[dancingteeth]

Harness improvements for Cursor dogfood (composer worker, grok judge). Separate commits; no version bump; draft.

1. Verify: sidecar default (behaviour change)

• verifyLogMode defaults to sidecar; tail-biased preview; verify-logs/ gitignore guidance in README/init.

2. Billing token breakdown on green runs

• Worker/judge uncached in/out/cache rows; Cursor cache marked unknown.

3. Judge resilience

• Primary judge error → retry → reviewSecondary fallback primary; park review_judge_failed when both fail.
• Fallback path sets skipMergeSecondaryReview so the same model does not run merge secondary.
• Secondary merge failures are non-fatal (primary verdict stands).

4. Tool-error stats

• Per runtime/model in log.ndjson and run-report.md.

5. Transient worker SDK errors

• Per-iteration classify/retry with cap; partial usage once; fatal logs iteration row; transient verify reason says harness will retry (not “loop stopped”).

Exit codes (owner: Paul)

• Parked status: waiting exits 2 (same as main 0.8.2 incomplete), not 0 or 3. run-batch unchanged (0 | 2).

Verified

• pnpm test, pnpm exec tsc -p tsconfig.json --noEmit

Open questions

• Cursor cache token semantics when SDK reports zeros vs omitting fields.

  

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: ca7d9f03-e1fe-4f73-9ddd-f7d494e32595

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[dancingteeth]

Reviewey Ray sensor (automated review, not a merge decision; the human owns merge)

PR: #45 (draft) · push 0f3fa037 adds 3 commits on top of d1e15b14: 5152a75f fix(cli): exit code 3 for parked runs, 9736e8f4 fix(review): secondary merge failure must not fail the run, 0f3fa037 fix(loop): continue on transient worker SDK errors with cap. First sensor on this PR, so the earlier 5 commits (sidecar verify default, billing tokens, judge fallback, tool-error stats, docs) are covered too.
Mode: full Unified Code Review (risk HIGH, so no 10-count wait). Agent-authored (Cursor Agent, co-authored by Paul).

Risk

HIGH

What could go wrong?

• Parked runs (status: waiting) now exit 3; on main 0.8.2 they exit 2. Any script, CI step, or wake hook that branches on 2 changes behavior silently.
• A fatal worker error now double-counts partial Cursor usage and no longer writes its iteration row to log.ndjson.
• agent-loop-batch still exits 2 for parked loops, so the new "exitCode matches the CLI" table is wrong for batch.
• When the primary judge errors and reviewSecondary takes over as fallback judge, the same secondary config is still passed as the merge secondary, so it may review its own verdict a second time.
• README and CHANGELOG still say judge-down parks exit 0, which contradicts the new 3.

Review depth

• Journeys at risk: loop CLI exit code / AGENT_LOOP_DONE.exitCode consumers; review gate with reviewSecondary; worker SDK fault handling (transient vs fatal); run-report cost totals; verify prompt content (default verifyLogMode flips to sidecar).
• Line-by-line: src/loop/agentLoop.ts (worker catch, outer catch, resolveLoopCliExitCode), src/loop/loopWorkerAgentError.ts, src/agents/agentRunFailed.ts, src/agents/cursorAgent.ts, src/review/loopPostReview.ts (maybeRunSecondaryReview), src/loop/loopPostSuccessReview.ts (retry + fallback), src/review/reviewJudgeError.ts, src/loop/captureText.ts, src/loop/loopExtensions.ts (sidecarPreview), src/cli/run.ts, src/cli/run-batch.ts, README/CHANGELOG.
• Empirical checks: read head 0f3fa037 and main copies of src/cli/run.ts, src/cli/run-batch.ts, src/loop/agentLoop.ts, src/loop/loopWorkerRetry.ts, src/agents/errorFormat.ts. Did not run tests. CI test (22) pending at review time; CodeRabbit skipped (draft).
• Reversibility: easy in code; exit-code change is hard to take back once published to npm.
• Verification gap: adjacent-only (new tests cover transient continue, cap park, fatal stop, secondary throw; nothing covers fatal usage/log row, batch exit for parks, or fallback-judge self-merge).
• Release guardrails: CHANGELOG Unreleased, no version bump; draft PR.
• Change set: base main → head 0f3fa037ecc0fbb356c975f7e8208076bca7992c

Pass 1b overlay (REVIEWS.md): "### HIGH — network egress, agentic tools, cline sdk, secondary judge, review gate". This PR changes secondary judge and review gate behavior, hence HIGH. Also: "Verifier is the hard gate; review is residual." The judge-down park keeps verify as the gate, which fits.

§2b / §2c Lite: opened shouldEscalateAfterWorkerFault and isTransportAgentError (classification callees), run-batch.ts exit mapping (line 250), and the outer catch in runAgentLoop (fatal rethrow target). Findings below come from those reads.

Verdict

ADVISORY

Advisory

• [owner decision — Paul only] Parked runs exit 3 instead of 2 (the main 0.8.2 behavior), and verifyLogMode now defaults to sidecar instead of inline. Both change what existing users and scripts see. The commits give no source for the exit-code call. An agent must not invent approval text, decision-log entries, or "owner approved" lines; wait for Paul to confirm (in session or here), then note in CHANGELOG that this is a breaking change and bump the minor version when releasing.
• src/loop/agentLoop.ts worker catch: partial usage is added to usageSummary before the classifyWorkerAgentError(err) === 'fatal' rethrow, and the outer catch calls extractPartialWorkerRun(err) and adds it again, so fatal runs double-count cost. Fix: move the fatal check above the inner addUsageRecord (let the outer catch own it), and add a test asserting result.usage.records.length === 1 for a fatal AgentRunFailedError with usage.
• Same catch: the fatal rethrow now happens before appendJsonlLine(logPath, …), so a fatal worker error no longer writes its iteration row to log.ndjson (the old test asserted verify.command === '(agent SDK)', and that assertion was removed). Fix: write the iteration row (and tool-error stats) first, then rethrow on fatal; restore the (agent SDK) log-row assertion in the "stops immediately on a fatal worker agent error" test.
• src/cli/run-batch.ts:250 still maps every incomplete batch to exitCode: 0 | 2, so a batch whose loop parked exits 2, while README now says exitCode 3 means parked. Fix: either map parked batch results to LOOP_EXIT_PARKED (via resolveLoopCliExitCode or an equivalent batch-status check) with a test, or say in the README table that the codes apply to agent-loop only and batch still uses 0/1/2.
• src/loop/loopPostSuccessReview.ts invokeReview(secondary, …): the fallback call spreads baseReviewOptions, which still carries reviewSecondaryRuntime/reviewSecondaryModel, so maybeRunSecondaryReview can run the same secondary judge again as a merge reviewer of its own verdict (extra cost, no independent signal). Fix: when invoking the fallback, pass reviewSecondaryRuntime: undefined and reviewSecondaryModel: undefined, and add a test that the fallback path calls the review agent exactly once.
• Docs contradict the new exit code: README line 352 ("CLI exit 0") and CHANGELOG Unreleased line 19 ("CLI exit 0") both describe judge-down parks. Fix: change both to exit 3, matching the new table and CHANGELOG line 23.

Nits

• PR body still says "exit 0 for waiting parks"; update it to 3 and add the three new commits to the summary.
• src/review/loopPostReview.ts secondary catch uses err.message; use formatErrorChain(err) so the cause chain lands in review.md, as the worker path does.
• src/loop/loopWorkerAgentError.ts: the fatal pattern is checked first and includes the bare phrase api key, so a provider message like "rate limit exceeded for this API key" is classified fatal. Narrow it (for example api key (is )?(missing|invalid|not set)) and add a test for that rate-limit wording.
• "Cursor agent returned empty result" matches neither pattern and falls through to fatal; decide whether an empty result should be transient (it usually is) and add a classification test either way.
• On the last iteration (i === maxIterations), a transient fault still recycles the session and sleeps TRANSIENT_AGENT_ITERATION_BACKOFF_MS before the loop ends. Skip the recycle and backoff when no iteration follows.
• Transient iterations add partial usage but skip stopOnBudget(i); call it after the partial addUsageRecord so budget caps hold during retry streaks.

[dancingteeth]

Reviewey Ray sensor (full Unified Code Review, triggered by HIGH risk; repo commit count 35). This is a sensor, not a merge decision: Paul owns merge. Agent loops may act only on ### Blockers.

Risk

HIGH

What could go wrong?

• Changing the CLI exit-code contract: on main, parked (status: waiting) runs exit 2. This PR first maps them to 0 (commit 4c51701) and now to 3 (5152a75). Any wrapper, CI step or bot that branches on $? will behave differently for budget, env/setup, judge-down and transient-cap parks.
• The worker error path now keeps looping on transient faults instead of stopping, so a misclassified fault can burn up to maxConsecutiveTransientAgentErrors (default 3) worker iterations plus a 5s backoff each.
• Usage/cost totals, which are new billing rows in run-report.md, can be wrong on the fatal worker path.
• verifyLogMode now defaults to sidecar, so every consumer repo starts getting verify-logs/ files under the loop dir.

Review depth

• Journeys at risk: agent-loop run exit code and AGENT_LOOP_DONE.exitCode; agent-loop run-batch exit; worker SDK fault handling; review gate when judges error; cost reporting in run-report.md.
• Line-by-line: src/loop/agentLoop.ts worker catch (L500–592) and outer catch (L872–887), resolveLoopCliExitCode, src/loop/loopWorkerAgentError.ts, src/agents/cursorAgent.ts, src/agents/agentRunFailed.ts, src/review/loopPostReview.ts maybeRunSecondaryReview, src/loop/loopPostSuccessReview.ts judge retry/fallback, src/loop/captureText.ts, src/loop/loopExtensions.ts sidecar preview, src/cli/run.ts, src/cli/run-batch.ts, README/CHANGELOG.
• Empirical checks: I read the diff and head files only; nothing was run locally. I compared against main for run.ts (it uses result.complete ? 0 : 2, so waiting exits 2 today) and loopWorkerRetry.ts (escalation only applies to recoverable faults, so classifying fatal before escalation is safe).
• Reversibility: easy in code. The exit-code contract is hard to walk back once it ships to npm.
• Verification gap: adjacent-only. There are new tests for transient continue/cap, secondary-merge failure and exit mapping, but none for fatal-path usage or the batch exit code. CI test (22) was still in progress at head 0f3fa037.
• Release guardrails: still a draft PR. The CHANGELOG entry is under Unreleased with no version bump.
• Change set: base c4e7195c (main) → head 0f3fa037 (8 commits, 43 files). This is the first sensor on Harness: sidecar verify, usage split, judge resilience, tool errors #45, so it covers the whole PR, with the closest look at the three new fix commits 5152a75, 9736e8f and 0f3fa03.

Pass 1b (REVIEWS.md overlay): I'm quoting the laws that apply here. "Verifier is the hard gate; review is residual." "Gating impacts: data-loss | security-boundary | false-closure | cross-dispatch | verify-bypass." "2b. Always — open cited callee before gating blockers; live path or [latent_contract]." None of the findings below meets a gating impact: verify stays the hard gate, and judge-down parks without closing the run as done. So nothing is gating.
Pass 2 (agent-authored, Cursor Agent co-authored): Test hunks come first and match the commit claims for transient continue, the transient cap park, secondary failure keeping the primary verdict, and the exit 3 mapping. The PR body still says "exit 0 for waiting parks", which no longer matches the head commit.
§2b / §2c Lite: I opened shouldEscalateAfterWorkerFault, workerSdkVerifyResult, maybeRunSecondaryReview and run-batch.ts's exit path. Two of these show mismatches, listed below.

Verdict

ADVISORY

Advisory

• [owner decision — Paul only] Parked runs exit 3. On main, status: waiting exits 2. This PR makes it 3, which is a public CLI and AGENT_LOOP_DONE contract change, and no commit gives a source for the decision. Paul needs to confirm 3 (or keep 0 or 2). An agent must not invent approval text, decision-log entries or "owner approved" lines. Once Paul decides, the docs fix below follows.
• README and CHANGELOG contradict the new exit code. README.md L352 and CHANGELOG.md L19 still say judge-down parks exit with "CLI exit 0", while the new table and the CHANGELOG L23 entry say 3. Update both lines (and the PR body's "exit 0 for waiting parks") to match the decided code.
• Fatal worker errors double-count partial usage. In src/loop/agentLoop.ts, the inner catch (around L514) adds partialRun.usage to usageSummary and then rethrows fatal errors. The outer catch (around L874) extracts and adds the same partial usage again. Add it in only one place: drop the add in the outer catch, or move the inner add below the fatal rethrow. Then add a test that a fatal AgentRunFailedError with usage is counted once in the run report.
• run-batch never emits the parked code. src/cli/run-batch.ts L250 still computes const exitCode: 0 | 2 = result.complete ? 0 : 2, so a batch whose loop parks exits 2. The README now says exitCode matches the CLI meaning. Either derive the batch code from per-loop statuses (3 when any loop parked and none failed), or document that batch only uses 0/2. Add a test either way.
• Misleading synthetic verify reason now reaches the next prompt. On transient faults without escalateModel, workerSdkVerifyResult (src/loop/loopWorkerRetry.ts L57) still says "Worker timed out or made no tool progress — loop stopped.", but the loop now continues. That text is pushed into priorFailures and fed to the next iteration. Reword it to something like "— retrying next iteration after backoff."

Nits

• When the primary judge fails and reviewSecondary is used as the fallback judge, invokeReview(secondary, …) still passes reviewSecondaryRuntime/reviewSecondaryModel. A gating verdict from the fallback can then trigger maybeRunSecondaryReview with the same agent reviewing itself. Clear the secondary fields in the fallback call's options.
• Judge errors drop partial usage. The judge retry/fallback catches in loopPostSuccessReview.ts don't call partialAgentRunFromError, so a Cursor judge session that fails after reporting tokens is missing from the new billing rows. Add it the same way the worker path does.
• Partial usage added on the transient path skips stopOnBudget, so a budget can be overshot by up to the transient cap. Call stopOnBudget(i) after adding partial usage in the worker catch.
• With sidecar now the default, confirm that the consumer-facing .gitignore guidance and any init template include .cursor/loops/**/verify-logs/, as this repo's own .gitignore does. If none exists, add a README line.
• TRANSIENT_AGENT_ERROR_PATTERN includes both status=error and Cursor agent run failed, so every Cursor status=error is transient. That's fine with the cap, but add a comment saying so, so a later edit doesn't widen fatal handling by accident.

[dancingteeth]

Owner decision — per Paul in chat (2026-10-21 16:32 Tbilisi):

1. verifyLogMode default sidecar: ship it. Note it in the CHANGELOG as a behaviour change, and add verify-logs/ to the consumer gitignore guidance.
2. Parked runs (status: waiting) keep exit code 2, the same as main 0.8.2. Don't change it to 3 (or 0). The README exit-code table, README line 352, CHANGELOG Unreleased line 19 and the PR body should all say 2 for parked, including judge-down parks. run-batch.ts already uses 2.

The other advisories from the sensor still stand: #45 (comment)

— Reviewey Ray (sensor; human owns merge)

[dancingteeth]

Reviewey Ray sensor (full Unified Code Review, triggered by HIGH risk; repo commit count 36). This is a sensor, not a merge decision: Paul owns merge. Agent loops may act only on ### Blockers.

Risk

HIGH

What could go wrong?

• This push puts the CLI exit-code contract back to main's behaviour (parked status: waiting exits 2), so if any doc or caller still says 3, wrappers that branch on $? would be misled.
• The review gate's fallback path changes: when reviewSecondary stands in for a failed primary judge, the merge-secondary pass is now skipped for that verdict.
• On the worker fault path, usage accounting and the per-iteration log moved around, and budget stops now run after transient faults.

Review depth

• Journeys at risk: agent-loop run and run-batch exit codes and AGENT_LOOP_DONE.exitCode; the review gate when judges error; worker fatal and transient fault handling; billing rows in run-report.md.
• Line-by-line: I read all of b5afa596 (18 files), plus the head src/loop/loopPostSuccessReview.ts judge retry and fallback block (L299–364).
• Empirical checks: I read the diff and head files only and ran nothing locally. CI test (22) was still pending at head b5afa596. CodeRabbit skipped because the PR is a draft.
• Reversibility: easy. Exit 2 for parks matches the 0.8.2 contract, so nothing new ships to npm consumers there.
• Verification gap: adjacent-only is closed for the prior advisories. There are new tests for single-counted fatal usage plus its log line, the transient retry reason text, the exit-2 mapping, the fallback skipping the merge secondary (at both layers) and the classifier edge cases. Nothing yet covers judge-error partial usage.
• Release guardrails: still a draft. The CHANGELOG entry is under Unreleased.
• Change set: 0f3fa037 → b5afa596, one commit: "fix(loop): parked exit 2, worker/review sensor fixes per PR feedback" (co-authored by Paul).

Pass 1b (REVIEWS.md overlay): These laws apply: "Verifier is the hard gate; review is residual." "Gating impacts: data-loss | security-boundary | false-closure | cross-dispatch | verify-bypass." "2b. Always — open cited callee before gating blockers; live path or [latent_contract]." Nothing below meets a gating impact. Judge-down still parks without closing the run as done.
Pass 2 (agent-authored): The commit's claims match its hunks. The owner decision (parked exit 2, sidecar default ships) is sourced to Paul in chat, as posted at #45 (comment), so it is resolved and not re-flagged. The PR body now says parked exits 2 and run-batch stays on 0 | 2.
§2b / §2c Lite: I opened shouldSkipSecondaryReview, workerSdkVerifyResult, the fatal and transient worker catch in agentLoop.ts, and the judge catches in runPostSuccessReviewPhase.

Prior sensor items (#45 (comment)):

• The owner decision on the parked exit code is resolved: it's 2, per Paul. The code, the 0 | 1 | 2 types in run, run-batch and notify, the README table and L352, and the CHANGELOG all now say 2. The 3 headline line was removed. Fixed.
• Fatal worker usage was double-counted. Fixed: the inner catch now skips the add when the error is fatal, and a test asserts exactly one usage row.
• run-batch and the parked code are moot under the decision. Batch already uses 0/2.
• The misleading "loop stopped" retry reason is fixed by the continuing flag, and a test covers it.
• Nit: fallback judge reviewing itself. Fixed with skipMergeSecondaryReview.
• Nit: stopOnBudget after transient partial usage. Fixed.
• Nit: verify-logs/ gitignore guidance. Fixed in the README and CHANGELOG, and src/cli/initGitignore.ts references it.
• Nit: comment on the broad transient pattern. Fixed. The fatal api key match is also narrowed to (missing|invalid|no) api key, with a test.
• Nit: judge errors drop partial usage. Still open (see below).

Verdict

PASS

Nits

• Judge errors still drop partial usage. In src/loop/loopPostSuccessReview.ts, the primary-judge catch (L313) and the fallback catch (L342) never call partialAgentRunFromError(err), so a Cursor judge session that fails after reporting tokens is missing from the billing rows. In both catches, add const partial = partialAgentRunFromError(err); if (partial?.usage) usageSummary = addUsageRecord(usageSummary, partial.usage), the same way the worker path does. Then add a test where the primary judge rejects with an AgentRunFailedError that carries usage, and assert the record is counted.
• Judge error classification only reads the top-level message. The same two catches use err instanceof Error ? err.message : String(err), while this push moved maybeRunSecondaryReview to formatErrorChain(err). A wrapped cause, such as a provider 5xx inside "Cursor agent run failed", won't reach classifyReviewJudgeError. Use formatErrorChain in both catches for consistency.
• LOOP_EXIT_PARKED now equals LOOP_EXIT_INCOMPLETE. Since parked and incomplete are both 2, the LoopCliExitCode union has a duplicate member and the two constants can't be told apart. Either drop LOOP_EXIT_PARKED and use LOOP_EXIT_INCOMPLETE with a comment, or keep it but note in the JSDoc that callers must read status (or AGENT_LOOP_DONE.complete plus reason) to tell a park from a verify miss.
• Wait for CI. test (22) was pending at b5afa596. Confirm it's green before un-drafting.

[dancingteeth]

Reviewey Ray sensor (automated code-review sensor, not a merge decision; human owns merge). Push 7085ddae on PR #45. Full UCR because risk is HIGH.

Risk

HIGH

What could go wrong?

• The judge-failure path in the review gate now records partial usage and classifies errors from the full cause chain. If usage were double-counted or classification flipped, park reasons and billing totals would be wrong.
• REVIEWS.md lists "secondary judge, review gate" under HIGH, and this commit edits exactly that fallback path in src/loop/loopPostSuccessReview.ts.

Review depth

• Journeys at risk: review gate retry, then secondary fallback, then park (status: waiting, exit 2); non-gated review failure (non-blocking success)
• Line-by-line: src/loop/loopPostSuccessReview.ts (both catch blocks), src/loop/agentLoop.ts exit constants, the new test
• Empirical checks: opened the callees partialAgentRunFromError / AgentRunFailedError (src/agents/agentRunFailed.ts), formatErrorChain (src/agents/errorFormat.ts), classifyReviewJudgeError (src/review/reviewJudgeError.ts), and the throw sites in src/agents/cursorAgent.ts (lines 160–185). Each failed attempt is a separate SDK run, so one usage record per attempt is correct and nothing is double-counted. formatErrorChain puts the top-level message first, so the timeout regex still matches, and transport text in a cause now correctly classifies as provider instead of other. Usage is recorded before the non-gated early return too. Tests not run locally; CI test (22) was queued at review time.
• Reversibility: easy
• Verification gap: adjacent-only (primary retry path is tested; the secondary-fallback usage path and cause-chain classification are not)
• Release guardrails: draft PR; parked exit stays 2 per Paul's 2026-10-21 decision, and LOOP_EXIT_PARKED is now an alias of LOOP_EXIT_INCOMPLETE
• Change set: base b5afa596 → head 7085ddae

Overlay (REVIEWS.md, Pass 1b): "Shared helper / multi-caller change → open callee once; assumption mismatch → gating blocker with file:line." Callees were opened and no mismatch was found. "Verifier is the hard gate; review is residual."

Agent-authored (Pass 2): co-authored commit. The intent comes from the prior sensor's nits (partial judge usage, formatErrorChain, duplicated exit constant), and all three are addressed.

Verdict

PASS

Nits

• In src/loop/loopPostSuccessReview.test.ts, the new test asserts records.length > 0 and totalInputTokens >= 1200. With two failing attempts it should produce exactly 2 records and 2400 input tokens, so assert those exact values to pin "one record per attempt, no double count".
• Add a test where the primary judge fails twice and the secondary also throws AgentRunFailedError with usage, and assert that all three usage records land in usageSummary (the fallback catch's recordJudgeErrorUsage(fallbackErr) is untested).
• Add a classifyReviewJudgeError(formatErrorChain(err)) test where the top-level message is generic but cause carries fetch failed / ECONNRESET, and assert provider. That locks in the behavior change this commit introduces.
• src/agents/cursorAgent.ts (around line 182) rethrows CursorAgentError as new Error(\Cursor SDK error: ...`)without{ cause: err }, so formatErrorChaincannot see the code or cause on that path. Pass{ cause: err }`.

[dancingteeth]

Reviewey Ray sensor (short sensor, risk LOW; repo commit count 38). This is a sensor, not a merge decision: Paul owns merge. Agent loops may act only on ### Blockers.

Risk

LOW

What could go wrong?

• Almost nothing at runtime. This push only touches three test files, adding text: '' to the AgentRunFailedError mocks because AgentRunPartialResult (src/agents/agentRunFailed.ts) now picks text as a required field.
• These are exactly the three TS2345 errors that failed pnpm typecheck on 7085ddae, b5afa596 and 0f3fa037, so the earlier PASS sensors on those pushes were given while CI was red on typecheck.

Review depth

• Change set: base 7085ddae → head 32582ffa (3 test files, +11/−3)
• Empirical checks: read the failed CI log for 7085ddae (three TS2345 errors at agentLoop.test.ts:602, loopPostSuccessReview.test.ts:379, loopWorkerAgentError.test.ts:42); each one is fixed by this diff. CI test (22) on 32582ffa was still running at review time.
• Verification gap: named-unrun (CI in progress)

Verdict

PASS

Nits

• Confirm test (22) goes green on 32582ffa before merge; if typecheck still fails, any other AgentRunFailedError(..., { usage }) call sites (non-test too) need text as well.
• The four nits from the 7085ddae sensor are still open: pin exact usage counts (2 records, 2400 input tokens) in the judge-retry test, test the secondary-fallback usage path, test cause-chain classification to provider, and pass { cause: err } in the CursorAgentError rethrow in src/agents/cursorAgent.ts.