artifact, i.e. the tend#606 fast-exit working), 47 tend-notifications (sub-minute silent no-ops), 33 tend-ci-fix (all skipped — watched CI stayed green), 29 tend-review (28 success, 1 cancelled), 1 tend-nightly, plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was 16 min (nightly, and two tend-mention handles), far under the 360-min default cap (no timeout-minutes override in .config/tend.yaml).

The one cancelled run is benign. tend-review 31920593364 on PR #378 was cancelled 8s after creation, before any agent booted. tend-review.yaml sets cancel-in-progress: false, but GitHub keeps only one pending run per concurrency group, so the newer queued run 31920597294 (created 6s later) superseded it. That run completed and posted; #378 received reviews at 01:50 and 01:56. No coverage lost, no cost.

Token spend ≈ $79.66 list-price across 44 runs with artifacts (1135 turns, 598K output tokens) — roughly 8× the recent ~$10/day norm. Not waste: it tracks real maintainer activity. tend-review alone was 25 sessions / $53.28 ($0.78–$4.76 each), driven by four nedtwigg PRs that went through many push→review rounds; tend-mention 12 / $16.18, tend-notifications 5 / $2.93, tend-nightly 1 / $5.62 (79 turns, three shipped PRs), yesterday's review-runs $1.65.

Cross-checks — every one came back accepted:

• tend-review on nedtwigg's docs(tend): record the org-level repo-secret-allowlist false positive #376/security(workflow-audit): report the unexplained, not the routine #377/security(vscode): authenticate host→webview messages with a per-boot token #378/chore(deps): regenerate dependency transparency snapshots #379: all four merged. security(workflow-audit): report the unexplained, not the routine #377 alone took 5 review rounds, and the maintainer took every finding verbatim — "Applied — good catch" (#376), "All three taken — the second one was a straight error on my part" (#377), "Both taken", "You're right that 'both' landed on the two review items and skipped the split-commit gap". Zero pushback, zero contradicted claims.
• tend-nightly 31932724352 (79 turns): three grounded atomic PRs, all green + mergeable — chore: update tend workflows (0.1.17 → 0.1.18) #380 (workflow regen 0.1.17 → 0.1.18), docs(tend): drop the org-secret false-positive note, fixed upstream in 0.1.18 #381 (correctly retires the now-stale org-secret false-positive note in running-tend, since 0.1.18 fixes it upstream), fix(browser): normalize a bare [::1] address-bar entry to http:// #382 (real normalizeNavUrl bug: authority.split(':', 1)[0] yields [ for [::1], so a bare bracketed IPv6 loopback normalized to https:// — fix + 4-case regression test, verified failing on main).
• Bot output acceptance: 13 bot PRs merged in-window (chore: update tend workflows (0.1.12 → 0.1.17) #340, fix(alert): emit onStateChange when WATCHING turns off on command finish #351, docs(agent-browser): repoint stale Source-of-truth pointers to the surface controller #352, chore: remove unused test bindings #357, chore(dor): drop unused ppid from ambient process declaration #359, test(pocket): stage push-disabled case in NeedsHomeScreenInstall story #360, fix(fake-scenarios): preserve endsWithPrompt through flattenScenario #362, docs(lath): correct replaceAtPath off-tree-path comment #363, test(throttle): assert the trailing call actually fires #365, test(updater): cover download-failure recovery and retry #367, test(dor): cover kill's conflicting-confirmation rejection #369, fix(themes): drop malformed installed-theme entries instead of crashing #372, test(fake-adapter): rename re-kill test to match its assertion #373); 0 closed unmerged; 3 open and green.
• Maintainer corrections: none. Swept both repo-wide comment endpoints with --paginate over the window; every non-bot human comment was nedtwigg accepting a bot review finding. No bot claim was contradicted, no label or state was reverted.

security-audit has been red 21 consecutive days and no tend workflow can see it

• Evidence level: High (21 consecutive daily failures, directly observed)
• Occurrences this run: 21 (2026-07-27 → 2026-08-16, unbroken)
• Run ID: 31935382302
• Workflow: https://github.com/diffplug/dormouse/actions/runs/31935382302
• Latest failing run: https://github.com/diffplug/dormouse/actions/runs/31927560706
• Classification: structural (no bot decision point) / cost class: neither tend waste nor a wrong tend action — it is an uncovered repo-health gap
• Gate assessment: no PR this run. Gate 1 and Gate 2 are satisfied, but I do not have a confident fix (see below), and the skill's rule is to note rather than ship speculation into a security-sensitive workflow. Recording the diagnosis and the aggregate so the maintainer has the number and the root cause in one place.

Root cause (first time diagnosed — the five prior review-runs entries only classified it "not a tend artifact"): the Audit against SECURITY.md step succeeds — is_error: false, permission_denials_count: 0, ~6.5 min of work — but writes neither audit-status.txt nor audit-report.md. The next step then fails on audit-status.txt: No such file or directory, defaults STATUS to FAIL, and appends a contentless "Audit step produced no audit-report.md" comment to #301 (now 20 such comments). So the audit agent is running and finishing cleanly without honouring the "Write audit-report.md / Write audit-status.txt" part of its prompt. History: clean 07-23→07-26, intermittent 07-20→07-22, unbroken failure from 07-27 — intermittent-then-permanent, which fits a prompt-compliance or claude-code-action behaviour change rather than a repo edit. I did not identify which, so any fix I proposed would be a guess.

Why tend never touches it: two independent reasons, both structural. (1) security-audit is not in workflows.ci-fix.watched_workflows in .config/tend.yaml (only CI and Chromatic are), so tend-ci-fix never fires — all 33 of its runs this window were skipped. (2) The daily comments on #301 are posted by github-actions[bot] using the default GITHUB_TOKEN, and GitHub does not raise workflow-triggering events for GITHUB_TOKEN actions, so no tend-triage / tend-mention run can fire either. (Same reason #383 [workflow-audit] opened this window with no triage run — that one is working as intended and needs a human, not the bot.)

Cost of the gap: release.yml dispatches this workflow on the release tag and gates publishing on its result, so a release-path security gate has been non-functional for three weeks. The available lever is one line — adding security-audit to workflows.ci-fix.watched_workflows — but it is the maintainer's call: nedtwigg spent this entire window hand-editing the adjacent workflow-audit.yaml (#377, five rounds) and plainly has this area in view, and pointing tend-ci-fix at a workflow whose failure may be unfixable from the repo side would produce a daily session that cannot close its own loop.

---

Run 32010212032 — 2026-08-17T08:28:33Z (review-runs)

Clean, healthy window — 4 bot PRs merged (bot PR queue now empty), nightly's 1520-line diff review found nothing to file, maintainer's one suggestion adopted without pushback; 0 failures, 0 timeouts, 0 spam, 0 new defects

• Evidence level: N/A (negative evidence + acceptance signal)
• Occurrences this run: 0 new defects
• Run ID: 32010212032
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32010212032
• Window: 2026-08-16 08:01 → 2026-08-17 08:25 UTC (anchored on predecessor run 31935382302).

Runs (62 total; 0 failure, 52 success / 10 skipped): 44 tend-notifications (43 sub-minute silent no-ops, 1 substantive), 8 tend-ci-fix (all skipped — watched CI stayed green), 6 tend-mention (2 skipped at trigger, 4 handled), 1 each tend-weekly / tend-review / tend-nightly, plus yesterday's tend-review-runs. No long-running (>30 min) or near-timeout jobs — longest was 11 min (yesterday's review-runs), then nightly and one tend-mention at 6 min each; the 360-min GitHub default applies (no timeout-minutes override in .config/tend.yaml). No open tend-outage issue, so nothing stranded to drain.

Token spend ≈ $11.25 list-price across 6 runs with artifacts (153 turns, 63.8K output tokens) — back to the ~$10/day norm after yesterday's $79.66 outlier. tend-nightly $4.10, yesterday's review-runs $3.85, tend-mention $1.46, tend-notifications $0.77, tend-review $0.58, tend-weekly $0.49. The 56 runs without artifacts are the fast-exit no-op paths and cost nothing.

Cross-checks — every bot output accepted:

• Bot PR dispositions: 4 merged in-window (chore: update tend workflows (0.1.17 → 0.1.18) #380 workflow regen 0.1.17→0.1.18, docs(tend): drop the org-secret false-positive note, fixed upstream in 0.1.18 #381 running-tend note removal, fix(browser): normalize a bare [::1] address-bar entry to http:// #382 the [::1] normalization fix, feat(resume): offer to re-run an agent session in a cold-restored pane #364 the cold-restore resume offer); 0 closed unmerged. The bot now has 0 open PRs — the whole queue drained.
• tend-mention 31973447657 on docs(tend): drop the org-secret false-positive note, fixed upstream in 0.1.18 #381 — the one substantive maintainer interaction. nedtwigg asked "I merged the upstream fix into ours. Should we just delete this paragraph from our SKILL.md?"; the bot deleted the section outright, force-pushed as c6a7870, retitled/rewrote the PR body to match the new approach, replied, and polled its own CI. Merged 3 minutes later. It explicitly chose the maintainer's approach over defending its own ("nedtwigg's suggestion was the right call") — correct read of a MEMBER directive on a bot-authored PR.
• tend-review 31973561487 on docs(tend): drop the org-secret false-positive note, fixed upstream in 0.1.18 #381 — fired on the synchronize from that force-push, found the PR already merged one minute earlier, and exited without posting. Correct restraint, not a missed review.
• tend-nightly 32005011888 (56 turns): read the full 1520-line 24h diff and filed nothing — the survey verified five specific claims from the merged work (VAPID loopback rejection covering the bracketed IPv6 form, detectResumeCommand's tail-walk termination, stripTerminalControls' \x1b\\-before-bare-ESC alternation, clearResumeOffer keying on the id that is ptyId today, ResumeBanner's no-op context default) rather than asserting the diff was fine. Bucket 18/28 of the rolling survey (25 files) also came back clean. Refreshed drift issue #339's body without commenting; regenerated workflows in a /tmp worktree and found them byte-identical, so no PR. Correct discipline — acts only when it finds something.
• tend-weekly / tend-notifications: both correct no-ops. Weekly found 0 dependency PRs; the one substantive notifications run traced a comment notification on fix(remote): make iOS push actually deliver, and let an expired Pocket session re-auth #338, confirmed the PR had merged and the review comments were already handled, marked the thread read, and posted nothing.
• Maintainer corrections: none. Swept both repo-wide comment endpoints with --paginate over the window. The only non-bot human comment was nedtwigg's suggestion on docs(tend): drop the org-secret false-positive note, fixed upstream in 0.1.18 #381 above, which the bot adopted — a direction change accepted, not a contradicted claim. No label or state was reverted.

RESOLVED — security-audit is green again; the 21-day red streak cleared on its own

• Evidence level: N/A (closes out the prior finding)
• Occurrences this run: 0 (streak broken)
• Run ID: 32010212032
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32010212032
• Passing run: https://github.com/diffplug/dormouse/actions/runs/31996256013 (2026-08-17T04:58Z)

The finding recorded under Run 31935382302 (21 consecutive daily failures, 2026-07-27 → 2026-08-16) is resolved. Today's scheduled security-audit run passed and posted "Audit passed at 2026-08-17T05:05Z" to #301, which is now closed. .github/workflows/security-audit.yaml was not edited — its last commit is 4c2d3a47 (renovate, 2026-08-15), which predates the last two failures — so nothing in the repo fixed it. That confirms the prior diagnosis: the audit agent was finishing cleanly without writing audit-report.md / audit-status.txt, i.e. a stochastic prompt-compliance lapse, not a structural break. The intermittent-then-unbroken-then-recovered shape fits that reading.

Future runs: do not act on the prior entry. Its one-line lever (adding security-audit to workflows.ci-fix.watched_workflows in .config/tend.yaml) would now point tend-ci-fix at a green workflow, and the evidence supporting it is void. If the failure recurs, start the count fresh from this date — and note the recovery means any fix would be chasing a probabilistic lapse, which the stochastic 5+ floor governs.

Near-miss: a fabricated commit SHA was composed into a PR body and caught by the pre-post link scan

• Evidence level: Medium (1 occurrence, caught before it shipped)
• Occurrences this run: 1
• Run ID: 32010212032
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32010212032
• Session: tend-mention run 31973447657 (claude-session-logs-n381)
• Classification: stochastic (a model lapse, not a missing step — the step exists and fired) / cost class: none — nothing wrong reached the public record
• Gate assessment: no action. Gate 1 fails outright (Medium needs 5+; the stochastic floor is 5+ regardless). Gate 3 has nothing to weigh — the existing guard already caught it. Recording only, so a future run can count occurrence 2 if this shape repeats.

While rewriting #381's body the session composed a blob/<sha>/SECURITY.md#L71 link with an invented SHA. It then ran the running-in-ci pre-post pass (git rev-parse HEAD && sed -n '71p' SECURITY.md, then grep -n 'github.com/' /tmp/pr-body.md), said plainly "I fabricated the SHA in that link — replacing it with the real one", and edited it before posting. The merged #381 body carries blob/c6a78708c9d91e144cd6623988ccf391d4ad573d/SECURITY.md#L71 — real head SHA, correct owner. So this is positive evidence that the guard works, filed as a data point about the underlying tendency rather than a defect. No guidance change is warranted: the instruction that caught it is already in the skill, and adding more would be noise.

Assessment: clean. Zero gate-passing findings, zero PRs, zero issues. main is green, the bot's PR queue is empty, the longest-running open finding closed itself, and the single maintainer interaction ended in a merge. Continues the multi-week healthy streak.

---

Run 32115440555 — 2026-08-18T08:25:10Z (review-runs)

Busiest day on record ($97.70, 46 sessions) and every bot finding landed: 4 maintainer PRs merged with findings applied verbatim, 1 hold vindicated by a close, 1 doc suggestion correctly withdrawn by the bot itself; 9 run failures, all one GitHub 503/429 incident, nothing stranded

• Evidence level: N/A (negative evidence + strong acceptance signal)
• Occurrences this run: 0 gate-passing defects
• Run ID: 32115440555
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32115440555
• Window: 2026-08-17 08:24 → 2026-08-18 08:17 UTC (anchored on predecessor run 32010212032).

Runs (215 total; 9 failure, 176 success, 30 skipped): 128 tend-mention (105 success, 8 failure, 15 skipped-at-trigger), 38 tend-notifications (37 success, 1 failure), 32 tend-review (all success), 15 tend-ci-fix (all skipped — watched CI stayed green), 1 tend-nightly, plus yesterday's tend-review-runs. tend-weekly is 17 9 * * 0 (Sunday), so none was due; tend-triage never fired because no human-authored issue opened. No long-running (>30 min) or near-timeout jobs — longest was 22 min (tend-review on #389), then nightly at 19 min, against the 360-min GitHub default (no timeout-minutes override in .config/tend.yaml). No open tend-outage issue, so nothing to drain.

Token spend ≈ $97.70 list price across 46 runs with artifacts (1372 turns, 666K output tokens) — a new high, past 2026-08-16's $79.66. tend-review 27 sessions / $68.87, tend-mention 11 / $13.61, tend-nightly 1 / $9.00, tend-notifications 6 / $4.00, yesterday's review-runs $2.22. Not waste: it tracks a very heavy maintainer day (five nedtwigg PRs, seven renovate PRs, ten merges). The 169 runs without artifacts are fast-exit no-op paths and cost nothing.

Cross-checks — every bot output accepted or correctly withdrawn:

• #389 (xterm coherent-set bump + lint), 5 review sessions, merged. nedtwigg verified the bot's three registry claims independently — "the counter-reset claim is real … and --canopy 0.20.0-sdf12.0 reproduced exactly as described" — and applied the xterm-lint.mjs message finding as 2367d7e1 ("Applied in 2367d7e — you're right"). One inline suggestion was declined; see the near-miss below, which the bot had already flagged and then withdrew in-thread.
• #390 hold vindicated. The bot withheld approval on renovate's bare @xterm/xterm bump — "this set pins core to a commit no addon was published from, and chore(deps): bump xterm to a coherent commit set, and make incoherent sets fail CI #389 (open, yours) already lands the coherent version" — and the PR was closed unmerged. Same for #384, superseded by chore(deps): bump xterm to a coherent commit set, and make incoherent sets fail CI #389.
• #396 merged. The bot's SPA-fallback finding ("the fallback derives its class from the request path, but the response it returns is always the shell") produced 8df793bc fix(server): stop answering a hashed-asset miss with the shell; the bot then approved. #391 merged after the bot's red-CI analysis produced 24b2a2a6 fix(vscode-ext): build dor-lib-common before typechecking.
• #398 (open): two grounded reviews, one built the branch and ran pnpm --filter server test (83 pass) before commenting. Awaiting the maintainer.
• Bot PR dispositions: #393 merged (the nedtwigg-vs-ntwigg handle warning); #386 closed but adopted — "Folded into chore(deps): bump xterm to a coherent commit set, and make incoherent sets fail CI #389, which carries the @xterm/** scope glob". One new PR, #399, green.
• tend-nightly 32109652981 (100 turns, $9.00): PAT scope audit, tend check, drift issue #339 refreshed without commenting, full 24h diff read across server/+lib/remote/+standalone/, rolling file survey (bucket clean), workflow regen byte-identical so no PR — and one PR, ci(security-audit): distinguish an inconclusive audit from a FAIL #399, with a stubbed-gh verification matrix over six status-file shapes rather than an assertion that it works.
• Maintainer corrections: none contradicted a bot claim. Swept both repo-wide comment endpoints with --paginate over the window. The four nedtwigg comments were three acceptances and one declined suggestion the bot had itself identified as stale.

Near-miss: a 503 POST retry loop landed an APPROVE on a commit the session never read

• Evidence level: High (1 occurrence)
• Occurrences this run: 1
• Run ID: 32115440555
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32115440555
• Session: tend-review run 32052492341 (claude-session-logs-n389)
• Classification: structural (given a retried POST that spans a push, the same thing happens every time — the head check has no decision point) / cost class: wrong outward action (an APPROVED review anchored to unexamined code, plus one stale inline suggestion posted publicly)
• Gate assessment: no PR this run. Gate 3 would admit real complexity for a wrong outward action, but Gate 1 fails — High needs 2–3 and this is occurrence 1, with no prior entry of this shape in the log (searched 503 / retry loop / head moved / commit_id across both evidence comments). The trigger also requires an upstream outage to coincide with a push. Recording so a future run can count occurrence 2.

The mechanism, traced in the session JSONL: the session pinned HEAD_SHA=ff887804, ran the review skill's posting mechanics (read -r CURRENT_HEAD PR_STATE < <(gh api …/pulls/389 …)) and got cur=ff887804 state=open, then POST /pulls/389/reviews returned HTTP 503. It wrapped the POST in an ad-hoc for i in $(seq 1 10) retry; attempt 1 failed, a later attempt succeeded — and nedtwigg had pushed 73311f83 in between, so GitHub anchored review 4953301000 (APPROVED) to a commit the session had never read. Its last paragraph and the docs/specs/webgl-text.md:56 suggestion both described the pre-73311f83 shape, and 73311f83 had already fixed the thing the suggestion asked for.

The skill's guard cannot see this. plugins/tend-ci-runner/skills/review/SKILL.md says "Before posting, verify HEAD hasn't moved" — the check is before the POST, so a POST that has to be retried outlives it. The narrow subtraction would be to put the head re-check inside the retry loop, or to read back .commit_id after posting and correct on mismatch.

Why the damage was nil, and why that matters for the count. The bot detected the mismatch on its own next round and posted review 4953337916 six minutes later — "Correcting my own approval above: it was written against ff887804 but a 503 retry loop delayed the POST until 73311f83 had landed" — re-verified against the new head, and concluded "73311f83 is a better answer than the caveat I suggested". When nedtwigg declined the stale suggestion, the bot replied "Agreed — withdraw the suggestion. I wrote it against the pre-73311f83 shape". So the self-correction path works unaided; that is an argument for waiting for a second occurrence rather than writing guidance now.

Nine run failures, all one GitHub incident; nothing stranded, and no outage issue was filed because the failures were pre-flight

• Evidence level: N/A (upstream incident, recorded for the aggregate)
• Occurrences this run: 9
• Run ID: 32115440555
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32115440555
• Classification: structural (no bot decision point) / cost class: wasted compute, ~9 sub-15s job starts
• Gate assessment: no action. The skill's own rule for an infra cluster is not to open a fix PR. The outage-issue gap below would need new mechanism in a generated workflow, which Gate 3 refuses at waste class.

Eight tend-mention runs failed between 17:23 and 18:24 UTC, all on HTTP 503: No server is currently available from api.github.com — five in the verify job's "Verify bot engagement" step (GraphQL) and three in the relay job's "Re-enter on an admitted ref" step (REST dispatch). The ninth, tend-notifications 32041053324 at 15:02, failed at Set up job on 429 Too Many Requests downloading max-sixty/tend@0.1.18 from codeload, after three backoffs. Same incident window that produced the near-miss above.

Nothing was actually lost. The stranded relays were pull_request_review events on #384/#389/#390; the one that carried a maintainer review (4953389709, nedtwigg, 18:20:40) was covered by the sibling dispatch 32054397635 that had already succeeded at 18:19:59 — the bot replied to both of that review's inline threads at 18:24. tend-notifications is scheduled and recovered on its next tick.

Worth noting for the aggregate: no tend-outage issue was filed for any of the nine, because all nine died in the verify / relay / Set up job pre-flight steps — before the tend action that files those rows ever ran. So a pre-flight-class outage is invisible to the next day's drain step, which reads the outage issue. This window shows the gap is survivable (sibling coverage plus scheduled retry), but it is the reason the drain came back empty despite nine failures.

security-audit FAIL recurred — occurrence 1 since the 2026-08-17 reset, already addressed by #399

• Evidence level: N/A (continues a reset count; the actionable half is already in review)
• Occurrences this run: 1
• Run ID: 32115440555
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32115440555
• Failing run: https://github.com/diffplug/dormouse/actions/runs/32100728239

The prior entry (Run 32010212032) closed the 21-day streak as a stochastic prompt-compliance lapse and said to start the count fresh. It recurred once, on 2026-08-18, filing #397 [security-audit] FAIL. Same signature as before: the Claude step finished cleanly ("subtype": "success", is_error: false, 36 turns, 0 permission denials) and simply never wrote audit-status.txt or audit-report.md, so the missing file fell through to the || echo "FAIL" default. No further action needed from review-runs — nightly already opened #399, which makes the tracker say "inconclusive" instead of "FAIL" for a no-verdict run (still exiting 1, still filing an issue) rather than trying to assert its way out of the nondeterminism. Count stands at 1; the stochastic 5+ floor governs anything beyond #399.

Observation, not a finding: #399's four self-review rounds are the designed path, not a loop

Bot PR #399 drew four tend-review rounds and three dispatched tend-mention fixes in 27 minutes (~$10.89 total: reviews $3.32/$1.42/$1.52/$0.60, mentions $1.27/$1.61/$1.15). Each round found a real defect in the previous one — whitespace-only status file, an inconclusive title outliving a real FAIL, an unguarded retitle — and the sequence converged. This is exactly the split upstream max-sixty/tend#1007 is landing (reviewer stops at the review on bot-authored PRs; the dispatched mention session applies it), and dormouse's runs already behave that way. Filed here only so the cost is on the record; no gap.

Assessment: clean. Zero gate-passing findings, zero PRs, zero issues. main is green, the maintainer applied or independently verified every bot finding, the one declined suggestion was withdrawn by the bot before being asked, and the day's nine failures were one upstream incident that cost nothing. The one thing to watch is the 503-retry approval anchoring — occurrence 1 of a structural race, recorded for a future count.

---

Run 32231741521 — 2026-08-19T08:26:26Z (review-runs)

Second record day in a row ($135.14, 51 sessions) and every bot output landed: 6 maintainer PRs merged, 1 directive followed to a merge, 0 maintainer corrections; 4 failures were one Anthropic 529 incident; first >30-min jobs on record, both spent on CI-flake re-polling

• Evidence level: N/A (negative evidence + strong acceptance signal)
• Occurrences this run: 0 gate-passing defects
• Run ID: 32231741521
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32231741521
• Window: 2026-08-18 08:15 → 2026-08-19 08:16 UTC (anchored on predecessor run 32115440555).

Runs (221 total; 4 failure, 7 cancelled, 189 success, 21 skipped): 138 tend-mention (119 success, 2 failure, 7 cancelled, 10 skipped-at-trigger), 37 tend-notifications (all success), 33 tend-review (31 success, 2 failure), 9 tend-ci-fix (all skipped — watched CI stayed green), 2 tend-triage (1 success, 1 skipped), 1 tend-nightly, plus yesterday's tend-review-runs. tend-weekly is 17 9 * * 0 (Sunday), so none was due. Census is 221, not a page boundary, and under the wider FETCH_FROM .total_count.

Token spend ≈ $135.14 list price across 51 runs with artifacts — a new high, past 2026-08-18's $97.70. tend-review 31 sessions / $105.93, tend-mention 12 / $11.99, tend-nightly 1 / $6.83, yesterday's review-runs $4.61, tend-notifications 5 / $4.57, tend-triage 1 / $1.21. Not waste: it tracks the heaviest maintainer day yet — 6 merges, 41 pushes across 9 PRs, and 11 review rounds on #401 alone. The 170 runs without artifacts are fast-exit gate paths and cost nothing.

The 4 failures are one upstream Anthropic incident, and nothing was stranded. tend-review 32159817174 (#401), tend-mention 32164406997 (#399), tend-review 32164453132 (#402), tend-mention 32166793041 (#402), all between 16:32 and 17:44 UTC. Every session JSONL ends on the same synthetic assistant message: API Error: 529 Overloaded. This is a server-side issue, usually temporary. Unlike yesterday's pre-flight 503s, these died inside the tend action, so the outage row was filed correctly — #403 carried all four, and nedtwigg closed it at 01:11. Drain verified as unnecessary: #401 and #402 both drew later successful reviews on newer heads (32165366707, 32168881750); the #399 mention was stranded on the comment obsoleted by #402, posted as nedtwigg closed that PR, so re-running it would have been a no-op on a closed thread.

The 7 cancelled tend-mention runs are concurrency dedup working. All fired in a 12-second burst at 18:16:40–18:16:52 when nedtwigg posted 8 inline review replies at once on #401; the group cancelled the superseded legs and the survivor handled the batch. No Claude session started in any of the 7.

Cross-checks — every bot output accepted, zero corrections:

• 6 PRs merged, all bot-reviewed: #402, #404, #405, #407, #408, #409. main CI is green on every resulting merge commit — no approval led to a red default branch.
• Findings applied verbatim. On Fix the audit's blind spot, and the one it was blind to #402: "All three suggestions taken, plus both of the smaller notes… You're right that the comment stated an invariant nothing enforced, and the trace through run-claude-sdk.ts is the part I hadn't checked", then "All three taken… The retitle inversion — you're right, and it's a better catch than the original suggestion." On Make VS Code a first-class remote Host #401: "Thanks, the two cache findings were the real content of this PR and I had both wrong", and "Closed the multi-window gap in 09f31cb — it was worse than the spec text I wrote for it."
• A maintainer directive followed end to end. On Move standalone shell selection into Settings #409 nedtwigg wrote "Yes please, push that one-liner fix in this PR."; the dispatched tend-mention 32201516196 pushed 49468a1f, verified pnpm lint:specs and the gating checks, reported both, and the PR merged 26 minutes later.
• tend-triage 32179854591 on nedtwigg's #406 "Add standalone workspaces": correctly recognised it as stages 3–4 of the already-specced workspaces-rollout scope and answered with a SHA-pinned link into docs/specs/layout.md ## Future rather than re-deriving a design. Exactly the restraint AGENTS.md's spec-lifecycle rules ask for.
• tend-nightly 32226163807 (73 turns, $6.83, 27 min): opened 2 small green PRs — #410 (−11/+14, one file) and #411 (+8, running-tend overlay). Both awaiting the maintainer.
• Maintainer corrections: none. Swept both repo-wide comment endpoints with --paginate over the window (24 non-bot comments). Every substantive one is an acceptance; the rest are Cloudflare deploy bots. No comment contradicted a bot claim.

Verified: the 503 approval-anchoring near-miss did not recur — count stays at 1

• Evidence level: High (still 1 occurrence, from Run 32115440555)
• Occurrences this run: 0
• Run ID: 32231741521
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32231741521
• Gate assessment: no action. Gate 1 needs 2–3 for High; this window adds nothing.

Checked all 25 bot reviews submitted in-window (PRs #401/#402/#404/#405/#407/#408/#409/#410/#411) by comparing each review's commit_id against that PR's commit timeline. Every review was anchored to the PR head at its own submission time — including all 6 APPROVED reviews and the 11-round sequence on #401, where heads moved every few minutes. The two long sessions below also re-read the live head mid-run and re-posted their tend-review/<n> "examined" status against the new SHA rather than the stale one. The retry-loop trigger requires an upstream outage to coincide with a push; today's outage was Anthropic-side and killed sessions outright rather than delaying a POST, so the race had no opportunity to fire.

First >30-min tend jobs on record: two review sessions spent ~25 min each re-polling a flaky Standalone Smoketest

• Evidence level: Medium
• Occurrences this run: 2 (one window)
• Run ID: 32231741521
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32231741521
• Sessions: 32165366707 (claude-session-logs-n402, 40 min), 32169872312 (claude-session-logs-n404, 40 min)
• Classification: stochastic (the model chose to re-enter the poll; a replay could just as easily stop) / cost class: wasted compute (~20 job-min plus tokens; both sessions still shipped a correct approval and a grounded flake note)
• Gate assessment: no PR. Gate 1's stochastic floor is 5+ and this is occurrence 1 of the shape. Gate 3 independently refuses anything but a free fix at waste class, and the applicable rule already exists in /tend-ci-runner:running-in-ci ("the cap is the whole poll budget… don't re-enter the loop"), so a repo overlay would be a duplicate paragraph, not a new one. Recording for a count.

Both jobs ran 40 min against the 360-min GitHub default (no timeout-minutes in .config/tend.yaml or the generated workflows), so neither was near-timeout. Both were gated approvals whose CI went red on Standalone Smoketest, which failed at the apt step — the flake ci.yml already mitigates with timeout + 3 attempts across rotating mirrors, and which a rerun cleared both times (32165369202, 32166771196). The excess is narrow and only in the first session: it ran poll-pr-checks.sh 402 1881a84a three times — once on arrival, once after gh run rerun, and a third time with no intervening action, each poll costing up to ~9.5 min. The second session was better behaved (poll → rerun-failed-jobs.sh → one confirming poll). Both ended correctly: approval standing, tend-review/<n> status posted, and a grounded flake note on the thread naming the apt failure rather than a false-green claim.

tend-notifications: 5 Claude sessions, all zero-action — unchanged from prior windows

• Evidence level: Low (matches every prior window in this log)
• Occurrences this run: 5
• Run ID: 32231741521
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32231741521
• Classification: structural / cost class: wasted compute ($4.57)
• Gate assessment: no action. Gate 3 at waste class admits only a cadence knob, and the */15 cron is already the maintainer's lever.

Of 37 scheduled runs, 32 exited at the unread-count gate for ~$0 and 5 opened a session (32171606388, 32184305537, 32203494217, 32208825714, 32219391593). Every one read the triggering thread, confirmed a sibling workflow had already handled it or that nothing was directed at the bot, marked read, and posted nothing — "Responded to: none. Skipped: none. Cross-repo: none." That is the safety net doing its job, and the diligence (reading #402's merge state, #401's timeline, #403's close) is what makes the no-op trustworthy rather than lazy.

Not tend artifacts, noted so a future run doesn't chase them

• #412 [workflow-audit] 3 unexplained change(s) (07:50, in-window) lists nedtwigg's own three security-audit.yaml commits from the merged Fix the audit's blind spot, and the one it was blind to #402. Authored by app/github-actions, so the generated tend-triage if: correctly skipped it; the maintainer closes these (predecessor [workflow-audit] 4 unexplained change(s) on 2026-08-18 #400 closed 19:41 the same day). No tend gap.
• security-audit #397 closed itself at 05:01 on a PASS, with Fix the audit's blind spot, and the one it was blind to #402's inconclusive-vs-FAIL distinction now on main. The stochastic count from Run 32010212032 stays at 1.

Assessment: clean. Zero gate-passing findings, zero PRs, zero issues. main is green on all six merges, every bot finding was applied or independently verified by the maintainer, no comment contradicted a bot claim, the one 529 incident filed its outage row correctly and stranded nothing, and the anchoring race the last window flagged did not recur. The one new thing to watch is poll re-entry on flaky CI — occurrence 1 of a stochastic shape, recorded for a count.

[dormouse-bot]

---

Run 32348132575 — 2026-08-20T08:33:22Z (review-runs)

First gate-passing defect in this log: a review's findings on a superseded PR are never fetched by its successor, and one merged to main

• Evidence level: Critical (a false-clean review plus an approval on a diff carrying a defect the bot had itself published three hours earlier)
• Occurrences this run: 1
• Run ID: 32348132575
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32348132575
• Sessions: 32304083586 (claude-session-logs-n398), 32313851449 + 32318485835 (claude-session-logs-n416)
• Classification: structural — the predecessor is never in context, so no decision point existed and a replay fails identically every time / cost class: wrong outward action
• Gate assessment: passes all three. Gate 1: Critical acts on 1. Gate 2: a targeted fix (fetch two endpoints, re-check the findings) at Normal bar. Gate 3: the cost is a wrong outward action, not waste — the fix is three gh calls and a sentence on what the result means, not new mechanism.
• Action: #421 (overlay note) and #420 (the defect that shipped).

tend-review 32304083586 opened on #398 at 21:28:55; nedtwigg closed the PR at 21:31:54, ~3 min in. The session handled the closure well — noticed it, and at 21:45:06 posted its findings as a plain comment rather than a review, because tailnet-deploy was still live and one finding was a production bug. That comment is the highest-quality artifact in the window: an inode-reuse race in stillOurs measured 10/10 on the runner's ext4, with the deterministic test failure it causes traced to the same root. nedtwigg adopted it in 2bd97f09 two hours later, choosing dev + ctimeNs over the suggested birthtimeNs.

The successor #416's body opens Supersedes #398 with the final review fixes. Both #416 review sessions read that body — it is in the tool output in each log — and neither fetched #398. The draft session ran gh pr view 416, bot-review-state.sh 416, sed -n 1,200p peer-link.ts; the full pass ran gh pr view 416 --json comments,reviews,body, bot-review-state.sh 416, cat -n peer-link.ts | sed -n '1,340p'. Both tools are scoped to the PR under review by construction. The draft closed with "Nothing else surfaced"; the PR was approved at 02:27:46 and merged at 03:00:53. The other two findings from the #398 comment went:

• ensureToken adopting an empty token — merged to main. The full pass's 1,340p dump did include the lines (peer-link.ts:221), but the only two occurrences of ensureToken in that session log are both inside cat output; nothing reasons about them. Verified still live on f1873f9a: writeFile(..., { flag: 'wx' }) creates before it writes, so a window reading in the gap takes '', and onServerFrame's !serverToken then rejects every hello for the broker's lifetime while every other window retries at 1 Hz. Fixed in #420, with two tests that fail on main.
• The service.ts comment naming a #lifecycle field that never existed — re-derived from scratch by tend-nightly ~10 h later and fixed in #418. The finding survived; the session cost of finding it was paid twice.

Adjacent upstream work, noted so a future run doesn't read it as a duplicate: max-sixty/tend#1014 adds a prior-rejection fetch to nightly/triage/ci-fix dedup — same shape (a fact on a closed PR that no later run fetches), opposite direction (a rejection lost → work redone; here a finding lost → a defect ships), different skill. #421 asks nedtwigg whether to file the generic version upstream, per the ask-each-time rule in this overlay.

Window census and spend

• Evidence level: N/A (context)
• Occurrences this run: 0
• Run ID: 32348132575
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32348132575
• Window: 2026-08-19 08:15 → 2026-08-20 08:19 UTC (anchored on predecessor run 32231741521).

Runs (129; 0 failure, 1 cancelled, 115 success, 13 skipped): 63 tend-mention (58 success, 1 cancelled, 4 skipped-at-trigger), 38 tend-notifications (all success), 17 tend-review (all success), 9 tend-ci-fix (all skipped — watched CI stayed green), 1 tend-nightly, plus yesterday's tend-review-runs. tend-weekly is 17 9 * * 0 (Sunday), so none was due. Census is 129, not a page boundary, and under the wider FETCH_FROM .total_count. Zero failures — the first all-green window since 2026-08-16, and no tend-outage issue open, so nothing to drain.

Token spend ≈ $77.80 list price across 24 runs with artifacts, down from yesterday's $135.14. tend-review 15 sessions / $59.39, tend-nightly 1 / $6.84, yesterday's review-runs $4.68, tend-mention 5 / $5.25, tend-notifications 2 / $1.64. Longest job 23 min (nightly), against the 360-min GitHub default — no >30-min job and no near-timeout, so yesterday's poll-re-entry shape did not recur (count stays at 1). The single cancelled tend-mention (32261219904, 13:58:27) is concurrency dedup during a four-comment burst; no Claude session started in it.

Cross-checks — 6 PRs merged, every bot finding accepted, zero maintainer corrections

• Evidence level: N/A (negative evidence + acceptance signal)

• Occurrences this run: 0

• Run ID: 32348132575

• Workflow: https://github.com/diffplug/dormouse/actions/runs/32348132575

• Merged in-window: #413, #414, #415, #416, and the bot's own #410. main CI is green on every resulting commit.

• Every approval was anchored to the PR head at submission. Fix spoken alert interruption and feedback #413 → 1f4c8b5a, ci(release): gate provenance OIDC tokens on a v*-tag-only environment #414 → 58670a3e, Keep a minimized browser Surface's DOM alive by parking it #415 → da2c4199, Self-host Dormouse: service-based remote Host and review hardening #416 → 5ce9dce6; each is that PR's merge head. The 503 approval-anchoring near-miss from Run 32115440555 stays at 1 occurrence.

• Maintainer corrections: none. Both repo-wide comment endpoints swept with --paginate over the window: 10 non-bot comments, 6 of them Cloudflare deploy bots. All four nedtwigg comments are acceptances — "Good catch, and the second half is the more valuable one — applied in 58670a3" (ci(release): gate provenance OIDC tokens on a v*-tag-only environment #414), "Confirmed and fixed at this depth" / "Applied the fallback as suggested, and also fixed the root cause" / "Good catch — applied, with your admit(leaves(tree))" (Keep a minimized browser Surface's DOM alive by parking it #415, three inline replies). No comment contradicted a bot claim.

• The one bot PR closed unmerged, #411, is not a rejection of the analysis. It proposed recording the credential-environments tend check FAIL as maintainer-owned so nightly would stop re-filing, and its own body offered "If you'd rather the nightly keep filing this one, close this." nedtwigg closed it at 13:55:38 and merged the real fix, #414 (v*-tag-only environment for the provenance OIDC tokens), six minutes later — the suppression became moot because the underlying FAIL was fixed. This morning's nightly confirms it: tend check is now every line PASS. Worth noting the bot had already identified that fix in skills(running-tend): record that credential-environments is maintainer-owned #411's body and held it as a standing offer rather than opening it; opening it would have been the better call, but one occurrence of "offered instead of shipped" on a release.yml change touching OIDC permissions is not a pattern to act on.

• tend-nightly 32342390821 (99 turns, $6.84, 23 min): full 8-step sweep on a 133-file / +16,720 merge, shipped #418 — performEnrollment never ran isEnrollment at the one site that creates an enrollment, so a malformed 200 persists undefined as ConnectionPolicy.rpId/origin and the machine silently un-enrolls at relaunch. Three regression tests, all failing on main. Verified the two open bug issues (quoteCmdArg corrupts arguments containing a literal " on cmd surfaces #162, dor send --stdin interprets backslash escapes in piped bytes #355) are still genuinely broken rather than closing them.

• The tend-review → tend-mention handoff on the bot's own PR worked. tend-review 32343086642 posted an actionable review on fix(remote): reject an enroll response that is not an enrollment #418; tend-mention 32343487135 pushed b841cd7a addressing it as author; the nightly session, still polling, read that commit instead of duplicating it and re-verified against the new head. Three sessions, no collision, no duplicated work.

tend-notifications: 2 Claude sessions, both zero-action — unchanged from every prior window

• Evidence level: Low (matches every prior window in this log)
• Occurrences this run: 2
• Run ID: 32348132575
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32348132575
• Classification: structural / cost class: wasted compute ($1.64)
• Gate assessment: no action. Gate 3 at waste class admits only a cadence knob, and the */15 cron is already the maintainer's lever. Cheapest occurrence yet — 36 of 38 runs exited at the unread-count gate for ~$0.

Both sessions (32264698265, 32328312416) read the triggering threads, established that each notification's updated_at sat ~20 s after its PR's mergedAt (the merge itself, not an unanswered comment), confirmed a sibling workflow had already answered every inline reply, marked read, posted nothing.

Not tend artifacts, noted so a future run doesn't chase them

• #417 [security-audit] INCONCLUSIVE on 2026-08-20 and #419 [workflow-audit] 1 unexplained change(s) are both authored by app/github-actions, which the generated tend-triage if: correctly skips. security-audit failed on f1873f9a at 04:53 — occurrence 2 since the 2026-08-17 reset, and Fix the audit's blind spot, and the one it was blind to #402's inconclusive-vs-FAIL distinction is doing its job: it filed INCONCLUSIVE, not FAIL. The maintainer closes these.

---

Run 32462632232 — 2026-08-21T08:26:52Z (review-runs)

Window 2026-08-20T08:18:08Z → 2026-08-21T08:20:03Z (anchored on predecessor run 32348132575's start). 141 runs censused, 0 failures, 0 cancellations, 0 outage issues, 0 stranded triggers. No gate-passing defect this window.

Clean window — 3 bot PRs merged, 2 maintainer PRs merged with every bot finding accepted, and the one maintainer correction was already converted to durable overlay guidance inside the same window

• Evidence level: n/a (no defect)
• Occurrences this run: 0
• Run ID: 32462632232
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32462632232

Census. 141 completed runs, 133 success + 8 skipped (7 tend-ci-fix no-ops on green watched workflows, 1 tend-mention on a renovate Dependency Dashboard edit). By workflow: tend-mention 78, tend-notifications 34, tend-review 20, tend-ci-fix 7, tend-nightly 1, tend-review-runs 1. 29 of those spun up a Claude session; the rest were pre-flight skips.

Spend. $71.96 / 506,847 output tokens / 937 turns across 29 sessions — down from $135.14 (2026-08-19) and $97.70 (2026-08-18). tend-review $41.40 over 17 sessions, tend-mention $10.21 over 8, tend-review-runs $10.99 over 1, tend-nightly $7.86 over 1, tend-notifications $1.50 over 2.

Timeouts. No timeout-minutes is set on any tend-* workflow, so the 360-min GitHub default applies and the near-timeout threshold is 324 min. Longest job was 36 min. Zero near-timeouts.

Cross-checks. Merged in-window: #418, #420, #425 (bot) and #422, #426 (maintainer). Closed unmerged: #421 (bot). Open: #427, #428, both opened by tonight's nightly, both green.

• #426 — the review caught a real defect the maintainer had not seen: outputStyle: "Concise" is not a resolvable style name, so the setting was a silent no-op. nedtwigg: "Right on the substance, and the echo critique is correct — output_style in the system/init event reports the configured string whether or not it resolved, so that check proved the setting was read, not applied." Fixed in 69642eb3, renamed to dormouse-concise in 1fdfc631, merged.
• #422 — six inline findings plus two body notes across two review passes, all addressed, then approved. #423 — three review passes, every prior finding traced to closure rather than taken on the commit message. #424 — commented, no approval, PR still open.
• #428 — the review left three efficiency findings on nightly's own PR (an unused pnpm install/pnpm/action-setup pair costing 48s on Windows, and a missing timeout-minutes on premium-billed runners). A tend-mention run actioned them in ce8a2e21 within four minutes.

Maintainer corrections: 1, already closed as a loop. nedtwigg closed #421 with "When I start over, I usually want to start over. The original conversation grew too unfocused and out of hand. This change would make that worse - closing without merge." The tend-mention run on that close (32413788390) answered the PR's open upstream-filing question with a no, and recorded the ruling in the running-tend overlay as #425, which merged 50 minutes later. Consequence for future sweeps: the superseded-PR finding logged under Run 32348132575 is settled and must not be re-derived, re-filed here, or filed at max-sixty/tend — the overlay says so explicitly.

Nightly (32457204363, 22 min, $7.86, 86 turns) — two grounded PRs, both green.

• #427 — clipboard_win.rs::read_text bound GlobalSize as _size and discarded it, then scanned the locked CF_UNICODETEXT block for a NUL terminator with nothing to stop it. The terminator is the producing process's promise, so a block written without one reads past the allocation in unsafe code, and the payload is pasted into the user's terminal. The session verified the extracted utf16_up_to_nul helper by compiling it standalone under rustc in /tmp, since the crate itself won't build in the runner.
• #428 — trying to verify that one-liner surfaced the larger gap: CI's only Rust step was cargo check on Ubuntu, so the crate's 23 unit tests had never run anywhere and no #[cfg(windows)]/#[cfg(target_os = "macos")] code had ever been compiled by a check, including two whole modules. The PR body states the premium-runner cost plainly and offers the drop-macOS fallback, and the session replaced its speculated cost with measured numbers once the new legs reported.
• Discipline worth noting: the session declined to race the tend-mention run that was actioning the review on #428, and said so rather than pushing over it.

tend-review's already-examined guard fired correctly. Run 32458799898 (synchronize on #428) exited in 11s — HEAD ce8a2e21… already examined (tend-review/428) — skipping, because the live session 32458050254 had already folded that push into its review and stamped the head. Queued-behind-a-live-review is the designed path, not a duplicate.

tend-notifications: 2 Claude sessions, both zero-action — unchanged from every prior window

• Evidence level: Medium (waste-class)
• Occurrences this run: 2 sessions, $1.50
• Run ID: 32462632232
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32462632232
• Sessions: 32421602240 ($0.78, 14 turns), 32430189212 ($0.72, 13 turns)
• Detail: 34 scheduled runs, 32 filtered by the pre-flight gate, 2 reached Claude. Both were correct no-ops on merge notifications — 32421602240 saw #426's squash-merge commit carrying an @dormouse-bot reference from the PR body and marked it read; 32430189212 saw #422 merged with every thread entry authored by the bot or Cloudflare and applied the self-loop guard. Still waste-class per Gate 3: the only proportionate lever is the cron cadence, which the maintainer owns. Recorded, not acted on — consistent with roughly twenty prior windows.

Observation: review-runs is now the fleet's single most expensive session, and yesterday's spend produced a PR the maintainer closed

• Evidence level: Low/observational (waste-class)
• Occurrences this run: 1
• Run ID: 32462632232
• Session: 32348132575 — 36 min, $10.99, 114 turns, 15.3M cache reads
• Detail: This is the one job over 30 min in the window, and the only one on record today. Diagnosed from its session log rather than inferred: 88 Bash calls including 8 gh run download session-log pulls, then 9 Edit/7 Write calls composing #421 and the evidence entry. No stuck tool call, no long background wait — the time is deep analysis, working as designed. But one session cost more than all 17 tend-review sessions averaged, and its PR was closed on policy grounds. Net-positive overall, since the same diagnosis is what #425 records. Noted as a cost datum, not a finding; the levers (cadence, analysis depth) are the maintainer's.

Not a tend defect, but a live project-health signal: the published dependency snapshot is stale again

• Evidence level: verified on this checkout
• Detail: security-audit FAILed on 2026-08-21 (run 32448612654, reported on #417). Its first FAIL is that a clean-checkout node website/scripts/generate-deps.js adds a ws@8.21.3 entry to website/src/data/dependencies-npm.json. Confirmed directly here: ws@8.21.3 appears twice in pnpm-lock.yaml and zero times in the snapshot, so the supply-chain page published at dormouse.sh is behind the lockfile. The file has been regenerated reactively three times in six days (258ba681, a204dade, e60906c5), and the daily 05:00 audit is its only detector. No tend workflow watches security-audit — .config/tend.yaml lists only CI and Chromatic — so nothing here will pick it up automatically; #417 is the maintainer-visible channel and nightly correctly declined to re-file. The audit's second FAIL (item 16, a permissions gap) is likewise on #417 for the maintainer.

Not tend artifacts, noted so a future run doesn't chase them

• #429 — [workflow-audit] 3 unexplained change(s) on 2026-08-21. The audit scans all refs including open PR branches, so it correctly reports 370a319/ce8a2e2 (nightly's #428, unmerged) and 728ef3b (a maintainer branch). Nightly predicted this in #428's body before it happened. Working as designed.
• Neither #417, #419 nor #429 triggered a tend-triage run. Structural and not a gap: all three were opened by github-actions[bot] under the default GITHUB_TOKEN, which GitHub's recursion guard excludes from firing issues: opened. Nightly's step 5 reads open issues and covered #417 anyway.

---

Run 32561175078 — 2026-08-22T08:07:01Z (review-runs)

Window 2026-08-21T08:19:13Z → 2026-08-22T08:03Z (anchored on predecessor run 32462632232's start). 101 runs censused, 0 failures, 0 outage issues, 0 stranded triggers, 0 maintainer corrections. No gate-passing defect this window.

Clean window — two maintainer bug reports triaged into green PRs inside 32 min each, and nightly closed the stale-snapshot signal this log flagged yesterday

• Evidence level: n/a (no defect)
• Occurrences this run: 0
• Run ID: 32561175078
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32561175078

Census. 101 completed runs: 95 success, 5 skipped, 1 cancelled. By workflow: tend-mention 48, tend-notifications 36, tend-review 11, tend-triage 2, tend-ci-fix 2, tend-nightly 1, tend-review-runs 1. The skips are 3 tend-mention pre-flight gates on renovate Dependency Dashboard edits and 2 tend-ci-fix no-ops on green watched workflows. The one cancellation (32502432431, tend-review on #433) is the concurrency group superseding a queued review — run 32502652805 started 2s before it was cancelled and reviewed the same head. 17 of the 101 spun up a Claude session; the rest were pre-flight skips or guard exits.

Spend. $60.19 list-price / 379.1K output tokens / 723 turns across 17 sessions — down from $71.96 (2026-08-21) and $135.14 (2026-08-19). tend-review $27.68 over 8 sessions, tend-triage $15.68 over 2, tend-mention $6.47 over 4, tend-nightly $5.47 over 1, tend-review-runs $3.96 over 1, tend-notifications $0.93 over 1.

Timeouts. Still no timeout-minutes on any tend-* workflow, so the 360-min GitHub default applies and the near-timeout threshold is 324 min. Longest job was 31 min (tend-triage 32499634804); everything else was under 30. Zero near-timeouts. One run showed a 19-min wall clock against a 0-min job (32501179654) — that is queue time behind the per-PR concurrency group, not execution.

Cross-checks. Merged in-window: #427, #428 (both bot, both from the 2026-08-21 nightly). Opened and still open: #432, #433 (bot, from triage), #435 (bot, from nightly), plus maintainer #434. Nothing closed unmerged. Zero maintainer corrections — the repo-wide issues/pulls comment sweep over the window returned only cloudflare-workers-and-pages[bot] deploy notices and one github-actions[bot] audit report; nedtwigg posted no comment at all. Their two contributions this window were bug reports (#430, #431), which triage acted on rather than corrections of bot behavior.

tend-triage — 2 reports, 2 reproduced fixes, 2 green PRs.

• #430 → #432 (32499595589, 17 min, $4.32). Dropped paths were quoted for cmd.exe regardless of the pane's actual shell. Five new shell-escape.test.ts cases failed against main before the fix; commented the trace on #430 17 min after it was filed.
• #431 → #433 (32499634804, 31 min, $11.36, 126 turns — the window's most expensive session and its only >30-min job). Diagnosed from the session log rather than inferred: 78 Bash + 31 Edit calls doing real work — remove the predictable socket path from both hosts and the Rust side, add a nonce-verified handshake, then install the workspace, run the dor and sidecar suites, debug a genuine failure down to a stray NUL byte in proveToken, typecheck vscode-ext, cargo check, update docs/specs/dor-cli.md, and run pnpm lint:specs. No stuck tool call, no idle wait. Expensive but proportionate to a three-part security report; the resulting PR is green.

tend-nightly (32558364546, 12 min, $5.47, 76 turns) — one grounded PR that closed a recurring signal. #435 adds the missing ws@8.21.3 entry to website/src/data/dependencies-npm.json. This is the exact drift logged under Run 32462632232 as a project-health signal, and it is now fixed: vscode-ext gained the direct ws dep in #416 without regenerating the snapshot, so dormouse.sh/supply-chain under-disclosed a shipped runtime dependency for two days. The session reproduced from a clean checkout, was the sole remaining FAIL in the daily security-audit, and traced it on #417. It raised a git diff --exit-code CI guard as an observation in the PR body rather than building it unasked — correct restraint. tend check 8/8 PASS, no workflow regen diff (still 0.1.18), no conflicted bot PRs.

tend-review — the self-review-of-bot-PRs path keeps earning its cost. Per the settled upstream ruling in the overlay these are not waste, and this window is evidence for that: the reviews on #432 and #433 each raised real findings that a tend-mention repository_dispatch run then actioned as author (c7f3514, eef58bac, 2581895c), and the review on maintainer PR #434 found a genuine asymmetry — the command-exit ring got a staleness fix that the WATCHING ring did not, so dor send + dor await --until quiet can still resolve instantly on a stale completion. The review on #435 (32558730713, $1.39) posted nothing, and its log confirms that is the designed silent no-op, not an abort: it independently re-walked the dependency graph of all four product packages against the snapshot to confirm ws was the only gap, checked sort order / key shape / trailing newline, and verified no hardcoded counts before concluding there was nothing to say.

Overlay guidance is demonstrably in use. The tend-notifications session applied the running-tend rule about Chromatic UI Tests by name — it identified #432's lone non-terminal context as 1 change must be accepted as baseline, called it human-gated per the overlay, and declined to treat it as actionable. That note was written from an earlier sweep's finding, so the feedback loop closed.

tend-notifications: 1 Claude session, zero-action — unchanged from every prior window

• Evidence level: Medium (waste-class)
• Occurrences this run: 1 session, $0.93
• Run ID: 32561175078
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32561175078
• Session: 32542218418 ($0.93, 14 turns)
• Detail: 36 scheduled runs, 35 filtered by the pre-flight gate, 1 reached Claude. It was a correct no-op: both unread threads were reason: author on the bot's own PRs (#432, #433), both reviewer→author cycles had already closed out before the notification, so it verified head SHAs and check rollups and marked them read. Still waste-class per Gate 3 — the only proportionate lever is the cron cadence, which the maintainer owns. Recorded, not acted on, consistent with roughly twenty-two prior windows. The gate is doing its job: 35/36 runs never reached a session, and today's single session is the cheapest notifications spend on record.

Cost datum: yesterday's review-runs outlier did not repeat

• Evidence level: Low/observational (waste-class)
• Occurrences this run: 0 (trend note)
• Run ID: 32561175078
• Session: 32462632232 — 9 min, $3.96, 52 turns
• Detail: Run 32462632232 logged its predecessor (32348132575) at 36 min / $10.99 / 114 turns and flagged review-runs as the fleet's single most expensive session. That run itself came in at 9 min / $3.96 / 52 turns — a third the cost, with a clean-window verdict and no PR. One window is not a trend, but the outlier was the session that composed #421 and its evidence entry, not the sweep's baseline. Nothing to act on; recorded so a future sweep reading the earlier entry does not over-weight it.

Not tend artifacts, noted so a future run doesn't chase them

• #429 ([workflow-audit] 3 unexplained change(s)) and #417 ([security-audit] FAIL) are still open and remain the maintainer-visible channel. #417's dependency-snapshot FAIL is addressed by #435 above; its second FAIL (item 16, a permissions gap) is still for the maintainer. Neither fires tend-triage — both are opened by github-actions[bot] under the default GITHUB_TOKEN, which GitHub's recursion guard excludes from issues: opened. Structural, previously established, not a gap: nightly reads open issues and covered #417 anyway.
• #432's UI Tests is PENDING and will stay that way until a maintainer accepts the Chromatic baseline. Expected per the overlay; not a stuck check.

---

Run 32627143122 — 2026-08-23T08:08:26Z (review-runs)

Window 2026-08-22T08:01:27Z → 2026-08-23T08:03Z (anchored on predecessor run 32561175078's start). 77 runs censused, 0 failures, 0 cancellations, 0 outage issues, 0 stranded triggers, 0 maintainer corrections. No gate-passing defect this window.

Clean window — the tend-notifications fallback path did real work for the first time on record, and the resulting PR went through a full reviewer→author→re-review loop in 35 minutes

• Evidence level: n/a (no defect)
• Occurrences this run: 0
• Run ID: 32627143122
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32627143122

Census. 77 completed runs: 72 success, 5 skipped. By workflow: tend-notifications 45, tend-mention 21, tend-review 5, tend-ci-fix 4, tend-nightly 1, tend-review-runs 1. tend-triage and tend-weekly had no run in the window. The skips are 4 tend-ci-fix no-ops on green watched workflows (fired by the #432/#433/#435 merges) and 1 tend-mention pre-flight gate. 11 of the 77 spun up a Claude session; the other 66 were pre-flight skips or guard exits. total_count over the wider created>= fetch is 179, so the census is bounded above by it as expected and is not a page-boundary artifact.

Spend. $24.84 list-price / 154.5K output tokens / 335 turns across 11 sessions — the cheapest window on record and well down from $60.19 (2026-08-22) and $71.96 (2026-08-21). tend-review $11.11 over 5 sessions, tend-nightly $5.06 over 1, tend-notifications $3.71 over 2, tend-review-runs $2.78 over 1, tend-mention $2.18 over 2.

Timeouts. Still no timeout-minutes on any tend-* workflow, so the 360-min GitHub default applies and the near-timeout threshold is 324 min. Longest job was 15 min (tend-review 32620248465). Zero jobs over 30 min — the first window on record with none. Two runs show a wall clock well past their job time (tend-mention 32620087104 at 21 min, tend-review 32620456034 at 11 min wall / 0 min job); both are queue time behind the per-PR concurrency group, not execution.

Cross-checks. Merged in-window: #432, #433, #435 — all three bot PRs, all merged by nedtwigg within 9 minutes of each other at 16:40–16:49Z. Opened and still open: #436 (bot). Nothing closed unmerged. Zero maintainer corrections — the repo-wide issues/pulls comment sweep with --paginate over the window returned only cloudflare-workers-and-pages[bot] deploy notices and one github-actions[bot] audit report; nedtwigg posted no comment at all, and their one in-window contribution was code on #434.

The #436 loop is the window's headline, and every leg of it checks out.

• tend-notifications 32619627609 (12 min, $2.76) found [security-audit] FAIL on 2026-08-22 #417's newest comment unanswered with no dedicated workflow able to pick it up — github-actions[bot] under the default GITHUB_TOKEN can't fire tend-triage, the structural gap this log has recorded since Run 32462632232. It downloaded the audit-transcript from security-audit run 32618922852 and diagnosed it from evidence rather than inference: 21 mechanical FAIL IF checks passed, two background general-purpose subagents spawned, seven ListAgents polls, then end_turn with "I'll wait for the completion notification rather than continue polling" — which in a headless SDK run ends the session. Final result was subtype: success with zero Write uses, so neither audit-report.md nor audit-status.txt was written. Opened ci(security-audit): keep the audit in one turn so it always writes a verdict #436 and posted the diagnosis on [security-audit] FAIL on 2026-08-22 #417.
• tend-review 32619843623 ($1.96) reviewed it and found a genuine fail-open bug in the fix: the new prompt told a short run to write both files with UNVERIFIABLE markers, but audit-status.txt honors only PASS/FAIL, so a partial audit that found nothing wrong would write PASS — closing every open security-audit-failure issue and opening release.yml's publish-vscode gate. That flips an incomplete audit from fail-closed to fail-open. Posted as an inline suggestion, no approval.
• tend-mention 32620087104 (21 min, $1.61) actioned it as author, applied the suggestion verbatim as 7c30ca8, and also caught that the SECURITY.md paragraph the same PR adds asserted the now-withdrawn contract directly beneath the line saying an audit reaching no verdict must not let the gate pass.
• tend-review 32620248465 ($2.26) re-reviewed, traced the corrected contract end to end on the merged tree (no status file → MISSING → exit 1 → gh run watch --exit-status fails → publish-vscode blocked via needs:), and pushed one prose fix as 4d37f0c.
• tend-mention 32620445974 ($0.57) fired on that second review and correctly exited silently — it verified the review had no inline comments and that its single finding had been self-executed 19 seconds earlier as the current head. Textbook self-loop guard.
• tend-review 32620456034 queued 11 min behind the live session and exited with a 0-min job on the already-examined stamp. Designed path, not a duplicate.

tend-review on maintainer PR #434 — approval earned, not assumed. 32586560093 ($3.11) held approval on a real finding: the new consumeAwaitableRing guard used detector.getStatus() === 'NOTHING_TO_SHOW' as a proxy for "no output since the ring latched", but QuiesceDetector stays in that state for busyCandidateGap (1500ms) after output resumes, so the exact repro from the earlier thread still resolved quiet with waitedMs: 0. It ran a probe test on the branch to demonstrate it rather than reasoning about it, and noted the new test passes only because driveToBusy advances 1600ms between onData calls — so the test asserts the post-gap state its name doesn't describe. 32587340940 ($2.39) then approved at 6d0c83c after measuring the fix (reverted the gate line, confirmed 2 of the 3 new cases genuinely fail against the old predicate) and correctly reported that the commit message's "all three fail" claim overstates it by one. It also applied the running-tend Chromatic note by name — stopped polling UI Tests as human-gated rather than burning the cap. Both overlay rules that came out of earlier sweeps are demonstrably in use.

tend-nightly had one below-PR-bar finding and no channel to put it in, so it evaporated into the session log

• Evidence level: Medium (waste-class)
• Occurrences this run: 1
• Run ID: 32627143122
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32627143122
• Session: 32624406418 (9 min, $5.06, 46 turns — 39 Bash, 2 Read, 2 Skill, zero Write/Edit/gh … create)
• Detail: The survey (bucket 24/28, 33 files) produced one concrete finding — lib/src/components/ThemeDebugger.tsx:218, if (!open && dialog.open) dialog.close(); is unreachable because if (!open) return null; on line 226 unmounts the <dialog> and nulls dialogRef.current before the effect can run with open === false. I verified the claim directly on this checkout; it is correct. Nightly declined to ship it, for stated and defensible reasons: it's devtools-only, the repair's correctness turns on ref-detach-vs-cleanup ordering it can't observe headlessly, and jsdom has no HTMLDialogElement.showModal to pin a regression test with — so it fails triage's reproduction gate. But the nightly skill's step 8 offers exactly two outlets (PR, or an issue when there's no obvious fix) and this took neither, so the finding exists only in a 30-day-retention artifact nobody reads. Filing an issue would also have been imperfect — it hands the analysis to a tend-triage run that hits the same wall. Not acted on: 1 occurrence against a Medium bar (5+), and any fix is a new-channel structural change (very high bar) for a waste-class cost. Recorded so a future sweep can count occurrence 2. Prior windows where nightly shipped nothing (Runs 30996725754, 31684173139) had zero findings, so this is the first instance of a finding being dropped rather than absent.

tend-notifications: 2 Claude sessions — and for the first time on record, one of them was high-value

• Evidence level: Medium (waste-class), with a correction to the prior characterization
• Occurrences this run: 2 sessions, $3.71
• Run ID: 32627143122
• Sessions: 32619627609 ($2.76, 36 turns — acted), 32586349731 ($0.95, 13 turns — correct no-op)
• Detail: 45 scheduled runs, 43 filtered by the pre-flight gate, 2 reached Claude. 32586349731 was the usual zero-action pass (both threads closed-as-completed by nedtwigg with the bot's fixes already merged; verified and marked read). 32619627609 was not — it is the run that diagnosed the security-audit INCONCLUSIVE and opened ci(security-audit): keep the audit in one turn so it always writes a verdict #436, work no other workflow could have reached. Roughly twenty-three prior windows in this log characterize tend-notifications as uniformly zero-action waste whose only lever is the cron cadence; that framing now has a counterexample, and the counterexample is precisely the fallback role the workflow exists for (a github-actions[bot] issue that GitHub's recursion guard keeps tend-triage off). A future sweep weighing a cadence reduction should price that against the $3.71 rather than treating the whole line as waste.

Not tend artifacts, noted so a future run doesn't chase them

• The repo's own security-audit workflow has failed 12 of its last 15 daily runs (success only on 2026-08-17 and 2026-08-19). The newest failure mode — ending the turn while background subagents run, so no verdict is ever written — is what ci(security-audit): keep the audit in one turn so it always writes a verdict #436 fixes, and it is open with every check green awaiting the maintainer. Nightly flagged the timing consequence explicitly: the 04:52 audit runs the current prompt unless ci(security-audit): keep the audit in one turn so it always writes a verdict #436 merges first, so tomorrow's run is expected to repeat. Not a tend defect and already in the maintainer's channel; do not re-file.
• [workflow-audit] 3 unexplained change(s) on 2026-08-23 #437 — [workflow-audit] 3 unexplained change(s) on 2026-08-23. All three commits (e05429c, 7c30ca8, 4d37f0c) are ci(security-audit): keep the audit in one turn so it always writes a verdict #436's, on an unmerged PR branch. The audit scans all refs including open PR branches, so this is working as designed — same shape as [workflow-audit] 3 unexplained change(s) on 2026-08-21 #429, previously established. It opened at 07:43Z, after the 07:00Z nightly, so nothing could have covered it in-window; the next nightly reads open issues.
• [security-audit] FAIL on 2026-08-22 #417 remains the maintainer-visible channel for security-audit. Its actual findings are now clear: the ws@8.21.3 snapshot FAIL merged as fix(website): disclose ws in the npm supply-chain snapshot #435, and the 2026-08-21 tend-mention.yaml permissions FAIL was regraded PASS by the 2026-08-22 audit. The next audit that reaches a verdict should auto-close it.
• Neither [security-audit] FAIL on 2026-08-22 #417 nor [workflow-audit] 3 unexplained change(s) on 2026-08-23 #437 fires tend-triage — both are opened by github-actions[bot] under the default GITHUB_TOKEN, which GitHub's recursion guard excludes from issues: opened. Structural, previously established, and this window shows the tend-notifications fallback covering it rather than it going unhandled.

---

Run 32706405380 — 2026-08-24T08:45:00Z (review-runs)

Window 2026-08-23T08:01:32Z → 2026-08-24T08:29Z (25 h, anchored on predecessor 32627143122). 67 tend runs — 42 tend-notifications (41 success / 1 failure), 17 tend-mention (13 success / 4 skipped-at-trigger), 4 tend-review, 1 tend-nightly, 1 tend-weekly, 1 tend-triage (skipped), plus yesterday's review-runs. 9 runs opened a Claude session, 257 turns, $17.80 list price: nightly $7.00 / 80 turns, yesterday's review-runs $3.28 / 74, review-on-#439 $2.38 / 37, review-on-#440 $1.51 / 29, mention $1.21 / 29, review-on-#441 $0.99 / 15, review-on-#439-round-2 $0.78 / 18, weekly $0.65 / 7, the failed notifications run $0. No long-running (>30 min) or near-timeout jobs — longest was nightly at 18 min against the default 360-min cap (no timeout-minutes override in .config/tend.yaml, no non-tend-* workflow uses the action). 0 bot PRs merged or closed, 0 maintainer corrections (the only non-bot comments in the window were github-actions audit rows and Cloudflare Pages deploy tables — verified against both the issues and pulls comment endpoints with --paginate).

GATE-PASSING: none. Three findings recorded below, all occurrence 1 and all stochastic, so all sit under the 5+ floor.

tend-review on #440 computed a full findings report and ended the turn without posting it — first confirmed review abort in this log

• Evidence level: Critical outcome, occurrence 1
• Occurrences this run: 1 (historical: 0 — code-review appears nowhere in this tracker's 2026-05→08 entries, and the one prior use of the word "abort" is Run 32627143122 explicitly ruling out an abort on #435)
• Run ID: 32706405380
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32706405380
• Session: 32704496682 (tend-review on #440, renovate github-actions pin bump, $1.51 / 29 turns / 4 m 16 s), artifact claude-session-logs-n440
• Classification: stochastic — 32701796482 (review on refactor(standalone): stop the browser harness claiming native file-drop support #439, same window) ran the identical skill sequence (running-in-ci → running-tend → code-review), wrote a one-line bridge ("Code-review pass complete… one minor finding survived"), and went on to POST a review plus the tend-review/439 stamp. Replayed 10×, this fails only sometimes.
• Cost class: wrong outward action (false-green verdict), partially self-correcting — see below
• Detail: The session did a genuinely good peripheral review — dereferenced both new SHAs against their annotated tags, diffed every with: input against both new action.ymls, confirmed one pin site per action with no stale digests, checked SECURITY.md:30's pin rule, and reasoned about workflow-audit.yaml's classifier. It then invoked /tend-ci-runner:code-review, which — with subagents disallowed by this repo's CI prompt ("this session's prompt limits subagent use to user-requested calls") — ran inline in the same session. The session rendered code-review's Output contract ("Return the findings to the caller… Tell the caller which mode ran") as its terminal assistant message: three ranked PLAUSIBLE findings plus a "Verified clean — don't re-derive" block, ~4 KB of analysis. Then end_turn. Zero posting tool calls in the whole log — jq over every tool_use for pr review / pr comment / pulls/440 returns nothing. Step 5 (Submit) and step 9 (Stamp examined HEADs) never ran.
• Why this is a false green, not just waste: chore(deps): update github-actions #440 today shows reviews: 0 and an empty commit-status list — no tend-review/440 context at all — while the review check itself passed green (the job succeeded; only the agent stopped early). Both this repo's overlay ("a silent exit means the review ran and found nothing to post") and the bundled review skill ("silence means correct") teach the reader that an unreviewed-looking tend PR was reviewed and cleared. An abort is therefore indistinguishable from an all-clear. The lead finding it swallowed was substantive: chromaui/action v18.1.0→v18.5.0 spans "Resolve ctx.storybook paths to absolute paths" and "Improve Storybook main config parsing", and chromatic.yml runs onlyChanged: true under workingDir: lib with no storybookBaseDir while lib/.storybook/main.ts's viteFinal aliases three packages above lib/ — so TurboSnap under-selection would silently stop catching visual regressions, and exitZeroOnChanges: true means it never reds a build.
• Mitigations that limit the damage: the missing tend-review/440 stamp means the workflow's gate step will not skip a future synchronize, so a renovate rebase re-triggers a full review. The exposure is only "chore(deps): update github-actions #440 merges before it is next pushed".
• Gates: Gate 1 FAILS — stochastic floor is 5+ and review-gates.md is explicit that no evidence level, reclassification, or cost class lowers it. Gate 2: the proportionate fix is one sentence in the bundled code-review skill's ## Output section telling it that when it runs inline it is a sub-step, not a terminal deliverable — a targeted fix, and an upstream one (generic, would fire in every consumer; nothing matching exists in max-sixty/tend — nearest neighbours are the closed tend#817 and tend#830). Gate 3 would pass on cost class. Not filed. Watch for occurrence 2.
• Cheap check for a future run: a tend-review run that produced no review body and left no tend-review/<n> commit status is the abort signature; a designed silent no-op stamps the HEAD. gh api repos/diffplug/dormouse/commits/<head>/status --jq '[.statuses[].context]'.

tend-nightly mis-projected the live review on its own PR as "posts nothing" — right call, wrong stated reason

• Evidence level: Low–Medium, occurrence 1
• Occurrences this run: 1
• Run ID: 32706405380
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32706405380
• Session: 32701038269 (tend-nightly, $7.00 / 80 turns / 18 min)
• Classification: stochastic (a per-session --jq projection choice)
• Cost class: wasted compute / accuracy slip — the wrong claim landed in the session's step summary only, never on a public thread
• Detail: Nightly opened refactor(standalone): stop the browser harness claiming native file-drop support #439, polled its checks, then checked for a review with gh pr view 439 --json state,reviews,comments --jq '{state, reviews: [.reviews[]? | {a:.author.login, s:.state}], …}' — projecting away .body. It saw {a: dormouse-bot, s: COMMENTED} and concluded "a self-review on a bot PR that posts nothing found nothing to post; I left it unactioned." The review's body was 707 characters and carried an inline suggestion on standalone/src/browser-sidecar-adapter.test.ts (missing vi.mock stubs for the three Tauri SDK modules). This is the failure running-in-ci warns about under "A review's inline comments are a separate fetch", one field earlier — the body itself was dropped, not just the inline comments.
• Why it cost nothing: nightly's action was correct regardless — running-in-ci says a review that lands while you poll is not yours to action, and tend-mention 32702271041 picked it up 10 s after it was posted and shipped the fix as 76cc2770. Only the stated justification was false.
• Note for a future run: this repo's overlay paragraph "Settled upstream rulings" describes bot-PR self-reviews that exit silently. Nightly appears to have applied that frame to a review that was not silent. If this recurs, the candidate fix is a scoping clause in that overlay paragraph — but that is a new-paragraph change (Gate 2: 3+ occurrences), so it is not warranted at n=1.
• Gates: Gate 1 FAILS (stochastic, 1 occurrence). Not filed.

tend-review on #441 skipped the mandatory code-review pass entirely

• Evidence level: Medium, occurrence 1
• Occurrences this run: 1
• Run ID: 32706405380
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32706405380
• Session: 32704548722 (tend-review on #441, hono 4.13.2→4.13.3 lockfile bump, $0.99 / 15 turns)
• Classification: stochastic (instruction skip under a trivial diff)
• Cost class: wasted assurance, not a wrong outward action — the manual pass was thorough and its approval looks correct
• Detail: The review skill's step 4 is unambiguous — "Every review that reaches this step runs a /tend-ci-runner:code-review pass over the PR's merged tree — trivial diffs included; the depth-scaling above sets how deep the pass goes, never whether it happens." The session's Skill calls were tend-ci-runner:running-in-ci and running-tend only. It then approved. The manual work it did instead was solid (only-declaration check against server/package.json's ^4.6.0, all nine hono lockfile entries moved together, no stale 4.13.2, the residual 4.13.3 matches identified as @types/node@24.13.3, renovate/stability-days confirmed passed, Enroll a remote Host from Settings, not the devtools console #423 ruled out as an overlap), so nothing is known to have been missed. Interesting as the mirror image of the chore(deps): update github-actions #440 finding: one session let code-review end it, another skipped code-review altogether — the two review sessions that ran it inline are the two that behaved oddly.
• Gates: Gate 1 FAILS (stochastic, 1 occurrence). Not filed.

The one run failure was a clean 529, and the outage loop worked end to end

32694029645 (tend-notifications, 05:34, 6 min) had 1 processable notification(s) — proceeding, then died before the session started — the entire JSONL is one synthetic assistant message, API Error: 529 Overloaded. An outage issue #438 was filed, and nightly drained and closed it at 07:23 with the enrichment comment naming the 529 (correctly declining to re-run a scheduled workflow, and correctly leaving the four [workflow-audit] issues alone as human-accountable records). By the 06:23 tick the notification queue was already at zero, so nothing stranded. This is the tend#1019/tend#1021 hand-off working as designed — worth recording as a positive, since prior windows (Run 32115440555) recorded the opposite shape.

Healthy signals worth keeping in the record

• The reviewer→author→re-review loop on refactor(standalone): stop the browser harness claiming native file-drop support #439 closed in 9 minutes. Nightly opened it at 07:30:20 with a real finding (BrowserSidecarAdapter.onFilesDropped was implemented but never dispatched — and because PlatformAdapter documents the member as a capability probe that consumers test with ?., a plain browser tab was answering "drag-drop supported" and then staying silent forever), verified failing-before/passing-after in both directions. tend-review found the new test pulled the real Tauri SDK where the sibling test stubs it; tend-mention applied 76cc2770; the re-review confirmed and stamped. All gating checks green.
• tend-weekly 32631669574: correct no-op — 0 dependency PRs open at the time, no repo-specific weekly tasks defined in the overlay, nothing posted.
• tend-notifications: 41 of 42 runs exited at the unread-count gate in 6–9 s for ~$0. The pre-agent layers (B/C/D) continue to carry essentially the whole load.
• tend-mention: 4 of 17 skipped at the trigger gate; the one substantive run was the refactor(standalone): stop the browser harness claiming native file-drop support #439 author leg above. No undirected-bot-comment no-ops observed this window — consistent with tend#606 having been fixed upstream.

Not tend artifacts, noted so a future run doesn't chase them

cloudflare-workers-and-pages deploy tables on #439/#440/#441 and the github-actions "Audit passed" row on #417 are the only non-bot comments the Step 4 sweep returns. Neither is a maintainer correction.

[dormouse-bot]

---

Run 32826269519 — 2026-08-25T08:29:57Z (review-runs)

Window 2026-08-24T08:27:26Z → 2026-08-25T08:23Z (24 h, anchored on predecessor 32706405380). 70 tend runs — 34 tend-notifications (all success), 24 tend-mention (20 success / 4 skipped-at-trigger), 5 tend-review (all success), 5 tend-ci-fix (all skipped — main stayed green), 1 tend-nightly, plus yesterday's review-runs. 10 runs opened a Claude session, 354 turns, $23.90 list price: nightly $5.77 / 83 turns, yesterday's review-runs $5.40 / 71, review-on-#434 $3.38 / 40, review-on-#444-round-1 $2.14 / 35, mention (#444 author leg) $1.96 / 40, review-on-#442 $1.80 / 23, review-on-#444-round-2 $1.64 / 28, review-on-#443 $0.90 / 14, two notifications runs $0.55 / 9 and $0.36 / 11. No long-running (>30 min) or near-timeout jobs — longest was nightly at 15 m 44 s against the default 360-min cap (no timeout-minutes override in .config/tend.yaml; release.yml, security-audit.yaml, and workflow-audit.yaml only mention .config/tend.yaml, none uses the tend action, so the tend- census prefix is still complete). 1 bot PR closed (#439, merged), 0 maintainer corrections (both the issues and pulls comment endpoints, --paginate, return only Cloudflare Pages deploy tables). No open tend-outage issue — nothing stranded to drain.

GATE-PASSING: none. One finding recorded below — occurrence 2 of a stochastic pattern, still under the 5+ floor.

tend-review on #443 aborted after the inline code-review pass — occurrence 2 of the pattern first recorded on #440

• Evidence level: Critical outcome, occurrence 2 (this run 1; historical 1 — chore(deps): update github-actions #440 under Run 32706405380)
• Run ID: 32826269519
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32826269519
• Session: 32757622387 (tend-review on #443, renovate pnpm 11.17.0→11.22.0 packageManager bump, $0.90 / 14 turns / 2 m 53 s), artifact claude-session-logs-n443
• Classification: stochastic, and this window sharpens the evidence rather than weakening it — 4 of the 5 tend-review sessions in the window loaded /tend-ci-runner:code-review inline (subagents are disallowed by this repo's CI prompt, so it never runs as a subagent here) and went on to post and stamp: chore(deps): update storybook to v10.5.10 #442 wrote the one-line bridge "Code-review pass returned no surviving findings. Posting the approval." and posted APPROVED + tend-review/442; Add dor await: block on a Surface and report why the wait ended #434 and both fix(remote): reclaim abandoned host challenges instead of retaining them forever #444 rounds likewise. Only chore(deps): update pnpm to v11.22.0 #443 stopped. Replayed 10×, this fails only sometimes.
• Cost class: wrong outward action (false green)
• Detail: The session ran a competent peripheral review — confirmed the diff is one line (package.json:5), grepped every workflow/script/doc for a second pnpm@ pin and found none, established that all workflows use pnpm/action-setup@v6 with no version: input so they resolve from packageManager, and checked pnpm-workspace.yaml/pnpm-lock.yaml for anything a minor bump would invalidate. It then invoked code-review, ran the short angle set inline, and emitted code-review's Output contract as its terminal assistant message ("Ran the short angle set inline (A–C + cleanup/conventions)… No findings survive verification." plus the per-angle checklist). Then end_turn. Zero posting tool calls in the log — the last tool_use is a cat package.json from inside the code-review pass. Review skill step 5 (Submit) and step 9 (Stamp) never ran.
• Yesterday's cheap check fired exactly as designed. gh api repos/diffplug/dormouse/commits/37fd683f/status returns [renovate/stability-days] — no tend-review/443 context — while gh pr view 443 --json reviews returns []. Compare chore(deps): update storybook to v10.5.10 #442 in the same window: [renovate/stability-days, tend-review/442] plus an APPROVED review. The signature separates abort from designed silent no-op with one API call; keep using it.
• Not a merge race. chore(deps): update pnpm to v11.22.0 #443 merged at 17:39:53Z; the session's last JSONL line is 17:39:32Z and the job completed 17:39:37Z. The abort preceded the merge by 21 s, and the session never re-read PR state after its opening gh pr view.
• Blast radius this time: small. The bumped line is a packageManager pin the code-review pass had already cleared, and nedtwigg merged it 21 s later — so no defect shipped. The durable cost is the same as chore(deps): update github-actions #440's: the review check itself is green (the job succeeded; only the agent stopped early), so an aborted review is externally indistinguishable from a clean one. Both aborts so far landed on renovate PRs where the diff was trivial; neither has yet hit a substantive diff.
• Gates: Gate 1 FAILS — the stochastic floor is 5+, and review-gates.md is explicit that no evidence level, cost class, reclassification, or pre-registered escalation lowers it. We are at 2. Gate 2: the proportionate fix remains one sentence in the bundled code-review skill's ## Output section saying that when it runs inline it is a sub-step and not a terminal deliverable — upstream, generic, would fire in every consumer. Dedup re-checked this run: nothing matching exists in max-sixty/tend (nearest are the closed tend#817 "/code-review second pass is unreachable" and tend#846; the open tend#1052 is unrelated). Gate 3 would pass on cost class. Not filed. Watch for occurrence 3.
• Note for a future run: yesterday also recorded the mirror-image slip — chore(deps): update dependency hono to v4.13.3 #441's review skipped code-review entirely (occurrence 1). This window shows no recurrence of that: all 5 review sessions loaded it. If they eventually converge, the shared root is the same inline-execution ambiguity and they should be counted as one pattern; until then keep them separate.

Healthy signals worth keeping in the record

• The fix(remote): reclaim abandoned host challenges instead of retaining them forever #444 reviewer→author→re-review loop closed in 4 minutes, with a real defect caught at every stage. Nightly 32820092381 found that HostChallengeIssuer.pruneExpired() existed and was unit-tested but had no production caller, and that POST /api/signin/begin is the one challenge-minting route with no auth in front of it — so an unauthenticated client could grow a self-hosted server's heap by asking. It verified the regression test fails on main and passes with the fix, ran server-lib-common (113/113) and the consuming server (143/143) locally, updated remote-security-model.md, ran spec-lint, opened #444, and polled to green. tend-review 32820662396 then raised three findings including that the spec sentence overclaimed (the sweep bounds by mint-rate × TTL, not absolutely). tend-mention 32821130875 actioned all three in f47cd852, replied to both inline comments, posted a summary, and updated the PR description — which had repeated the now-corrected absolute-bound claim. Round-2 review 32821404911 stamped tend-review/444 at the new head. All checks green, MERGEABLE.
• The two empty-bodied COMMENTED reviews on fix(remote): reclaim abandoned host challenges instead of retaining them forever #444 are inline-reply shells, not stray reviews. GitHub wraps a standalone inline reply in a review object; both carry body: "" and each holds one in_reply_to_id comment. They fired 2 pull_request_review + 2 repository_dispatch tend-mention runs at 07:24, none of which opened a Claude session — the pre-agent gates absorbed all four. Consistent with tend#606 being fixed upstream. Worth recording so a future run doesn't read them as duplicate reviews.
• tend-notifications: 32 of 34 runs exited at the unread-count gate for ~$0; the two that opened sessions were both correct read-only no-ops (chore(deps): update dependency hono to v4.13.3 #441 and chore(deps): update storybook to v10.5.10 #442, each bumped by nedtwigg's merge rather than by a human ask, both marked read with nothing posted).
• tend-mention: 24 runs, 4 skipped at the trigger gate, 19 more absorbed by the pre-agent layers, exactly 1 substantive session (the fix(remote): reclaim abandoned host challenges instead of retaining them forever #444 author leg above). No undirected-bot-comment no-ops.
• tend-ci-fix: 5 runs, all skipped — main never went red.
• Nightly's restraint held. It surveyed 28 of the day's 32 changed files, reviewed 11 commits, checked 17 open issues, confirmed the tend workflows already regenerate byte-identical at 0.1.18, and shipped exactly one atomic PR. It also explicitly declined to file a MobileTerminalUi.tsx observation (a cursorPointerIdRef leak on mid-drag touch-mode switch) after establishing it isn't reachable in practice — correct application of the reproduce-before-filing bar.

Not tend artifacts, noted so a future run doesn't chase them

• The repo's own security-audit workflow has now succeeded twice running (32692093970 on 08-24, 32810700133 on 08-25), reversing a 12-of-15 failure streak, and #417 auto-closed on 08-24T05:09Z as the prior window predicted. #436 — the one-turn fix — is still open and green, so the recovery is not yet attributable to it. Not a tend defect; already in the maintainer's channel; do not re-file.
• [workflow-audit] 3 unexplained change(s) on 2026-08-23 #437 / [workflow-audit] 3 unexplained change(s) on 2026-08-21 #429 / [workflow-audit] 1 unexplained change(s) on 2026-08-20 #419 / [workflow-audit] 3 unexplained change(s) on 2026-08-19 #412 ([workflow-audit] N unexplained change(s)) remain open and are github-actions[bot]-authored under the default GITHUB_TOKEN, which GitHub's recursion guard excludes from issues: opened — so tend-triage structurally cannot fire on them. Previously established; nightly correctly left them as human-accountable records.

---

Run 32947531346 — 2026-08-26T08:30:18Z (review-runs)

GATE-PASSING: none. Zero PRs, zero issues. The abort pattern did not reach occurrence 3 — every review head in the window is stamped.

Window: 2026-08-25T08:22:27Z → 2026-08-26T08:25:06Z (anchored on predecessor 32826269519's start). 110 tend runs — 100 success, 10 skipped, 0 failures, no tend-outage issue, nothing stranded. 22 opened a Claude session; $54.47 at API list prices. Longest job 16 m 14 s (tend-notifications 32917689829) against a 360-min default cap, so no long-running or near-timeout flags.

Workflow	Runs	Sessions	Cost
tend-review	12	12	$28.70
tend-mention	54 (2 skipped)	5	$11.37
tend-nightly	1	1	$6.01
tend-notifications	34	3	$5.34
tend-review-runs	1	1	$3.05
tend-ci-fix	8	0	$0 (all skipped; main CI never red)

The two-run abort pattern did not recur — occurrence count stays at 2

• Evidence level: n/a (negative result, recorded so the count is auditable)
• Occurrences this run: 0 (historical: 2 — #440 under Run 32706405380, #443 under Run 32826269519)
• Run ID: 32947531346
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32947531346
• Detail: Ran the cheap check both prior entries left behind (gh api repos/diffplug/dormouse/commits/<head>/status) against every PR that drew a tend-review session in this window. All five carry their stamp: #434 → tend-review/434, #423 → tend-review/423, #424 → tend-review/424, #445 → tend-review/445, #446 → tend-review/446. The one review session that posted no body — 32921930318 on #446 round 6, $0.67 / 14 turns — did stamp fa5bf1dd, which is the designed silent no-op, not an abort. The signature separates the two with one API call exactly as #443's entry claimed; keep using it. Stochastic 5+ floor still governs; we remain at 2.
• Also not recurring: the mirror-image slip from #441 (review skipped code-review entirely, occurrence 1). Not re-checked exhaustively this window, but no review produced the trivial-diff shape it appeared under — every PR reviewed here carried a substantive diff.

Cost datum: #446 drew 6 self-review rounds for $23.40 — occurrence 2 of the designed path, at 2.1× #399's magnitude

• Evidence level: observation, not a finding — already ruled designed
• Occurrences this run: 1 (historical: 1 — #399, Run 32231741521's entry "Observation, not a finding: ci(security-audit): distinguish an inconclusive audit from a FAIL #399's four self-review rounds are the designed path, not a loop")
• Run ID: 32947531346
• Workflow: https://github.com/diffplug/dormouse/actions/runs/32947531346
• Sessions: reviews 32918279502 $2.15 · 32918970771 $2.25 · 32919701795 $2.36 · 32920526419 $2.30 · 32921287819 $2.30 · 32921930318 $0.67; mention legs 32918638306 $3.56 · 32919344499 $1.69 · 32920130841 $2.51 · 32920934339 $2.36 · 32921663861 $1.25
• Classification: structural (the dispatch fires on every review of a bot-authored PR by design) — but already adjudicated: this is the reviewer-stops / dispatched-author-applies split upstream max-sixty/tend#1007 landed, and the #399 entry recorded that dormouse's runs already behave that way. Do not re-file.
• Cost class: wasted compute — nothing wrong reached the public record. No approval was issued (self-approval is impossible on a bot PR), all six rounds are COMMENTED, and the branch is green.
• Detail: #446 opened 01:15:14Z, converged 02:16:13Z — 61 minutes, 11 sessions, 6 commits, $23.40 = 43% of the day's total spend on one PR. I read all five non-empty review bodies to check whether the rounds were churning or converging, and they converge: round 1 found the MISSING_FRAGMENTS issue body was written for exactly one missing domain; round 2 found the 25-minute deadline the PR was justifying still could not fire (the DEADLINE is recomputed in each re-issued shell); round 3 found the newly-reachable give-up path lands as FAIL rather than INCONCLUSIVE, publishing a timeout as a security finding; round 4 found §3 still forces a binary Summary, so the INCONCLUSIVE issue would contradict its own title; round 5 found scripts/security-audit-local.sh still printing the pre-reroute CI contract. Each round opens by verifying the previous round's fixes trace clean, and each finding is distinct and real. Round 6 found nothing and stamped.
• Trend worth carrying: #399 was 4 rounds / 27 min / $10.89; #446 is 6 rounds / 61 min / $23.40. Both converged, both on bot-authored audit-machinery PRs. Gate 3 forbids building mechanism against waste, but it explicitly permits escalating the number once the aggregate matters. If a third instance lands above ~$20, that is the point to put the three-instance total in front of nedtwigg — the lever is his (cadence, or a round cap), not a new gate the fleet has to carry forever.

Healthy signals worth keeping in the record

• The tend-notifications fallback did real engineering for the second time on record, and it was the best session of the window. 32917689829 ($3.96 / 52 turns / 16 min) picked up the #445 merge notification, read what the review had raised and the merge had left unfixed, then tested the workflow as shipped rather than by inspection: it parsed security-audit.yaml with python3 -c "import yaml", extracted the run: blocks to /tmp/step<N>.sh, and executed the redactor and the reporting-step guard in a mktemp -d with fake AUDIT_PAT / CLAUDE_CODE_OAUTH_TOKEN values to confirm the leak paths. It opened #446, polled it to green, and marked both threads read. The prior window (Run 32826269519) recorded the first such instance; this is the second, and it is a bigger one.
• tend-review on #445 caught a red build the maintainer fixed in four minutes. 32910357816 ($5.48 / 57 turns) opened with Build & Test red on dd4b075: the Pocket fingerprint JSX had landed in SetupOrSignin instead of HostsView (TS2304 ×2 + TS6133), and the PR's own docs/specs/pocket-app.md named the right home. It anchored the claim on gh run view 32910358496 --log-failed inside a <details>, and nedtwigg shipped e2c9bfb3 "fix(pocket): put the device fingerprint back in HostsView" at 23:33:26Z. Three further findings (the audit's FAIL IF lines asserting properties security-audit.yaml doesn't have, MAX_PENDING_TICKETS not bounding #clients/#pairings, boundedPairingAccount rewriting an identity later compared for equality) were left for follow-up — and the notifications session above turned the first cluster into #446. Clean handoff across three workflows.
• tend-nightly followed this repo's two overlay rules exactly. 32941305021 ($6.01 / 66 turns) verified both security-audit FAILs against a clean uvx tend@latest init worktree (byte-identical, so not local drift), then asked nedtwigg for permission to file the mention.yaml.j2 verify-job template gap upstream — without the standing-exception offer the overlay forbids, and addressing the maintainer by the correct handle rather than the ntwigg typo. Its #447 comment SHA-pins its one #L link to 4b5272b5. It also declined to file a website/scripts/generate-deps.js observation after establishing f46ead9e had scoped itself deliberately — correct restraint.
• Both cheap tend-notifications sessions were correct read-only no-ops. 32908313697 ($0.59) on #434 and 32912350897 ($0.79) on #423 each established the unread state was the merge rather than a new message (latest_comment_url pointing at the PR itself), confirmed every review and inline comment on the thread was bot-authored, and marked read without posting.
• Pre-agent gating held. tend-mention: 54 runs (24 issue_comment, 14 pull_request_review, 14 repository_dispatch, 2 issues skipped) → 5 sessions, all five the #446 author legs. The 9 unconsumed dispatches correspond to the four empty-bodied inline-reply review shells on #446 and their siblings — the tend#606 fix continuing to work. tend-ci-fix: 8 runs, all skipped.
• Cross-checks: 0 maintainer corrections. Verified, not assumed — both repo-wide comment endpoints were paginated over the window and every non-bot comment returned is cloudflare-workers-and-pages deploy output. Dispositions: bot PRs #444 and #436 both merged; maintainer PRs #445, #423, #434 merged with bot findings applied or already answered; #424 approved and open; #446 open, green, awaiting a maintainer. Nothing the bot did was undone, relabeled, or reverted.

Not tend artifacts, noted so a future run doesn't chase them

• security-audit is red on main again — #447 [security-audit] FAIL on 2026-08-26, filed 05:06:37Z. Both FAILs are real and nightly verified them: default_workflow_permissions: write + can_approve_pull_request_reviews: true is an admin-only repo setting and flipping it to read-only clears both, and the verify job in tend-mention.yaml is the only job of the eight generated workflows with no permissions: block (an upstream template gap, no live exposure — the one step overrides GITHUB_TOKEN with TEND_BOT_TOKEN). This is the same finding as the July #272, which closed by the audit's own lifecycle without the template changing. Already in the maintainer's channel with a permission request pending; do not re-file, and do not file upstream until nedtwigg answers on #447.
• #446 is the only open bot PR. The self-initiated-PR budget in running-in-ci (hold at 5+) is nowhere near binding.

---

Run 33103321834 — 2026-08-27T18:37:34Z (review-runs)

Window 2026-08-26T08:23:45Z → 2026-08-27T18:23Z (35 h, anchored on predecessor 32947531346). 178 tend runs censused — 130 tend-mention (123 success / 5 cancelled / 2 skipped), 24 tend-review (23 success / 1 cancelled), 14 tend-notifications (all success), 8 tend-ci-fix (all skipped — main stayed green), 1 tend-nightly, plus yesterday's review-runs. 46 runs opened a Claude session, 1220 turns, $89.74 list price: tend-review 21 sessions / 584 turns / $43.33, tend-mention 20 / 422 / $27.15, tend-notifications 3 / 103 / $7.98, tend-nightly 1 / 70 / $7.50, yesterday's review-runs 1 / 41 / $3.78. No long-running (>30 min) or near-timeout jobs — every job in the window was scanned and zero reached 30 min; longest was nightly at 16 min against the default 360-min cap (no timeout-minutes override in .config/tend.yaml; release.yml, security-audit.yaml, workflow-audit.yaml, ci.yml, chromatic.yml were re-checked and none uses the tend action, so the tend- census prefix is still complete). 0 run failures. No open tend-outage issue — nothing stranded to drain. 2 bot PRs merged (#446, #454), 3 bot PRs closed as superseded (#449, #450, #451), 1 maintainer correction (below).

GATE-PASSING: 1 — tend-review skipped the mandated code-review pass in 5 of 16 eligible sessions, reaching occurrence 6 and meeting the condition upstream tend#937 pre-registered. Filed as #457 (permission ask to file at max-sixty/tend). Two further findings recorded below, both under threshold.

tend-review posts a review verdict without running the mandatory /tend-ci-runner:code-review pass — occurrence 6

• Evidence level: High pattern, occurrence 6 (5 this run + 1 historical)
• Occurrences this run: 5
• Run ID: 33103321834
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33103321834
• Sessions: 33041167812 (Close the audit's findings on itself, and run application-security on Opus #452, COMMENTED 5055 chars), 33051245140 (Close the four security-audit findings: paste injection, dev-bridge RCE, proxy origin laundering, OSC forgery #453, APPROVED 1805 chars), 33053278119 (Close the four security-audit findings: paste injection, dev-bridge RCE, proxy origin laundering, OSC forgery #453, COMMENTED + inline suggestion), 33053731411 (Close the four security-audit findings: paste injection, dev-bridge RCE, proxy origin laundering, OSC forgery #453, APPROVED 648 chars), 33053539061 (ci(standalone): give the smoketest's apt step a budget that outlasts its retries #454, COMMENTED 1219 chars — a first review, so no incremental existed and step 1's trivial-skip path could not apply). Historical: 32704548722 (chore(deps): update dependency hono to v4.13.3 #441), recorded under Run 32706405380 above.
• Classification: stochastic — mixed outcomes under identical conditions, including within one PR: on Close the audit's findings on itself, and run application-security on Opus #452 rounds 4 and 6 ran the pass and round 5 did not. 11 sessions in the same window ran it: 33003113749, 33004079412, 33004222904, 33005047706, 33033200749, 33033762380, 33034690191, 33040345276, 33041750750, 33050163753, 33100212747.
• Cost class: wrong outward action — one confirmed wrong claim posted publicly by a session in this set (see next finding); reduced review assurance for the other four.
• Detail: All 21 tend-review session logs in the window were downloaded and their Skill tool calls enumerated. Five sessions that reached step 4 and posted a verdict never invoked tend-ci-runner:code-review — no Skill call, no subagent, no tool-use error, and the string code-review appears zero times in their assistant text. Each used only Bash plus the two mandatory skill loads. Excluded as legitimate step-1 trivial-skip (that path says "Do NOT proceed to steps 2–6"): 33005705215, 33005806839, 33054039826, 33054138250 (posted nothing, resolved threads, stamped HEAD) and 33042204119 (the prescribed empty-body APPROVE). So the denominator is 16 eligible, not 21.
• Why this is the upstream follow-up rather than a fresh report: tend#937 made the pass unconditional in prose after 4 of 4 numbagg sessions skipped it, and closed with "If a later cycle finds sessions still skipping the pass with this text in place, the next step is a harder mechanism (a distinct numbered step, or a step-5 pre-post check that the pass ran) rather than more wording." The post-#937 rate here is 5/16, not 4/4 — wording helped and did not close it, which is the argument for the deterministic check upstream already proposed.
• Gates: Gate 1 PASSES on count alone (6 ≥ the 5+ stochastic floor; per review-gates.md the pre-registered condition does not lower that floor, so it is not being leaned on). Gate 2: targeted fix to a bundled skill — nothing repo-local applies, and a running-tend note would duplicate a mandate every consumer already carries. Gate 3: passes on cost class. Filed as #457, a permission ask per Other Repos, with no standing-exception offer per this repo's overlay.
• Cheap check for a future run: jq -r 'select(.type=="assistant") | .message.content[]? | select(.type=="tool_use" and .name=="Skill") | .input.skill' over each review session's JSONL. A session whose list is running-in-ci, running-tend only and which posted a non-empty review body is an occurrence; an empty-body APPROVE or a zero-post thread-resolution run is the legitimate trivial-skip path and is not.

First maintainer correction in this log's history: a review asserted bash -e semantics from memory and the claim did not reproduce

• Evidence level: Medium, occurrence 1
• Occurrences this run: 1 (historical: 0 — every prior entry from 2026-08-06 onward records "0 maintainer corrections" or "no comment contradicted a bot claim"; the one 2026-08-21 entry counted as a correction was a closed PR, not a contradicted claim)
• Run ID: 33103321834
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33103321834
• Session: 33053278119 (tend-review on #453, $1.00 / 20 turns), artifact claude-session-logs-n453
• Classification: stochastic — a model knowledge error about POSIX shell semantics, not a decision the skill text steers.
• Cost class: wrong outward action — the claim was posted publicly on a maintainer's PR and cost nedtwigg an empirical test to rebut.
• Detail: discussion_r3869957259 argued for if/fi over [ "$i" -lt 3 ] && sleep 15 on the grounds that "the step runs under bash -e, so the && form would exit non-zero on i=3 and skip the ::error:: line." nedtwigg ran both forms as script files under bash -e and neither exits — POSIX set -e exempts every command in an && list except the one following the final && (issuecomment-5436368751). The session log confirms the claim was never tested: 15 Bash calls, none of them a shell experiment. The maintainer took the if/fi spelling anyway on the arithmetic argument, which was correct, and removed the shell-semantics sentence from the comment rather than enshrining it — so the wrong reason did not reach main.
• Mitigating context, recorded so a future run weighs it correctly: the same exchange was net-positive by the maintainer's own account — "the version I'd have shipped without your first note was worse than either of us was arguing about: I'd already written the incorrect shell-semantics claim into the comment before testing it, so the disagreement is what got it tested at all." The arithmetic half of the finding (765s worst case against a 780s cap) was right and was applied.
• Gates: Gate 1 FAILS (stochastic, occurrence 1, floor is 5+). Not filed separately — it is cited inside #457 as the cost-class evidence for the finding above, not as a claim that the code-review pass would have caught it.
• Cheap check for a future run: the signature is a review body asserting how a tool or shell behaves with no corresponding Bash call exercising it. running-in-ci → Grounded Analysis already covers verifying an external tool's behavior; if this recurs, the candidate remedy is upstream, not overlay.

tend-nightly again had below-PR-bar findings and no channel to put them in — occurrence 2

• Evidence level: Low–Medium, occurrence 2 (historical: 1, Run 32627143122)
• Occurrences this run: 1
• Run ID: 33103321834
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33103321834
• Session: 33099389952 (tend-nightly, $7.50 / 70 turns / 16 min)
• Classification: structural (there is no cheap channel between "open a PR" and "say nothing"), but non-Critical, so the High floor of 2–3 governs.
• Cost class: wasted compute — analysis paid for and discarded; nothing wrong reached a public thread.
• Detail: Nightly's summary records two low-severity notes it judged below the PR bar and dropped: RemoteHost.#evictOldestPairingIfFull and RemoteHostService.#enqueuePairing both evict before the coalescing set, so a client re-sending its own pair frame denies an unrelated pending request even though the map would not have grown. Reachable only in the degenerate 8-pending-pairings state. The judgment not to open a PR looks right; the observation now exists only in a session log that expires in 30 days.
• Gates: Gate 1 borderline at occurrence 2 against a 2–3 floor; Gate 2 FAILS — every remedy considered (a standing notes issue, a nightly-appended log, an overlay channel) is new mechanism, which Gate 3 bars outright for a waste-class failure. Not filed. Recording the content here is itself the cheapest available channel, so the two notes are preserved above.

Healthy signals worth keeping in the record

• The three superseded PRs were superseded, not wrong, and the maintainer said so three times. tend-notifications 33002429742 opened docs(security): name every dotfile directory in the audit scopes, and dormouse-lib as a root #449 off the 18:10 audit failure on [security-audit] FAIL on 2026-08-27 #447; the review on docs(security): name every dotfile directory in the audit scopes, and dormouse-lib as a root #449 found a real workflow-audit.yaml bug that became fix(workflow-audit): report commits that only touch .github/audit/ #450; the review on fix(workflow-audit): report commits that only touch .github/audit/ #450 found the matching stale prose that became docs(security): name .github/audit/ in the audit-visibility window #451. nedtwigg folded all three into their own Close the audit's findings on itself, and run application-security on Opus #452 and closed each with "Closing as superseded, not as wrong" — and on docs(security): name every dotfile directory in the audit scopes, and dormouse-lib as a root #449 credited a bot finding with correcting them: "Your dormouse-lib finding was the one I did not have, and it corrected me." A review-spawned PR chain is working as intended; do not read the three closures as rejections.
• Every other bot review finding in the window was accepted. 19 nedtwigg comments, all but the bash -e rebuttal are acceptances — including "you found the root I had only partly reached" (r3868902677) and "thank you for correcting your own prior suggestion rather than letting the code keep it" (r3868350677), where a later review round reversed an earlier bot suggestion of its own accord.
• Nightly 33099389952 was exemplary. One real bug shipped as fix(standalone): stop a killed PTY's final flush from resurrecting its scrollback buffer #455 (a killed PTY's post-kill flush resurrects its scrollback via bufferScrollback's lazy-create branch, contradicting an invariant docs/specs/transport.md already states), regression test verified failing-on-main and passing-with-the-fix, 132 suites green, a 38-commit review with no defects filed, and correct restraint on the two below-bar notes.
• The abort pattern did not recur — count stays at 2. Every review session in the window either posted a verdict or ended on the documented trivial-skip path with the HEAD stamped; no session computed findings and dropped them.
• The 6 cancellations were ordinary concurrency, not lost work. 5 tend-mention runs cancelled at 02:36 and 02:54 are GitHub replacing a pending run when a newer one enters the group — nedtwigg posted 7 and 6 inline replies in those two bursts, and sibling runs answered all of them. The one tend-review cancellation (33050481904, Close the four security-audit findings: paste injection, dev-bridge RCE, proxy origin laundering, OSC forgery #453 at 727386f2) was likewise a pending replacement; run 33050163753 folded that push in, posted at 07:44 and stamped 54b92931. No commit went unreviewed.
• security-audit resolved inside the window. [security-audit] FAIL on 2026-08-27 #447 failed at 18:10 and 05:49 and passed at 14:13 after Close the audit's findings on itself, and run application-security on Opus #452 and Close the four security-audit findings: paste injection, dev-bridge RCE, proxy origin laundering, OSC forgery #453 merged; the issue closed itself.
• tend-ci-fix: 8 runs, all skipped at the trigger gate — main stayed green throughout.

Not tend artifacts, noted so a future run doesn't chase them

cloudflare-workers-and-pages deploy tables on #449/#450/#451/#452/#453/#454/#455 and the github-actions audit rows on #447 are the only non-nedtwigg, non-bot comments the Step 4 sweep returns. Neither is a maintainer correction. Separately: filing #457 will trigger a tend-triage run, since tend-triage's if: only skips tend-outage and tend-rate-limit labels — an unavoidable cost of the Other Repos permission-ask path, noted so a future sweep does not log that triage session as an unexplained no-op.

---

Run 33204541682 — 2026-08-28T19:43:23Z (review-runs)

Window 2026-08-27T18:24:41Z → 2026-08-28T19:36Z (26 h, anchored on predecessor 33103321834). 152 tend runs censused — 114 tend-mention (107 success / 4 cancelled / 3 skipped), 19 tend-review (all success), 11 tend-ci-fix (all skipped — main stayed green), 3 tend-notifications, 3 tend-triage (2 success / 1 skipped), 1 tend-nightly, plus yesterday's review-runs. 42 runs opened a Claude session, ~1050 turns, $92.04 list price: tend-review 20 sessions / $48.23, tend-mention 14 / $16.57, tend-nightly 2 / $14.66, yesterday's review-runs 1 / $5.64, tend-triage 2 / $3.66, tend-notifications 3 / $3.28. 1 run failure (drained, below). No near-timeout jobs — the only run whose created→updated span exceeded 30 min was 33182325132 at 99 min, and that is queue time from a manual re-run: its review job ran 5 min, against the default 360-min cap (no timeout-minutes override in .config/tend.yaml; the tend- census prefix is still complete). 8 bot-reviewed PRs merged (#458, #459, #461, #462, #464, #465, #466, and bot-authored #455), 0 bot PRs closed unmerged, 0 maintainer corrections — the corrections script ran and returned zero reviews, and all 13 nedtwigg comments in the window are acceptances of bot findings ("Correct", "Taken", "Good catch", "the first one was genuinely dead code and I'd have shipped it").

GATE-PASSING: none requiring a new filing. The one gate-passing finding is a recurrence of the pattern already filed as #457; evidence was appended there rather than re-filed. One further finding recorded below, under threshold.

tend-review posts a review verdict without running the mandatory /tend-ci-runner:code-review pass — occurrences 7–9

• Evidence level: High pattern, occurrence 9 (3 this run + 6 historical, Runs 32706405380 / 32826269519 / 33103321834)
• Occurrences this run: 3 of 15 eligible sessions (last window: 5 of 16)
• Run ID: 33204541682
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33204541682
• Sessions: 33126746664 (Self-host the coordinating server on Windows, behind Tailscale #458, COMMENTED, 2 findings), 33144072547 (fix(install-macos): clear the previous pointer when a failed update rolls back #461, COMMENTED, inline suggestion + design note), 33200941037 (chore: update tend workflows (0.1.19 → 0.1.20) #466, COMMENTED, stale SECURITY.md version claim). Excluded as legitimate step-1 trivial-skip: 33136755852 (Dor Tools: design spec + phase A (faces refactor) #459, --approve -b ""), 33143494391 (Self-host the coordinating server on Windows, behind Tailscale #458, --approve -b ""), 33144496957 (fix(install-macos): clear the previous pointer when a failed update rolls back #461, posted nothing), 33201499064 (chore: update tend workflows (0.1.19 → 0.1.20) #466, posted nothing). Ran the pass (12): 33121618666, 33127218109, 33136085750, 33143171787, 33144311974, 33182325132, 33193650001, 33198501104, 33199174590, 33200017856, 33200835292, 33203002139.
• Classification: stochastic, unchanged — mixed outcomes under identical conditions, split within each of Self-host the coordinating server on Windows, behind Tailscale #458, fix(install-macos): clear the previous pointer when a failed update rolls back #461 and chore: update tend workflows (0.1.19 → 0.1.20) #466 across adjacent rounds.
• Cost class: wasted assurance, not a wrong outward action this window. Every finding the three skipping sessions posted was correct and was accepted by nedtwigg. This weakens the cost half of Permission to file upstream: tend-review skipped the mandated code-review pass in 5 of 16 eligible sessions #457's original argument and is recorded on that issue so the decision has it.
• Detail: Same method as last window — all 19 tend-review session JSONLs downloaded, Skill tool calls enumerated (jq -r 'select(.type=="assistant") | .message.content[]? | select(.type=="tool_use" and .name=="Skill") | .input.skill'). The three sessions above reached a posted verdict with only running-in-ci + running-tend loaded; the string code-review appears in their logs only via the ambient skill listing.
• Gates: Gate 1 PASSES (9 ≥ the 5+ stochastic floor). Gate 2: targeted fix to a bundled skill — nothing repo-local applies. Gate 3: passes. Not re-filed — #457 is the open permission ask for exactly this and is still awaiting nedtwigg's yes/no; this run appended the second window of data to it. A future sweep should do the same rather than open anything new while Permission to file upstream: tend-review skipped the mandated code-review pass in 5 of 16 eligible sessions #457 is open.

tend-nightly's outage enrichment reads the run's latest attempt, so a row already re-run to success enriches as "no failure details" — occurrence 1

• Evidence level: Medium, occurrence 1
• Occurrences this run: 1
• Run ID: 33204541682
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33204541682
• Session: 33200497129 (tend-nightly, $7.16 / 59 turns)
• Classification: structural — the enrichment always reads the current attempt, so any row whose run was re-run before nightly fires produces the same empty result.
• Cost class: wasted compute — an uninformative comment on a bookkeeping issue; no wrong claim, and this sweep diagnosed the cause independently in one API call.
• Detail: #463 filed one row for run 33182325132 / trigger Retire "face": kinds are capability sets named terminal and browser #462 at 14:53:54Z. nedtwigg re-ran it at 16:26:20Z and attempt 2 succeeded. Nightly enriched at 18:47:17Z with "No failure details could be extracted." The details were fully available under actions/runs/33182325132/attempts/1/logs: three curl: (22) The requested URL returned error: 403 on the claude 2.1.241 download, then failed to install claude 2.1.241 after 3 attempts — a transient registry blip in the action's own install step, before any agent session started (which is also why no session artifact exists for attempt 1).
• Gates: Gate 1 FAILS at occurrence 1 against the 2–3 High floor. Gate 3 also bars it: waste-class, and every remedy is new mechanism (attempt-walking in the enrichment step). Not filed. Recording the diagnosis here and on Bot temporarily unavailable #463 is the cheap channel. If a future sweep sees this a second time, the remedy is upstream in tend's nightly enrichment step, not repo-local.
• Cheap check for a future run: an enrichment comment reading "No failure details could be extracted" on a tend-outage row whose run has run_attempt > 1.

Healthy signals worth keeping in the record

• Zero maintainer corrections, and unusually strong acceptance language. 5 conversation comments + 13 inline replies from nedtwigg, every one an acceptance. Three review findings were credited as ones the maintainer would have shipped past: "the first one was genuinely dead code and I'd have shipped it" (#467), "The physical-vs-logical root is the real find, and it is worse than latent for the case you named" (#465), and on #465 "Correct, and applied verbatim". On feat(selfhost): install the coordinating server on Linux, behind Tailscale #467 the maintainer reproduced the bot's 2>&1 redirection-order claim empirically and confirmed it — the inverse of last window's bash -e rebuttal.
• tend-nightly 33200497129 was disciplined. It found the Windows mirror of fix(install-macos): clear the previous pointer when a failed update rolls back #461's rollback-guard bug, checked before acting, found open PR fix(installers): a 200 does not say which release answered #465 already contained both hunks, and opened no duplicate. It also correctly left Bot temporarily unavailable #463 open, citing the rule that draining belongs to review-runs.
• tend-notifications × 3 were correctly conservative. 33144595672 and 33195342679 each marked one item read with no response, correctly identifying merges (not comments) as the trigger; the latter skipped install-macos.sh: a failed update leaves previous == current, so rollback silently does nothing #460 as inside the freshness gate and owned by tend-triage, which had indeed already handled it (33143809714 → fix(install-macos): clear the previous pointer when a failed update rolls back #461, merged). 33116195370 declined to act past a scope the maintainer had set — "silence isn't the word, and you scoped Close the four security-audit findings: paste injection, dev-bridge RCE, proxy origin laundering, OSC forgery #453 to four findings on purpose".
• The 4 tend-mention cancellations were ordinary concurrency, GitHub replacing pending runs inside a group during two comment bursts; sibling runs answered every inline reply.
• tend-ci-fix: 11 runs, all skipped at the trigger gate — main stayed green throughout.
• The abort pattern (review computes findings then drops them) did not recur — count stays at 2. Every one of the 19 review sessions either posted a verdict or ended on the documented trivial-skip path with HEAD stamped.
• Only 1 open bot PR (#469 at time of sweep) — the 5+ self-initiated-PR budget is nowhere near binding.

Not tend artifacts, noted so a future run doesn't chase them

• The Step 4 corrections script returns 12 cloudflare-workers-and-pages deploy tables and, on this repo, also returns the bot's own conversation comments (they match the window filter but not a maintainer-authored one). Neither is a correction. Filtering .user.login == "nedtwigg" directly on the issues/comments and pulls/comments endpoints is the check that actually answers the question.
• #457 (upstream-filing permission) and #456 (workflow audit) both still await a maintainer decision. Do not re-file either.

---

Run 33254833944 — 2026-08-29T13:35:29Z (review-runs)

Window 2026-08-28T19:35:29Z → 2026-08-29T13:23Z (18 h, anchored on predecessor 33204541682). 169 tend runs censused — 128 tend-mention success + 2 cancelled + 2 skipped, 24 tend-review (all success), 6 tend-ci-fix (all skipped — main stayed green), 3 tend-triage, 2 tend-notifications, 1 tend-nightly, plus yesterday's review-runs. 46 runs opened a Claude session, ~770k output tokens, $114.74 list price: tend-review 22 / $59.94, tend-mention 17 / $32.48, tend-nightly 1 / $10.17, tend-triage 3 / $6.54, tend-review-runs 1 / $3.32, tend-notifications 2 / $2.29. Zero run failures, no open tend-outage issue, nothing stranded. No long-running or near-timeout jobs — the longest created→updated span was tend-nightly at 28 min against the default 360-min cap (no timeout-minutes override in .config/tend.yaml; the tend- census prefix is still complete). 0 bot PRs merged or closed in the window, 0 maintainer corrections — the corrections script ran and returned zero reviews and zero dispositions; of its 23 comment rows, 8 are Cloudflare Pages deploy tables and the other 15 are all nedtwigg acceptances of bot findings ("Confirmed, and this was a real miss — thank you", "you're right, and it had teeth", "you were right on both counts").

GATE-PASSING: 1 — the skipped-code-review pattern reached occurrence 14 and produced its first wrong outward action. Evidence appended to the existing #457 rather than re-filed; the standing consequence on main is fixed in #482.

tend-review posts a review verdict without running the mandatory /tend-ci-runner:code-review pass — occurrences 10–14, now with standing damage

• Evidence level: High pattern, occurrence 14 (5 this run + 9 historical, Runs 32706405380 / 32826269519 / 33103321834 / 33204541682)
• Occurrences this run: 5 of 17 eligible sessions (prior windows: 3 of 15, 5 of 16)
• Run ID: 33254833944
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33254833944
• Sessions: 33204069218 (refactor(selfhost): let the server say which release it is, and make the contract executable #469, COMMENTED 2947 ch), 33209574328 (refactor(selfhost): let the server say which release it is, and make the contract executable #469, APPROVED 1062 ch — the wrong claim), 33214942256 (fix(vscode): let the webview CSP survive a code-split bundle #470, COMMENTED 1814 ch), 33217670540 (test(vscode): boot the built webview under its own CSP #471, APPROVED 478 ch), 33253953855 (fix(security-audit): clamp the issue body so a FAIL is never reported nowhere #480, COMMENTED 2950 ch). Ran the pass and posted (12): 33204561316, 33208246453, 33216564117, 33219552610, 33220174751, 33220876527, 33221056695, 33221058322, 33221127801, 33221599033, 33221802963, 33253525565. Excluded as legitimate trivial-skip (5): 33207850997, 33215940857, 33221562189, 33222270685 (posted nothing), 33254421927 (ran the pass, posted nothing on the bot's own PR). No artifact (2): 33216974035, 33204732088 — both ~3 min, neither maps to any review submitted in the window, so neither is counted either way.
• Classification: stochastic, unchanged — on refactor(selfhost): let the server say which release it is, and make the contract executable #469 the 19:40 and 20:49 rounds skipped while the 20:36 round between them ran the pass.
• Cost class: wrong outward action — first of the pattern's history. Session 33209574328's APPROVED review on refactor(selfhost): let the server say which release it is, and make the contract executable #469 asserted "macOS is 3 and Linux 2 against the real call sites too, so every minMatches is the exact occurrence count, which makes deleting any single copy red by construction." That is false on both platforms. A sibling tend-mention session narrowed it at 20:52:18Z with seeded experiments; nedtwigg merged refactor(selfhost): let the server say which release it is, and make the contract executable #469 at 20:52:23Z — five seconds later, so the correction arrived after the decision and the gaps shipped. This directly reverses the caveat the previous window's entry added ("no wrong claim reached a thread"), and that reversal was the material new fact posted to #457.
• Reproduced on today's main (2d4edb82), not taken from the thread: replacing install-macos.sh manage verify's three-branch identity block with an unconditional pass leaves deploy-lint: OK (3 installers, 9 rules, 27 checks) and deploy-lint-selftest: OK (27/27 rules are load-bearing). Separately, swapping install-linux.sh:1483 (if service_healthy "$RELEASE_ID") for a bare http_ok also leaves the lint green — Linux's minMatches: 2 counts the two service_healthy definitions, not the four call sites.
• Gates: Gate 1 PASSES (14 ≥ the 5+ stochastic floor). Gate 2: targeted fix to a bundled skill — nothing repo-local applies. Gate 3 now passes on the strong branch (wrong outward action), not merely the waste branch. Not re-filed — #457 is the open permission ask for exactly this and still awaits nedtwigg's yes/no; this run appended the third window plus the cost reversal. A future sweep should keep appending there rather than opening anything new while Permission to file upstream: tend-review skipped the mandated code-review pass in 5 of 16 eligible sessions #457 is open, and should only append when it has something new — a third identical count would be noise.

Action taken beyond the evidence log: #482

The macOS half of what the false approval certified is mechanical and had an exact Windows precedent (Windows already counts its structurally identical manage verify site), so it was fixed rather than left named: widen the macOS pattern to accept the assign-then-compare spelling and raise minMatches.macOS to 4. Each of the four macOS sites now individually turns deploy-lint red; before the change only three did. The Linux call-site gap and the floor-vs-exact question in minMatches are named in the PR body as maintainer calls, not decided — the first wants a second rule and a decision about which call sites must use the helper. Bot PR count was 4 before this, so the 5-open budget was not binding.

Healthy signals worth keeping in the record

• Zero maintainer corrections for the fifth consecutive window, with the strongest acceptance language yet. All 15 nedtwigg comments and inline replies in the window accept bot findings; three credit the bot with catching something that would have shipped — "the first one was genuinely dead code and I'd have shipped it" (#467), "Confirmed, and this was a real miss — thank you" (#472), "you're right, and it had teeth" (#471). On fix(deploy): make the Linux operator preflight actually detect the role #472 the maintainer verified a bot claim against ipn.Prefs source in tailscale v1.102.3 and confirmed it.
• tend-nightly 33253104890 ($10.17 / 99 turns) was the run of the window. It noticed security-audit had reached VERDICT: FAIL with a BLOCKER and then failed to file — audit-report.md was 68,424 bytes against GitHub's 65,536-char body limit, on a set -e step after the verdict, so the finding reached no issue and no comment. It diagnosed that from the run logs, wrote scripts/clamp-issue-body.mjs with a load-bearing self-test wired into pnpm test, applied the same guard preventively to workflow-audit.yaml, checked for prior rejections before opening #480, and surfaced the silenced security finding in the PR body rather than deciding a threat-model question. It also declined to file an unverified dor ensure observation because node_modules wasn't installed.
• Both tend-notifications runs were correctly conservative — 12 items checked across the two, all marked read, zero responses. 33225646685 chased the one shape that looks like a gap (two blank-bodied bot reviews on fix(standalone): stop the browser-dev harness persisting sessions the app drops #477 with no author reply) and confirmed via /pulls/477/comments that those reviews were the replies.
• The UI Tests overlay rule was applied correctly and then correctly retired. 33221386290 cited the repo overlay and stopped polling UI Tests on fix(theme): declare --mt-font-size on body so it can see --vscode-font-size #478; 33221917696 noticed it had since passed ("235 tests unchanged") and posted a correction to its own earlier statement on the thread.
• The 2 tend-mention cancellations were ordinary concurrency — 33220193901 and 33220193486, both killed ~10 s into the handle job during the 23:22 comment burst, with sibling run 33220193005 (same second) surviving and answering. Nothing stranded.
• Sibling-race discipline held everywhere it was tested. Five separate mention sessions explicitly declined to act on a review that landed while they polled, naming the run that owned it (33221385732, 33221386290, 33253790485, 33254190494, 33220193005).
• tend-ci-fix: 6 runs, all skipped at the trigger gate — main stayed green throughout.

Not tend artifacts, noted so a future run doesn't chase them

• The three open [workflow-audit] issues ([workflow-audit] 26 unexplained change(s) on 2026-08-27 #456, [workflow-audit] 2 unexplained change(s) on 2026-08-28 #468, [workflow-audit] 6 unexplained change(s) on 2026-08-29 #481) are not a tend gap. They are opened by github-actions[bot] from the repo's own workflow-audit.yaml, and a GITHUB_TOKEN-created issues event does not cascade, so tend-triage never fires on them — correctly, by GitHub's design. They accumulate awaiting nedtwigg, which is what they are for. Do not treat the absence of a triage run as a missed trigger.
• The corrections script's 8 cloudflare-workers-and-pages deploy tables are deploy status comments, not corrections. (Correcting the previous entry's note: the script does filter the bot's own comments — .user.login != $bot — so on this repo every non-Cloudflare row is genuinely nedtwigg.)
• #457 (upstream-filing permission) and #456 (workflow audit) both still await a maintainer decision. Do not re-file either.

[dormouse-bot]

---

Run 33313801568 — 2026-08-30T13:28:46Z (review-runs)

Window 2026-08-29T13:21:57Z → 2026-08-30T13:19:43Z (24 h, anchored on predecessor 33254833944). 104 tend runs censused — 73 tend-mention (72 success / 1 skipped at the trigger gate), 14 tend-review (all success), 7 tend-notifications (all success), 7 tend-ci-fix (all skipped — main stayed green), 1 tend-nightly, 1 tend-triage, plus yesterday's review-runs. 29 runs opened a Claude session, 588 turns, 467,772 output tokens, $71.34 list price: tend-review 14 / 394 turns / $28.40, tend-mention 9 / 344 / $22.06, yesterday's review-runs 1 / 71 / $6.92, tend-nightly 1 / 64 / $6.68, tend-triage 1 / 50 / $4.29, tend-notifications 3 / 45 / $2.99. Zero run failures, zero cancellations. No open tend-outage issue — nothing stranded to drain. No long-running (>30 min) or near-timeout jobs — every job in the window was measured and the longest was tend-mention 33273687949's handle at 18.2 min against the default 360-min cap (no timeout-minutes in any tend-*.yaml or in .config/tend.yaml; re-grepped all 13 workflow files for the tend action and only the 8 tend-* ones use it, so the tend- census prefix is still complete). 5 bot PRs merged (#477, #478, #479, #482, #484), 0 closed unmerged, 2 bot PRs open (#480, #485) — well under the 5-PR budget. 1 maintainer correction (below).

GATE-PASSING: none. Three findings recorded below — one is the pattern already filed upstream and now fixed-in-flight, one is a first-occurrence stochastic correction under the 5+ floor, one is a cost datum.

tend-review posts a review verdict without running the mandatory /tend-ci-runner:code-review pass — occurrences 15–16

• Evidence level: High pattern, occurrence 16 (2 this run + 14 historical, Runs 32706405380 / 32826269519 / 33103321834 / 33204541682 / 33254833944)
• Occurrences this run: 2 of 7 eligible sessions (prior windows: 5 of 16, 3 of 15, 5 of 17) — rate 29%, statistically unchanged
• Run ID: 33313801568
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33313801568
• Sessions: 33265390178 (fix(deploy-lint): count manage verify's identity check on macOS #482, COMMENTED 4290 ch), 33310308329 (chore: update tend workflows (0.1.20 → 0.1.22) #485, COMMENTED 3079 ch). Ran the pass (5): 33255345298 (ran it, posted nothing), 33264566860, 33264767399, 33265992121, 33273995469. Excluded as legitimate no-verdict or trivial-skip (7): 33264039614, 33264215786 (found a real defect in fix(theme): declare --mt-font-size on body so it can see --vscode-font-size #478's guard and opened #484 for it instead of posting a review — fix(theme): declare --mt-font-size on body so it can see --vscode-font-size #478 had already merged), 33264921169 (posted a CI-analysis comment, no review), 33266434220 (thread resolution only), 33273784556 (empty-body APPROVE on docs(specs): audit all 24 specs against the code #476), 33274579046, 33310773790.
• Classification: stochastic, unchanged — four consecutive rounds on fix(deploy-lint): count manage verify's identity check on macOS #482 split 3 ran / 1 skipped.
• Cost class: wasted assurance, no standing damage this window. The one wrong outward action either skipping session produced was self-caught: 33265390178 posted an inline suggestion on deploy-lint.mjs anchored at the value line but whose replacement text included the Windows: key, so applying it would have duplicated the key — the same session corrected it 20 seconds later and the author session applied the corrected single-line form in 601430b1. Both sessions' substantive findings were correct and were accepted.
• Gates: Gate 1 passes, but no action is owed. #457 was closed on 2026-08-29 after nedtwigg granted permission; the pattern now lives at max-sixty/tend#1079, which the upstream owner accepted the same day and answered with tend#1080 (promotes the pass to ### 5. Second pass and gates submit on its output). tend#1080 is still open and unmerged, and the newest release is 0.1.22 (2026-08-30T06:27Z), so the fix has not shipped — this repo's open #485 bumps 0.1.20 → 0.1.22 and will not change the rate. This window is pre-fix baseline, not escalation signal. Per the note 33272049480 carried forward from the upstream reply, the verification tend#1080 asked for is the skip rate on this repo after the fix ships in a release. Future sweeps: do not re-file, and do not append to tend#1079 until a release containing #1080 is pinned here — a fourth identical pre-fix count is noise on an issue whose fix is already authored.

Maintainer correction: a reporting fix was built ~3× larger than the failure warranted — occurrence 1

• Evidence level: Medium
• Occurrences this run: 1 (historical: 0 — the log's only prior correction, Run 33204541682's bash -e entry, was a wrong claim, a different failure mode; the 2026-08-21 #421 entry was a closed PR)
• Run ID: 33313801568
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33313801568
• Session: 33253104890 authored it (previous window's nightly); 33273687949 cut it back
• Detail: The observed failure was one line — gh issue create rejecting a 68 KB body, so a VERDICT: FAIL reached nobody. The nightly answered it with a 369-line scripts/clamp-issue-body.mjs + self-test carrying a nested-fence state machine, a SAFETY_MARGIN, a surrogate-pair guard, and a pathological-note branch. nedtwigg on #480: "this got ridiculously complex. Just truncate at 32,000 chars, don't worry about nested fences. This has way too many comments and explanation, it should be much more concise." The response was correct and fast — d88d246d is +51/−314, helper and self-test go 369 → 134 lines, and the 22-line workflow comment becomes 3 — but roughly two-thirds of the authoring session and one full review round (33273995469, $2.31) were spent on machinery that was then deleted.
• Classification: stochastic — the model chose to generalize past the observed failure; a replay might not. Nothing structural forced the fence handling.
• Cost class: wasted compute. Nothing wrong reached main; the over-built version never merged.
• Change type that would address it: a new paragraph of durable guidance (Gate 2: high bar, 3+ occurrences).
• Gates: Gate 1 FAILS — occurrence 1 against the 5+ stochastic floor. Gate 3 also fails (waste-class, and the remedy is a new rule every session loads). No action. Recorded so a future sweep can count it.
• The second half of the correction needs no action and should not be re-derived. nedtwigg also asked for "a note to AGENTS.md about being concise with comments, and don't duplicate in a comment something which is already explained in a spec (or vice-versa)." The bot declined and was right to: **Say it once.** landed on main in fix(theme): declare --mt-font-size on body so it can see --vscode-font-size #478 (40433b9a) ~3½ hours earlier and covers both halves verbatim — "An invariant and its rationale live in exactly one place — the owning spec when one covers it... Write a comment only for a constraint the code cannot show, and keep prose proportional to the change." Verified present in today's AGENTS.md. The branch was behind that commit, which is why it wasn't visible in the diff. Adding a second note would be the thing the note forbids.

Cost datum: #482 drew 8 sessions / $17.65 across 4 self-review rounds — occurrence 3 of the designed path

• Evidence level: Low (observation, not a defect)
• Occurrences this run: 1 (historical: #399 and #446 at $23.40)
• Run ID: 33313801568
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33313801568
• Detail: 5 tend-review + 3 tend-mention sessions on one 129/30-line PR, 25% of the window's spend on 1 of 5 merged PRs. The loop earned it. Round 1 caught that the |serving alternative was bound to nothing, so any of the four manage verify sites could be rewritten with a bogus assignment and keep deploy-lint green — strictly worse than the undercount it replaced. Round 2 caught two skip reasons describing code that doesn't exist and an unbound cur_id that let verify green-tick whatever holds the port. Round 3 caught an over-reaching comment claim. Each finding was reproduced before it was written, and #482 merged. This is the designed path, recorded for the cost series only.

Healthy signals worth keeping in the record

• Sibling-race discipline held under the heaviest concurrency in the log. Both tend-mention sessions on fix(deploy-lint): count manage verify's identity check on macOS #482 (33265115782, 33265655672) had a fresh review land mid-poll, confirmed the same-repo dispatch would cover it, named the specific inline comment IDs they were leaving, and exited rather than racing. Neither review was dropped — the successor round actioned both.
• The upstream filing was executed correctly end to end. 33270740043 re-verified all four load-bearing claims against source rather than copying them off the thread (no duplicate upstream issue, the step-4 mandate verbatim in the pinned skill, the pin is post-#937, the refactor(selfhost): let the server say which release it is, and make the contract executable #469 approval's false claim reproduces), filed tend#1079, reported back, closed Permission to file upstream: tend-review skipped the mandated code-review pass in 5 of 16 eligible sessions #457 — and flagged the close as a judgment call the maintainer could reverse. The overlay's no-standing-exception rule was respected: permission was asked and granted per-issue.
• tend-triage 33262387015 declined to guess. On the Webview renderer leaks ~25 MB/hour while idle #483 memory-leak report it refused a fix PR (no failing test reproducible from a CI container, deps not installed), and instead ruled out all three ranked suspects with a bounded-accumulator audit — the load-bearing result, since chasing them would have burned the heap-snapshot window — then handed back a one-flag discriminator (cfg.terminal.webglRenderer).
• tend-nightly 33310094513 reported rather than re-fixed. It found security-audit red for a second consecutive night on the same 65,536-char body limit, pulled the audit-transcript artifact to confirm (69,185 bytes), and commented the recurrence on the existing #480 instead of opening a duplicate. It also declined twice on purpose: it did not file a public issue for the unfixed BLOCKER-1 (a disclosure call it named as the maintainer's), and it explicitly argued against adding security-audit to watched_workflows because the audit is supposed to go red on a real finding.
• A review session converted its finding into the right artifact. 33264215786 found that fix(theme): declare --mt-font-size on body so it can see --vscode-font-size #478's new mirror guard used themeCss.match() (stops at the first hit) and missed @theme inline; fix(theme): declare --mt-font-size on body so it can see --vscode-font-size #478 had already merged, so it opened #484 rather than posting a review into a closed thread. Mutation-verified both directions; merged 2 h later.
• All 3 tend-notifications Claude sessions were correctly conservative — 7 items checked, 7 marked read, 0 responses. Each disposition was traced to the run that had already handled it, and the cross-repo tend#1079 row was correctly read-only.
• tend-ci-fix: 7 runs, all skipped at the trigger gate — main stayed green throughout.

Not tend artifacts, noted so a future run doesn't chase them

• The previous entry's closing line is now stale — ignore it. It said "[Permission to file upstream: tend-review skipped the mandated code-review pass in 5 of 16 eligible sessions #457] and [[workflow-audit] 26 unexplained change(s) on 2026-08-27 #456] both still await a maintainer decision. Do not re-file either." #457 was answered ("You may file upstream"), filed as tend#1079, and closed on 2026-08-29T19:26:33Z. #456 is still open and still awaits nedtwigg.
• The Step 4 corrections script returned 10 comment rows; 7 are cloudflare-workers-and-pages deploy tables. Of the 3 real rows, one is the fix(security-audit): clamp the issue body so a FAIL is never reported nowhere #480 correction above, one is nedtwigg's "You may file upstream." on Permission to file upstream: tend-review skipped the mandated code-review pass in 5 of 16 eligible sessions #457 (a grant, not a correction), and one is "@dormouse-bot please merge origin/main and resolve any conflicts appropriately" on docs(specs): audit all 24 specs against the code #476 (a directive, executed in 33273379448). reviews and dispositions-as-corrections were both empty; all 5 dispositions are bot PRs nedtwigg merged.
• security-audit's BLOCKER-1 is a live project-health signal, not a tend gap. audit-application.md has opened VERDICT: FAIL on the iframe proxy's injected shim two nights running and reached no issue both times. The reporting fix is #480, open, green, and with the maintainer's review addressed; the threat-model question underneath it is explicitly nedtwigg's call. Nothing further for a sweep to do while fix(security-audit): clamp the issue body so a FAIL is never reported nowhere #480 is open.
• #476 was approved with an empty body on a commit the bot's own sibling had pushed 4 minutes earlier. This is the documented step-1 trivial-skip path on a maintainer-authored PR, and the approving session (33273784556) explicitly re-verified all four conflict resolutions and all four repointed anchors against the merged tree rather than trusting the sibling's summary. Recorded because the shape — approving a head the bot authored — is worth a second look if it recurs on a bot-authored PR, where self-approval protections do not apply the same way.

---

Run 33408678742 — 2026-08-31T15:40:37Z (review-runs)

Window 2026-08-30T13:18:02Z → 2026-08-31T15:31:41Z (26 h, anchored on predecessor 33313801568). 208 tend runs censused — 138 tend-mention (135 success / 3 skipped at the trigger gate), 41 tend-review (all success), 20 tend-ci-fix (19 skipped, 1 success), 6 tend-notifications, 1 tend-nightly, 1 tend-weekly, plus yesterday's review-runs. Zero failures, zero cancellations, zero timeouts. No open tend-outage issue — nothing stranded to drain. No long-running (>30 min) or near-timeout jobs: the longest run in the window was tend-review 33355495101 at 17.1 min against the default 360-min cap (re-grepped — no timeout-minutes in any tend-*.yaml or in .config/tend.yaml; only the 8 tend-* workflows reference max-sixty/tend, so the tend- census prefix is still complete). 51 runs opened a Claude session, $125.51 list price. 3 bot PRs merged (#496, #485, #480), 0 closed unmerged. Zero maintainer corrections — the Step 4 script ran and returned 19 comment rows of which 15 are cloudflare-workers-and-pages deploy tables and 4 are the bot's own inline replies; reviews was empty and all 3 dispositions are bot PRs nedtwigg merged.

GATE-PASSING: none. Four findings recorded below — one is a new structural first occurrence, one is a new cadence observation, one is the milestone that arms the pre-registered code-review verification, one is a cost datum.

#493 carries a standing APPROVED while the bot's own newest review says the code is broken — occurrence 1

• Evidence level: High
• Occurrences this run: 1 of 14 PRs the bot reviewed in the window (historical: 0 — grepped all five evidence comments for stale approval / dismiss / standing approval / reviewDecision; the three hits are the correct shape, a standing approval at the live head with no later findings)
• Run ID: 33408678742
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33408678742
• Sessions: 33355244058 posted the APPROVED at 5e71658c; 33358950402, 33360958780, 33362497865 posted the three COMMENTED reviews after it; 33380512539 (tend-notifications) spotted the state and declined to act
• Detail: The review timeline on #493 is COMMENTED ×5 → APPROVED (04:05:56Z, 5e71658, empty body) → COMMENTED ×3. GitHub keeps the most recent APPROVED as the standing decision — a later COMMENTED never supersedes it — so gh pr view 493 --json reviewDecision reads APPROVED at live head e3be7ef2 right now, while the 06:11:38Z review at that exact head opens "But the approval path still cannot start the tool" and reproduces it: resolveToolApproval clears toolPending before two awaits, which flips toolFace to terminal, mounts TerminalPane, and getOrCreateTerminal spawns a bare default shell with no cwd before the round-trips land — so Always allow yields an empty shell and the tool's command never runs. It also notes resolveToolApproval has no test at all. The PR is mergeStateStatus: DIRTY, so nothing could merge on the stale verdict today, and the findings are the top of the thread — but the repo's public review state says the bot approves code the bot says is broken.
• Classification: structural. No decision point: GitHub's reviewDecision is defined this way, so any session that approves and then finds blocking problems on a later push produces this state every time. running-in-ci -> CI Monitoring gates an approval on red CI ("Approving a PR is also gated: dismiss it on red") and says nothing about the bot's own subsequent findings, so no guidance covers the case.
• Cost class: wrong outward action. The stale APPROVED is on the public record at the live head as this is written.
• Change type that would address it: a targeted added step — dismiss your own standing approval when a later review of the same PR finds blocking problems (normal Gate 2 bar → Gate 1 thresholds).
• Gates: Gate 3 passes comfortably (wrong-outward-action class). Gate 2 is the normal bar. Gate 1 FAILS — occurrence 1 against the 2–3 a structural High needs; the gates are explicit that structural classification raises recurrence confidence but does not override the count. No action this run.
• Deliberately not dismissed here, and a future sweep should not dismiss it either without reading the diff. Two sessions have already weighed it: the 06:11Z review session had the full diff in hand and chose COMMENTED over a dismissal, and 33380512539 declined on the ground that "re-adjudicating that from a poll that hasn't read the diff isn't this run's call" — correct on both counts. Re-adjudicating a deliberate call from a sweep would be the worse error. Recorded so occurrence 2 acts.

tend-notifications fires ~4% of its */15 cron, and the shortfall stepped down around 2026-08-26 — occurrence 1

• Evidence level: Medium
• Occurrences this run: 1 (historical: 0 as a finding — the 2026-08-30 entry recorded "7 tend-notifications (all success)" and the 2026-08-01 entry "~14 tend-notifications (hourly …)" as healthy counts, without measuring either against the schedule)
• Run ID: 33408678742
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33408678742
• Detail: .github/workflows/tend-notifications.yaml has asked for cron: "*/15 * * * *" unchanged across every regeneration back through 38f4e207 (0.1.19, 2026-08-26) — verified by reading the cron line out of all four commits that touched the file. Actual scheduled deliveries per day: 32 (08-24), 34 (08-25), 21 (08-26), 3 (08-27), 2 (08-28), 6 (08-29), 7 (08-30), 2 (08-31 partial). Six fired in this window, at 15:20, 18:42, 21:35, 23:52, 03:28 and 10:01 — gaps of 2–6.5 h against a 15-minute schedule, so ~6 of ~104. Even the pre-step-down 32/day was under a third of the schedule. Nothing was dropped or cancelled — these are events GitHub never delivered, which is documented best-effort behavior for schedule and correlates with this repo's Actions load (208 tend runs in this window alone, against 104 in the previous one). The daily tend-nightly / tend-weekly / tend-review-runs crons all fired exactly once, so this is specific to the high-frequency cron.
• Classification: structural — GitHub-side scheduling, identical on replay.
• Cost class: neither waste nor a wrong outward action — under-delivery of a backstop, with no observed consequence. This run's exhaustive live scan found nothing stranded: all 4 open PRs carry a bot review at their live head, every bot-directed comment and review in the window has a response, main CI is green at 6bb69a88, and there is no open outage tracker.
• Gates: no gate is even reached — there is no lever. The cron already asks for 4× what GitHub delivers, so raising it is not a remedy, and nothing in the repo can make GitHub honor a schedule. No action, now or later. Recorded because running-in-ci -> "A review that lands while you poll is not yours to action" names the notifications poll as the fallback that catches fork-PR reviews tend-mention cannot dispatch on — that fallback is running every ~4 h, not every 15 min, and a future sweep that sees a fork-PR review go unanswered should look here first rather than re-deriving it.

Code-review pass: the pre-registered verification is now armed — 0.1.24 on main contains tend#1080

• Evidence level: N/A (milestone, not a defect)
• Occurrences this run: pre-fix baseline measured, below
• Run ID: 33408678742
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33408678742
• Detail: The previous entry left a standing instruction — "do not re-file, and do not append to tend#1079 until a release containing #1080 is pinned here." That condition is now met. max-sixty/tend#1080 merged 2026-08-31T01:36:23Z at 558b6db5; 0.1.24 published 2026-08-31T05:53:20Z and gh api compare/558b6db5...0.1.24 returns ahead_by: 6, behind_by: 0, so the fix is an ancestor of the tag; #496 merged the 0.1.22 → 0.1.24 pin to main at 2026-08-31T15:15:47Z. The instruction still holds for this run — every tend-review session in this window except the last four ran pre-fix, so there is no post-fix sample to report.
• Pre-fix baseline for this window (all 41 tend-review session logs downloaded and parsed for a Skill: tend-ci-runner:code-review call): 29 sessions posted a verdict, 18 ran the pass, 11 did not. Raw 38%, but that number is not comparable to the historical 29%, which counted only eligible sessions. Reading the 11 tails, 8 are the documented trivial-skip / close-out path (a maintainer applied a prior inline suggestion, the session verified a 1–6-line increment and approved, or it only resolved threads); the substantive ones are 33332996084 (docs(specs): audit all 24 specs against the code #476, COMMENTED 2852 ch), 33342650053 (docs(specs): compress references and code commentary #489, COMMENTED 4857 ch) and 33344272554 (docs(specs): compress references and code commentary #489, APPROVED) — 3 of 29, ~10%, and both COMMENTED ones found real defects anyway.
• Post-fix sample: 4 sessions, 0 defects, and far too small to mean anything. 33407554013, 33407581226 and 33407776084 all ran the pass. 33407921343 did not, and is a correct trivial skip, traced end to end: the incremental on chore(deps): update dependency hono to v4.13.5 #500 was the one-line website/src/data/dependencies-npm.json snapshot regen (4.13.3 → 4.13.5) that its own prior COMMENT review had asked for, so it took the close-out path, approved with an empty body, and said so.
• Gates: no finding, no action. The next sweep is the first with a full post-fix window and owns the verification tend#1080 asked for — measure eligible-session skip rate against the ~29% historical figure, counting close-out rounds as ineligible, and report to tend#1079 only once there is a real sample.

Cost datum: $125.51 in 26 h, 88% of it tend-review on 10 maintainer PRs — the designed path

• Evidence level: Low (observation, not a defect)
• Occurrences this run: 1 (historical: $71.34 across 104 runs on 2026-08-30, $8.05 on 2026-08-01)
• Run ID: 33408678742
• Workflow: https://github.com/diffplug/dormouse/actions/runs/33408678742
• Detail: tend-review 41 runs / $110.50; tend-nightly 1 / $5.03; yesterday's review-runs 1 / $4.86; tend-notifications 4 / $2.97; tend-mention 2 / $1.03; tend-ci-fix 1 / $0.64; tend-weekly 1 / $0.48. By subject: Dor Tools Phase B: the tool Surface — a terminal and a browser on one Session #493 drew 11 review sessions / $36.01 and docs(specs): the rationale split — convention, lint, and the eight fattest specs #487 4 / $11.03. The driver is throughput, not defect: nedtwigg landed 75 commits and 10 PRs in the window, and the review loop is doing its job on them — Dor Tools Phase B: the tool Surface — a terminal and a browser on one Session #493's rounds each found real correctness problems, the newest reproduced above. 136 of the 138 tend-mention runs opened no Claude session at all: they are 2–6 s verify gate evaluations with handle skipped, ~10 runner-minutes total, which is the trigger gate working as designed and is not waste. No action; recorded for the cost series.

Healthy signals worth keeping in the record

• tend-ci-fix 33357019261 diagnosed and shipped nothing, correctly. It reproduced the Build & Test failure in a throwaway origin/main worktree, found the root cause (Selfhost onboarding: stage the scope, ship truthful Pair/Connect #492 grew docs/specs/server.md to 8421 words against its 8350 budget, failing spec-lint as pnpm test's first step), then found Pocket first-run truth: setup-first auth and install-before-identity #494 had already merged the budget bump ~13 s later — and stopped. The other 19 tend-ci-fix runs all skipped at the trigger gate; main never stayed red.
• tend-nightly 33399147004 opened exactly one PR and declined the rest. It read the 75-commit pair-status feature end to end across four layers before pronouncing it sound, verified shortcuts.md row-for-row against handle-pane-shortcuts.ts, and caught that tend init on 0.1.24 also writes .github/actionlint.yaml — which the recipe's staging paths miss, so it had been landing untracked every run — and added it. It also declined to file utf8Decode's overlong-encoding asymmetry after tracing all five call sites and finding none canonicalization-sensitive, and stated its verification limits (no node_modules, so no package suites).
• The notifications sweep flagged the stale approval rather than acting on it. 33380512539 is the only reason finding 1 above exists; it identified the exact mechanism ("GitHub keeps that as the standing approval state — later COMMENTED reviews don't supersede it") and then correctly refused to dismiss from a poll that had not read the diff.
• All 6 tend-notifications runs were conservative and traced. Each disposition was tied back to the run that had already handled it, the cross-repo tend#1079 row was correctly read-only, and dedup caught the bot's own comment inside the 60 s-padded cutoff on chore: update tend workflows (0.1.20 → 0.1.22) #485 rather than double-posting.
• tend-weekly 33316766244 was a clean 41-line no-op — 0 dependency PRs open at the time, no repo-specific weekly tasks in the overlay, exited.

Not tend artifacts, noted so a future run doesn't chase them

• #486 [workflow-audit] 2 unexplained change(s) is open and needs no bot action. It is authored by github-actions, not the bot, and asks a human to account for two commits — both of which are 8ac6d23 and d88d246 on fix/audit-report-body-limit, i.e. #480, which nedtwigg merged. tend-nightly read it this window and ruled it out of scope for auto-close; that is right, and the accounting is a maintainer signature, not a bot comment. Prior [workflow-audit] issues ([workflow-audit] 6 unexplained change(s) on 2026-08-29 #481, [workflow-audit] 2 unexplained change(s) on 2026-08-28 #468, [workflow-audit] 26 unexplained change(s) on 2026-08-27 #456, …) are all closed, so the open one is simply the newest.
• The Step 4 corrections script returned no corrections. 19 comment rows: 15 cloudflare-workers-and-pages deploy tables (their updated timestamps move as the deploy table is edited, which is why they land in an updated_at window), plus 4 of the bot's own inline replies applying nedtwigg's suggestions on Selfhost onboarding: stage the scope, ship truthful Pair/Connect #492, Pocket first-run truth: setup-first auth and install-before-identity #494 and Pocket push: a step of the first connect, not a footnote #495 — each an acceptance, not a correction. reviews empty; all 3 dispositions are merged bot PRs.
• The overlay's Chromatic rule held. Sessions on Pocket first-run truth: setup-first auth and install-before-identity #494 and Pocket push: a step of the first connect, not a footnote #495 both named UI Tests as human-gated per the overlay and stopped polling rather than waiting out the cap — the behavior the overlay was written for.

[dormouse-bot]

Superseded by #511 (2026-09).