Add CI (config/script validation, no auto-deploy) - #30
Open
sergiocoding96 wants to merge 28 commits into
Open
Conversation
- tower/docs: architecture, agent access, Discord permissions, Krati setup - tower/scripts: provision Krati, align personal agents, Discord hub fixes - tower/SERGIO-CHANGELOG.md: summary of live sergio config (no secrets) - .gitignore: agents-auth.json and perf audit artifacts
Adds CloakBrowser-backed scraping service at localhost:9378 as the primary stealth-scrape path for DataDome / Cloudflare Turnstile / Imperva-protected domains. Camofox stays for the agent's interactive browser tool (snapshot/click/type) until Phase 2 mirrors those endpoints. Why --- Existing stack failed on DataDome targets (Idealista, Ticketmaster, Glassdoor): Firecrawl playwright-service uses stock Chromium that gets blocked, and the Camofox interactive service was silently broken (better-sqlite3 native module crash, health check didn't catch it). Benchmarked Cloak vs Camoufox vs stock Playwright; Cloak passes the same fingerprint checks Camoufox does at lower latency and RAM, and exposes a Playwright-compatible API that lets us collapse the dual stack onto one engine. What ships ---------- - tower/docs/browser-stealth-benchmark-2026-05-16.md Full benchmark methodology + numbers (engine launch / per-URL DCL / peak RSS / Sannysoft / Cloudflare behavior). - memos-setup/learnings/2026-05-16-cloak-deprecate-camofox.md Decision doc: why this change, what shipped Phase 1, what's deferred to Phase 2, rollback path, cost projection. - skills/web-research/SKILL.md New §4 "Cloak Stealth Service" + domain routing table updated to send DataDome / Turnstile / Imperva domains to localhost:9378. - CLAUDE.md Architecture, "when to use which tool", and commands sections updated to reflect new stack. Health-check commands added. Deliverables / evidence ----------------------- - perf-audit-results-idealista/marbella-top20-analysis.md Worked example: 20 Marbella properties scraped via Cloak, per-property analysis, sub-area heatmap, value/trophy picks. End-to-end proof the new pipeline works on a real DataDome target. - perf-audit-results-idealista/marbella-top20-raw.json Structured data behind the analysis (reproducible). - perf-audit-results-idealista/idealista-estepona-page*.pdf PDF renders of the first 2 Estepona listing pages. - perf-audit-results-idealista/idealista-estepona-page*-CHALLENGE.html Evidence of DataDome captcha-mode state from the IP-burn portion of testing; kept as documentation of failure mode. Service infrastructure (on Tower host, not in this repo) -------------------------------------------------------- - /home/openclaw/.hermes/cloak-service/cloak_service.py (~280 LOC FastAPI service: per-domain persistent context, request pacing, asset blocking, DataDome detection, CapSolver fallback) - ~/.config/systemd/user/cloak-service.service (enabled, auto-restart) - ~/.config/systemd/user/cloak-service.service.d/capsolver.conf (CAPSOLVER_API_KEY env, chmod 600, gitignored by being outside repo) Interim fix bundled ------------------- Camofox node_modules/better-sqlite3 was failing to load (native module not registered), making /tabs return 500 silently while /health reported ok. Rebuilt via `npm rebuild better-sqlite3` on the Tower host; service restored. Phase 2 retires Camofox entirely. Phase 2 (not in this PR) ------------------------ Mirror Camofox's /tabs/* interactive endpoints (snapshot, click, type, scroll, press, back, screenshot) in the Cloak service, then flip CAMOFOX_URL to point at port 9378 and disable the Camofox systemd unit. Documented in the decision doc. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ox deprecated to warm-rollback Same-day continuation of #29. Cloak service (`localhost:9378`) now exposes the full Camofox HTTP surface used by `tools/browser_camofox.py`, and `CAMOFOX_URL` env has been flipped to point at it. All gateway services restarted and confirmed active. What's now live on port 9378 ---------------------------- - POST /tabs create tab + optional initial nav - POST /tabs/:id/navigate go to URL - GET /tabs/:id/snapshot ariaSnapshot YAML + [eN] ref markers - POST /tabs/:id/click click by ref (auto-refresh on stale) - POST /tabs/:id/type type into element by ref - POST /tabs/:id/scroll mouse-wheel up/down - POST /tabs/:id/press keyboard key - POST /tabs/:id/back go_back + refresh refs - GET /tabs/:id/screenshot PNG (base64) - DELETE /sessions/:userId close all tabs + context Plus the Phase 1 surface (/v1/scrape, /v1/save-pdf, /v1/balance, /health) which is unchanged. Format-compatibility verification --------------------------------- Snapshot output diff-tested against Camofox on the same URL at the same wall-clock: - Cloak: refsCount=222, totalChars=41004 - Camofox: refsCount=222, totalChars=41004 - 182-line diff was 100% content drift (HN updated vote counts and timestamps between fetches), 0% format drift - `e15` etc. resolve to the same DOM nodes in both engines End-to-end agent flow tested: POST /tabs -> GET /snapshot -> POST /click(e3) -> GET /snapshot -> POST /navigate -> POST /back -> GET /screenshot -> DELETE /sessions — all green. Cutover state ------------- - ~/.hermes/.env backed up to ~/.hermes/.env.pre-cloak-flip - CAMOFOX_URL flipped from http://localhost:9377 to http://localhost:9378 - Restarted: hermes-gateway, hermes-gateway-arinze, -hr-agent, -krati, -research-agent, -sergio — all `active` - Camofox systemd service still running (warm-rollback only, no agent traffic). Stop+disable scheduled after ~1 week of clean operation. Rollback -------- ``` sed -i 's|^CAMOFOX_URL=http://localhost:9378$|CAMOFOX_URL=http://localhost:9377|' ~/.hermes/.env systemctl --user restart hermes-gateway hermes-gateway-arinze hermes-gateway-hr-agent hermes-gateway-krati hermes-gateway-research-agent hermes-gateway-sergio ``` Service source (not in this repo — lives on Tower) -------------------------------------------------- - /home/openclaw/.hermes/cloak-service/cloak_service.py (Phase 1, scrape) - /home/openclaw/.hermes/cloak-service/interactive.py (Phase 2, ~450 LOC, new) - INTERACTIVE_ROLES + SKIP_PATTERNS + MAX_SNAPSHOT_NODES constants kept in sync with Camofox server.js lines 98-108, 256 - _build_refs_from_yaml mirrors Camofox _buildRefsInner (lines 1313-1389) - _annotate_yaml_with_refs mirrors Camofox annotation (lines 1933-1965) - _ref_to_locator mirrors refToLocator (lines 1412-1421) - _window_snapshot mirrors lib/snapshot.js windowSnapshot Phase 3 (deferred) ------------------ After ~1 week clean: stop+disable camofox.service, mark tools/browser_camofox.py deprecated in module docstring, optionally move Cloak service source into this repo under tower/services/. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ackend
Triggered after Phase 2 E2E was green; user opted out of the warm-rollback
week. Camofox service stopped + disabled. Cloak now serves 100% of the
agent's web stack (stealth scraping + interactive browser tool).
Tower-side changes
------------------
- `systemctl --user stop camofox.service` (0 active tabs/sessions at stop time)
- `systemctl --user disable camofox.service` (won't restart on Tower reboot)
- Port 9377 released
- `~/.hermes/hermes-agent/tools/browser_camofox.py` module docstring
rewritten as deprecation notice — file name retained because renaming
would touch 17 call sites in `browser_tool.py`; HTTP calls land on
Cloak via `CAMOFOX_URL=http://localhost:9378`
- Camofox install left at `node_modules/@askjo/camofox-browser/` for
reference / future rollback (~140 MB; remove only if disk pressure)
Repo changes (this commit)
--------------------------
- CLAUDE.md: Camofox bullet updated to "retired", clarifies the
browser_camofox.py name retention
- skills/web-research/SKILL.md: Phase 3 status block
- memos-setup/learnings/2026-05-16-cloak-deprecate-camofox.md: full
Phase 3 section appended (what landed, post-stop verification,
re-enable recipe, deferred Phase 4 considerations)
Verification after Camofox stop
-------------------------------
- Cloak service still active
- All 6 hermes-gateway services still active
- Full E2E flow (`/tmp/e2e_cloak_as_camofox.py`) re-run with Camofox
dead: 9/9 endpoint calls passed
- `tools.browser_camofox` Python symbols still importable; with
CAMOFOX_URL=http://localhost:9378 in env, `is_camofox_mode()` →
True and `check_camofox_available()` → True
Re-enable Camofox if a regression appears
-----------------------------------------
```
systemctl --user enable camofox.service
systemctl --user start camofox.service
sed -i 's|^CAMOFOX_URL=http://localhost:9378$|CAMOFOX_URL=http://localhost:9377|' ~/.hermes/.env
systemctl --user restart hermes-gateway{,-arinze,-hr-agent,-krati,-research-agent,-sergio}
```
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
15-slide widescreen deck summarizing the same-day Camofox → Cloak migration for sharing with the team. Covers: executive summary, problem state, benchmark methodology + results, DataDome IP-score discovery, architecture decision, cost & scaling tiers, Phase 1/2/3 implementation, final architecture, Marbella worked example, ops state + rollback recipe, and links. Generated with pptxgenjs (Node) using a custom slate-900 / cyan accent design system. Visual QA pass via subagent on rendered JPGs caught 4 high-severity layout issues (title wrap collision, architecture-row footer overflow, table row clipping, glyph rendering) and 1 cosmetic imbalance; all 4 fixed and re-verified before shipping. Files: - tower/docs/cloak-migration-brief-2026-05-17.pptx (684 KB, editable) - tower/docs/cloak-migration-brief-2026-05-17.pdf (314 KB, viewable) Source generator script kept at /tmp/cloak-report/build/generate.js (not in repo — one-shot artifact; rerun by re-installing pptxgenjs + react-icons + sharp deps). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ent vision
Adds 7 new slides (15–22) extending the migration brief into a strategic
"what this enables" Part 2. The original 14 Part 1 slides about the
Cloak migration are unchanged.
New slides
----------
15. PART 2 divider — From plumbing to product
16. The product — discovery agent for property investors (who/what/why)
17. Pipeline + monetization — Tier 1 free baseline + Tier 2 Nota Simple pay-per-use
18. The data moat — what accumulates day 1 / week 2 / month 1+ / month 6+
19. The embedding moat — why specialized property embeddings matter
(with concrete BGE-vs-trained-encoder comparison)
20. Roadmap — Phase 0 done / Phase 1 next / Phase 2 history / Phase 3 embedding
21. Optionality — three buckets of future business models the data unlocks
22. Links + current state + deferred — replaces old slide 15
Cover subtitle updated to bridge the two parts.
QA: independent subagent visual review caught 5 critical layout issues
(decorative shape colliding with footer, title-wrap collisions on 2
slides, eyebrow mid-word break, footer-text collision) plus 2 minor
cosmetic issues. All 7 fixed and re-verified before shipping.
Files
-----
- tower/docs/cloak-migration-brief-2026-05-17.pptx (923 KB)
- tower/docs/cloak-migration-brief-2026-05-17.pdf (416 KB)
The PPTX source generator is at /tmp/cloak-report/build/generate.js
(one-shot artifact; not in repo).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…dal embeddings, storage cost, meaning-vs-pixels, embedding-vs-LLM
Adds 6 new slides + refreshes the final state slide. Deck now 28 slides
total (was 22). The new content captures the strategic conversation
since the v2 push.
New slides (20-25 in the 28-slide deck)
---------------------------------------
20. Wayback bootstrap — real coverage (~13M snapshots = ~3-5M unique
listings, ~10-25% national coverage) + legal posture (EU TDM
exception, Idealista litigation history, recommended hybrid path).
21. Data foundation after 6 months — 2-3M unique listings, ~50M photos,
per-city annual flow breakdown across top 10-12 Spanish cities.
22. Multimodal embeddings — text-only vs CLIP-class side-by-side,
SigLIP 2 / OpenCLIP / PaliGemma / Qwen2.5-VL as recommendations.
23. Storage cost comparison — raw photos vs text embedding vs
multimodal per-photo / per-property / hybrid. Hetzner pricing.
Compute panel: ~$15/mo + one-time $300-500.
24. Meaning vs pixels — LOST vs PRESERVED table + the photo-URL-not-
photo insight (link Idealista's CDN, don't host).
25. Embedding vs LLM — catalog vs analyst metaphor. Why fine-tuning
Gemma is the wrong frame for retrieval. Optimal stack = both.
Updated slides
--------------
- Cover: subtitle bridges Part 1 / Part 2
- Roadmap (slide 26, was 20): Phase 2 + 3 refined with Wayback +
multimodal embedding language
- Final state slide (slide 28): adds Wayback Phase 2, Multimodal
Phase 3, lawyer review + Idealista B2B license in Deferred
QA: subagent visual review caught 4 critical issues (title-wrap
collisions, footer overflows) + 2 cosmetic glyph issues (fl ligature,
≠ character). All fixed.
Files
-----
- tower/docs/cloak-migration-brief-2026-05-17.pptx (1.2 MB, editable)
- tower/docs/cloak-migration-brief-2026-05-17.pdf (486 KB, shareable)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…nges
Strategic direction change captured per CLAUDE.md governance rule:
- Updated CLAUDE.md active-sprint header to reflect v2-only architecture
(Paperclip/CEO retired, v1 MemOS server stopped, single-machine, single
daemon, row-level namespace tenancy, BGE-large embedder)
- New decision doc: memos-setup/learnings/2026-05-17-v2-only-bge-shares.md
- Architecture decks in docs/architecture/:
* 2026-05-17-memory-system-brief.{pdf,pptx} (merged canonical brief — 34 slides)
* 2026-05-17-memory-system-decisions.{pdf,pptx} (source A: architectural review)
* 2026-05-17-session-summary.pptx (source B: one-day overhaul summary)
* 2026-05-17-memory-system.{pdf,pptx} (intermediate consolidation draft)
- Supporting artifacts: perf-audit harness + results, v2 data-integrity audit
report, memory-viewer / UMAP-viewer screenshots used as evidence
- Operational scripts: tower-upgrade-hermes-runtime.sh, worktree initiation prompt
- .gitignore: exclude .playwright-mcp/ MCP scratch dir
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ools/
Trackable artefacts from the 2026-05-17 v2-only memory overhaul that had been
sitting only on the box because tools/ was blanket-ignored.
Headline changes:
- tools/memory-map.html — Memory Viewer v3 (D3 force-directed light theme,
fill=kind / border=owning-agent, topic hubs as DeepSeek-labelled clusters)
- tools/memos-explorer.py — graph-export with HDBSCAN + DeepSeek cluster labels
- tools/memory-graph-watcher.py — WAL-driven debounced auto-refresh daemon
- tower/services/memory-graph-watcher.service — systemd unit template for it
- tools/plugin-patches/ — staged v2 plugin patches (5 logical patches across
13 files; postinstall-patches.sh re-applies + verifies after npm update)
- tools/d3.v7.min.js — local d3 dep (d3js.org was unreachable from the box)
Gitignore: keep tools/ tracked; ignore only generated data + screenshots
(memory-graph.json, vec-map(-3d).{json,png}, screenshots/). Also tighten
repo-root ignores for playwright test PNGs and perf-audit JSON dumps.
Decision doc: extend memos-setup/learnings/2026-05-17-v2-only-bge-shares.md
with the Viewer v3 design, cluster-labelling pipeline, and watcher unit
docs. Drop the duplicate "Cluster summarisation + auto-refresh watcher"
section now covered upstream in the Memory Viewer section.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the 2026-05-21 decision doc and updates the active-sprint header for the new CTO agent (lives in ~/Coding/Hermes-CTO, joins the v2 plugin as the `cto` profile). Re-introduces peer cross-agent interaction but keeps the v2-only stance: no central orchestrator, no cross-machine routing. Also documents the dirty-episode cold-boot fix (7 episodes marked reward.skipped) that resolves the slow bridge boot behind the process leak. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…rable daemon Root causes (2026-05-26): (1) a non-idempotent capture retry re-inserted the same turns tens of thousands of times across multiple agents (95k rows for ~320 real turns; arinze session cost only ~$0.01 in tokens — the damage was DB bloat), and (2) each gateway bridge cold-loads BGE-large; concurrent boots starved the CPU past the 30s session.open timeout → respawn/leak/CPU spiral. No data was lost; the v2 store simply only goes back to its 2026-05-12 birth. This change: - Purge + global dedup of all runaway episodes (95,174 → ~320 traces), FTS rebuilt, VACUUM (1.8GB → 566MB). [DB ops, not in tree] - traces_idempotent_turn BEFORE INSERT guard (RAISE(IGNORE)) so a duplicate (episode_id,turn_id) insert is silently skipped — structural fix for the re-insert loop. [applied to DB; documented in decision doc] - Raise bridge timeouts in memos_provider (session.open 30→120s, reconnect 4→90s) so a cold boot finishes before the fatal timeout. - Replace the dangerous `pkill -f 'bridge\.cts'` band-aids (killed the :18800 daemon + matched admin shells) with scripts/lib-bridge-safe-cleanup.sh (preserves the daemon, targets only node bridges) + staggered restarts. - Make the :18800 daemon a systemd --user unit (was a fragile manual SSH-session process). - Fix viewer console errors in hermes-profile-switcher.js (mixed-content guard for the cross-daemon probe + auth gate before diag/namespace). - Reconcile CLAUDE.md: extraction + skillEvolver LLMs are DeepSeek (reverted from the Gemini stopgap on 2026-05-25), not Gemini. - Wire this Claude Code session to memos under a 'claude-code' profile (.mcp.json; the .claude/settings.json hook config is gitignored — contents recorded in the decision doc). Decision doc: memos-setup/learnings/2026-05-26-capture-outage-rca-recovery.md Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The 3D map intentionally renders only skills/policies/topics; individual memories were folded into topic clusters. Add a 'Memories' checkbox (default OFF so it doesn't re-noise the view) that plots trace nodes as small dots, with legend + styling. Frontend-only; web/dist symlinks to this file. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…uction + WM embedding-merge Map (tools/memory-map.html, memos-explorer.py): - reward-heat: explorer exports per-trace value/priority; map colors Memory dots red→grey→green by value (heatColor). - bubble drill-down: clicking a Topic lists its member memories; click one → detail. Induction #16 (l2-induction.ts, l2/induce.ts, l2/types.ts): - L2 prompt now classifies experience_type (success_pattern|failure_avoidance) from the supporting traces' outcomes/value and emits anti_patterns; induce.ts persists experienceType/evidencePolarity + seeds decisionGuidance.antiPattern. Fixes the 155:1 success:failure imbalance — negative episodes induce avoid-policies. World-model merge #18 (l3/merge.ts): - gatherMergeCandidates now also gathers by embedding similarity (cosine ≥ 0.86), not only policy-overlap/domain-tag, so semantically-identical world-models from different agents merge instead of creating near-duplicates. Decision threshold raised to the strict dedup cutoff (clusterMinSimilarity=0.3 was too loose). Validated: daemon boots clean (:18800 200), stress-test 34/36 (2 fails are the idle-stopped Firecrawl/SearXNG docker stack, unrelated). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…acking (#17) #15 reward (core/reward/backprop.ts): - error_signatures are captured per trace but were never a reward term — a step that hit tool failures/loops scored the same as a clean one. backprop now subtracts a 0.35 outcome penalty from V for any trace with non-empty error_signatures, pushing failures below neutral (lower rank + seeds failure_avoidance induction). The LLM R_human scorer already covers episode-level satisfaction, so this adds the missing objective signal. #17 skills (core/config/defaults.ts, core/pipeline/memory-core.ts): - crystallization bar: skill.minSupport 1→3 (a pattern must recur across ≥3 episodes, not crystallize from one) + minGain 0.02→0.03. Fixes the hyper-specific one-off skills that never re-matched. - usage tracking: record skill usage when retrieval surfaces a skill into the prompt (was only bumped by explicit getSkill → ~all at 0). Verified: usage_count sum rose 2→8 across a stress-test run. Validated: daemon boots clean (:18800 200), stress-test 34/36 (2 = idle Firecrawl/SearXNG docker, unrelated). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Comprehensive 2026-05-26/27 session report: outage RCA + recovery, memory-quality audit, the 6 quality fixes (#14-#19), and what works / what doesn't / what's left. Rendered 16:9 PDF via Chrome headless from the HTML source. A Google Doc copy of the same content was uploaded to the operator's Drive folder. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the failure-induction (l2-induction/induce/l2-types), WM embedding-merge (l3/merge), outcome-reward (reward/backprop), and crystallization-bar (config/defaults) files to the patch FILES + MARKERS lists so they survive an npm reinstall and --check verifies them. All 22 markers present. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Full memory-system story Apr→May 2026: multi-tier origins (Paperclip/CEO + MemOS v1), the v1/v2 pivots, the 2026-05-17 single-tier overhaul (BGE embedder, share-scope policy, orphan-cron root cause), the 22 plugin patches, the LLM provider saga, the CTO agent, and the 2026-05-26/27 outage recovery + quality overhaul — with current state and roadmap. Brings the 2026-05-17 brief current. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Roadmap for post-recovery work (Phase 1 stability → Phase 2 feedback loop → Phase 3 shared daemon). Adds tools/build-memory-report.sh to regenerate the comprehensive report PDF from its HTML source, and records the standing process that every memory change updates that report. CLAUDE.md header marks the 05-26 pending items as done and points to the roadmap + report. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Each gateway spawns its own bridge.cts that cold-loads BGE-large (~2 cores); concurrent boots starve the CPU past session.open → respawn → leak/CPU spiral (the recurring sergio failure). Add bridge_boot_lock() in daemon_manager.py — an flock on ~/.hermes/memos-plugin/.bridge-boot.lock held around the spawn+open in _reconnect_bridge so only ONE boot runs at a time (best-effort: proceeds after a timeout rather than blocking capture forever). Reusing a warm bridge stays fast. Verified: 3 gateways restarted simultaneously → boots serialized (1 at a time, load capped ~4 vs 13+ before), no spiral, :18800 healthy. (Phase 3 shared-daemon will remove this lock.) Also fixed the postinstall footer's dangerous pkill-based restart hint to use the systemd unit + staggered gateway restarts. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First application of the standing process: a memory change (P1.2 flock) updates the comprehensive report. Moved serialize-boots from 'what's left' to 'what works', noted per-gateway model loads remain (Phase 3), rebuilt the PDF via tools/build-memory-report.sh.
…ack SKIP P1.3: tools/prune-stale-skills.sh archives (reversible, never deletes) skills with usage_count=0 + null last_used + age>grace(21d) + low gain. Hides them from retrieval (tier-1 = active+candidate only). DRY-RUN by default; meant to run weekly so skills accrue usage data (now tracked) before being judged. Archives 0 today (all skills <21d old) — correct. P1.4: stress-test marks Firecrawl/SearXNG as SKIP (not FAIL) when the on-demand stack is idle-stopped → 34/34 GREEN, 2 skipped (was '2/36 failing'). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
init() awaited recoverDirtyClosedEpisodes() before the HTTP server started — one dirty episode's reflect took 154s, leaving :18800 unavailable during the worst cold boots (and starving concurrently-booting bridges). Now fire-and-forget: the reflection backfills reward/value after startup while capture/health serve immediately. Self-capped (the recover fn is sequential). The periodic autoRescoreDirtyClosedEpisodes() path is unchanged. Verified in the boot log: server.started now fires the instant the 9-episode rescore is kicked off; reflect.done events land AFTER server.started. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Boot-reflection now backgrounded (P1.1), boots serialized (P1.2), dead-skill auto-prune (P1.3), stress-test green/skip (P1.4). Rebuilt PDF.
… P2.1 deferred) Verified the feedback loop is more wired than assumed: the LLM R_human judge already sees every user/agent turn pair (buildTaskSummary), so P2.2 (implicit next-turn signal) is redundant — dropped to avoid double-counting in reward code. Tool failures already feed error_signatures (P/#15 penalty) and the subscriber auto-schedules runRepair on bursts; the gap is the adapter not calling recordToolFailure + no feedback bridge RPC — a cross-layer change deferred to a focused, tested session (learning-corruption risk). Auto-prune (P1.3) marked done in the report.
Gateways → thin HTTP clients of the one :18800 daemon (one BGE-large total), retiring per-gateway bridge.cts. Enablers already shipped: per-request namespace (patch #2 X-As-Profile) makes the daemon multi-tenant; --daemon self-serves LLM. Gap = expose the capture write path (session.open/turn.start/turn.end) over HTTP (reads + memory/search already are). Covers auth (loopback bypass), adapter HTTP transport behind MEMOS_TRANSPORT flag, namespace correctness, migration/rollback, risks, verification. Removes the P1.2 boot-lock once landed.
First safe, additive slice of the shared-single-daemon work (roadmap Phase 3, spec memos-setup/learnings/2026-05-27-p3.1-shared-daemon-spec.md). - New server/routes/turns.ts: POST /api/v1/turn/start -> core.onTurnStart, POST /api/v1/turn/end -> core.onTurnEnd. Registered in registry.ts buildRoutes. (session/episode open/close already existed.) Lets the :18800 daemon accept the capture WRITE path over HTTP, the prerequisite for converting gateways into thin HTTP clients (one BGE-large load total). - Namespace correctness (spec work item 4): namespaceFor() in core/pipeline/memory-core.ts now falls back to effectiveNamespace() (request-ALS aware) instead of bare activeNamespace, so X-As-Profile is honored on writes. Outside an HTTP request getRequestNamespace() is undefined -> resolves to activeNamespace exactly as before (stdio bridge unaffected). Without this, a turn written over HTTP landed under 'default'. Auth: these routes are gated by the existing viewer session check, NOT added to the public allowlist — a loopback bypass is unsafe because Tailscale Serve proxies tailnet -> 127.0.0.1. The per-boot internal token for the adapter + the HTTP-client transport cutover (behind MEMOS_TRANSPORT) remain the next steps. Verified: login -> open session/episode -> turn/start + turn/end with X-As-Profile: claude-code -> trace lands owner_profile_id='claude-code', share_scope='private'. stress-test 36/36 green; postinstall-patches.sh --check clean (now 26 tracked files, markers registerTurnRoutes / 'POST /api/v1/turn/start'). Comprehensive report deck + PDF updated (standing process). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
deploy/install.sh overwrites live fleet config, so this repo intentionally has no deploy workflow. CI validates YAML, Python syntax and shell scripts. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds a CI workflow that validates deploy YAML, Python syntax and shell scripts.
Deliberately no deploy workflow:
deploy/install.shoverwrites the live~/.hermes/config.yamland profile SOUL/config files, and this repo's snapshot has drifted from the running fleet — auto-deploying would revert the fleet's model configuration.🤖 Generated with Claude Code