Memory Browser

Live view of /home/dev/.claude/projects/-home-dev/memory/ — 83 memory files. Rendered 2026-10-22 10:52 · refresh

INJECTED INTO CONTEXT EVERY SESSION

MEMORY.md — the index

These 79 lines load into Claude's context at the start of every session. This is the budget to trim. Each bullet points at a full memory file below.

  • Dev Skills Dashboard shape — /opt/dev-skills-dashboard, Flask+gunicorn on 5001, per-task tmux sessions named task-<id[-8:]>
  • LingoClaw turn-review board build — per-turn feedback→isolated fix job→auto-merge into lingo backend, built into dev-lifecycle-board; branch lrb/lingoclaw (228 tests), e2e rig, sandbox/propose/gate design, Bence-gated deploy + deferred consolidation
  • Alt-Y chat component + sandbox — vanilla alt-y.js panel vs React /AltY; isolated mock-broker+Playwright sandbox at /home/dev/alty-sandbox (port 5210); broker on 5009 burns Claude quota; video→Gemini loop (creds depleted)
  • Alt-Y deploy branch + CI/CD (now automated) — live Alt-Y source is the ai branch not main; as of 2026-05-31 the dev→live loop auto-pushes (per win) + auto-rebuilds/restarts ai.idsys.ee (ai-northstar-watch.service + deploy-northstar.sh); long-running agents (mia/online/ybar) run in the Alt-Y cockpit (AgentLeftPanel), NOT the :5090 Agent Factory app; customer-live promote still manual
  • ai.idsys.ee superadmin login — toth.bence@idsys.ee / Test1234; headless-Playwright login to test auth-gated /AltY cockpit; save+reuse storage_state (login lockout risk)
  • New-customer React-native triad + build plan — isolated online+broker+alty envs (ports 5311/5312/5313, *-newcust units) for the editable React-native OneSheet doc build for NEW customers; plan at forge.../onesheet-plan/; Codex 5.5 worktree fan-out; alty git topology (doc port on origin/main, live ai on 65.109.84.219)
  • Codex CLI auth and session resume — ~/.codex/auth.json format, ~15min token cache, rollouts at ~/.codex/sessions for codex resume
  • OnSave write-back forward-sync gotcha — capture row values BEFORE trigger runs; resolveValueLinks:true on asset link forward-syncs empty values onto the row
  • dev.sh port allocation TOCTOU + flock fix — concurrent dev.sh start raced on port pick; added flock + /var/lib/dev-start-reservations
  • /compact EACCES on root-owned project dir — chown ~/.claude/projects/<slug>/ to dev:dev; resume via tmux task-<id8>
  • systemd Restart=always semantics — Restart=always only restarts on crash, NOT after clean systemctl stop. Pair with a watchdog if you need always-running.
  • Dev instance watchdog — dashboard auto-recovers stopped dev instances; pause via Instances tab or touch data/watchdog_disabled
  • Agent refusing /do-task on safeguard false-positive — symptom: "system reminder says I can't improve code"; fix is trust preamble already added to prompts; recover via send-raw nudge
  • Dashboard owns dev-instance lifecycle on ship — agent must NOT run dev.sh stop / SSH cleanup on ship; dashboard auto-tears down on shipped.flag; force refresh via GET /api/tasks/<id>
  • Self-improve cgroup gotcha + sandbox/skills config — KillMode=process so workers survive merge-triggered restart; sandboxes bind 0.0.0.0 with iptables 5101:5199; skills_sandbox/ snapshot lets self-improve edit ~/.claude/skills
  • testing.idsys.ee login lockout — ~5 attempts → 5min cooldown that extends on each retry; verify URL ≠ /Login before saving Playwright storage state
  • /regression-test skill — Playwright suite for testing.idsys.ee w/ proof overlays; template + SKILL.md at ~/.claude/skills/regression-test/; CMMSPRE doc-type cheat sheet
  • Testing customer DBs drift (toila, 1idsys) — recurring "Unknown column 'u.DocumentId'" UpdateService errors are pre-existing per-customer migration drift, not a new regression
  • Viewable browser for user login on headless server — Xvfb+x11vnc+websockify behind nginx HTTPS on a nip.io subdomain + patchright Playwright; let Bence finish OAuth/MFA/captcha steps mid-automation
  • Sun City Subs — security history — all 5 hardening items resolved 2026-05-02 (backdoor removed, SECRET_KEY rotated, /manage admin-only, db chmod 600, login rate-limit). Re-flag only if reverted.
  • Sun City Subs — A2P 10DLC IDs and resubmit recipe — brand/MS/QE sids; FAILED→DELETE+POST under same brand; MESSAGE_FLOW must reference public /sms-consent
  • Sun City Subs — deploy model + safe-test recipe — edit /opt/sun-city-subs + systemctl restart (git stale, don't commit); canonical domain suncitysubs.org; blank settings creds on a data.db copy to test w/o texting real instructors; admin assign/hand-over feature shipped 2026-06-17
  • ntfy.sh push to operator's phone — curl -d "..." https://ntfy.sh/agent-factory-bc7f3a — operator wants notifications anytime the AI judges they should know (autonomous loop milestones, high-conviction theses, blockers, blotter verdicts, anything ~1h-relevant)
  • Agent factory: OUTPUT not design — measure delivered output (landed math runs), not the design rubric. demand-supply-math-v1 is THE primary deliverable. 80/80 design = necessary not sufficient.
  • Codex/Claude usage: subscription framing — operator pays subscription; reason about cost as % of 5h+7d rate-limit windows, not $/call. Codex usage from ~/.codex/sessions/.../*.jsonl rate_limits field; Claude has /context for per-session.
  • Audit container state before touching it — when changing anything inside a stateful container (modal/form/wizard), read one ring outward: list sibling listeners and ask whether your change can perturb them. Data/Edit upload-clobbers-save bug was the trigger.
  • Hill-climb orchestrator (Claude vs Codex) — /home/dev/orchestrator/, port 8766, bubble panel UI; propose→score→plan→score→execute→summarize with ntfy on conf<7
  • Orchestrator Tree — bottom-up emergent task tree, https://tree.204.168.210.83.nip.io, /home/dev/orchestrator-tree/, port 8770
  • Orchestrator Tree rate-limit deadlock + fix — Codex usage limit pinned all 15 worker slots in auditing; fixed global-ratelimit reclassification + cooldown gate + wall-clock ceiling in app.py
  • Situational Awareness LP 13F watch — CIK 0002045724; option value field is NOTIONAL not premium (no 4-6× scale-up); aggregate by CUSIP[:6]; watch script at /home/dev/saw-13f-watch/
  • Default to background agents for anything non-trivial — operator wants the main convo responsive; launch agents for multi-file edits, research, feature builds; inline for quick clarifications
  • Rehvid InvoiceCustomer casing filter 500 (2026-05-27) — Root cause: MySQL collation mismatch (Orders.* utf8mb4_general_ci vs MetaItems.* utf8mb4_unicode_ci) in CASE+LIKE. Fix = EF.Functions.Collate(..., "utf8mb4_general_ci") in AgGridColumnService.cs for the four merged columns.
  • LingoClaw GCP access — story app live at story-webapp-o5tnurdf4q-uc.a.run.app; Cloud Run project phrasal-aegis-378922 (tothbenc60@gmail.com); access via SA key at ~/.config/gcloud-keys/
  • Atlas research orchestrator v2 — ground-up redesign of orchestrator-tree: top-down PLAN → MECE pillars → breadth-first → earned depth. /home/dev/atlas (stdlib backend :8780 + board/tree UI); validated; next = dock as alty-app React Canvas
  • LingoClaw subscription LLM CLIs — route LLM calls via subscriptions not metered APIs: agy=Gemini 3.5, claude -p=Opus 4.8; agy keyring auth replicated on agent-server; helper subscription_cli.py; Cloud Run can't use CLIs
  • Gemini via Vertex AI (multimodal works here) — AI-Studio Gemini key is credit-depleted + agy vision HANGS; use Vertex AI via LingoClaw SA. Helper /home/dev/tools/gemini-vertex/gemini.py (text/image/video) + gemini-rate skill. Rate UI similarity on a side-by-side VIDEO not screenshots (static conflates content w/ UX)
  • LingoClaw xAI image+video API recipe — funded key (…6fSMj); image grok-imagine-image; video POST /v1/videos/generations → poll GET /v1/videos/{id} until status=done → video.url; provider modes grok/grok_video (API) vs grok_browser (subscription)
  • LingoClaw mastery/selection logic — selection.py priority (overdue→reinforce-cap→new→free-play); patterns-before-vocab; reactive components (no curriculum) so all-correct=free-play every turn & review only after wrongs; real reason in logs.txt not /user/components; LINGO_DISABLE_MEDIA=1 for free dogfood
  • LingoClaw overseer meta-loop — two-tier self-improvement: hourly fix loop (cron :17) + overseer meta-loop (cron :47 every 6h) that reads trends and retunes the loop's prompt/schedule; lock recipe to run a manual fix session without fighting the cron loop
  • LingoClaw AI model/provider map — what model each function uses + the cost lever: phrase-gen & component-extraction burn a METERED hardcoded sk-ant Anthropic key (opus-4-8/4-7) when a subscription claude -p exists; in-app Gemini AI-Studio key depleted across TTS/transcription/image/voice-match fallbacks
  • LingoClaw engine philosophy + NO-determinism mandate — operator mandate: LLM-generated learner content must be right via DIRECTIVE not Python guards; no hardcoded examples; emergent patterns + reuse-when-same-skill (no regular/irregular tag); vocab credit on base form; remove the extractor's deterministic guards + rewrite prompt
  • LingoClaw jury harness tested WRONG story — playthrough.py mis-clicked the card when --story-text didn't exact-match the title → graded taquitos under a different label → bogus "theme mismatch" HIGH; fixed via story_id→exact-title map + fail-loud; verify the real story via storyType: in logs.txt
  • lingoclaw-walkthrough skill + public page — /lingoclaw-walkthrough plays a session → publishes per-turn public HTML (Q/A + targeted vs actually-used components + mastery bars) at https://lingo-walkthrough.204.168.210.83.nip.io; renderer render_walkthrough.py reuses the jury's transcript.json
  • FMP stable endpoints — v3 retired 2025-08-31 (403); use /stable/ for live prices & multiples
  • FMP MCP plan-gating + EDGAR fallback — FMP MCP gates 13F + statement line-items (ACCESS DENIED); metrics-ratios-ttm works & gives full margin waterfall; for 13F curl raw SEC EDGAR infotable.xml (WebFetch=403)
  • FORGE dev-factory concept deck — public deck at https://forge.204.168.210.83.nip.io/ ; src /home/dev/dev-factory-deck (xAI hero art + PIL-annotated real Alt-Y/AgGrid screenshots; static nip.io vhost)
  • /company-onepager skill — ticker in → deep plain-English investment one-pager out, via Atlas ecosystem harness + disciplined synthesis (the honesty rules from the META session)
  • AI-earnings-scan project — /home/dev/ai-earnings-scan; ranks non-sw/non-semi/non-supplier "adopter" stocks by AI earnings-call upside; consensus (Opus4.8+gpt5.5+blind referee) on impact+credibility, config-driven model tiers; :5188 plain python via nginx nip.io; run.py N + AES_* env flags; gotchas: local inspect.py shadow (run scripts from /tmp) + numeric DB cols MUST be REAL not TEXT; MODEL_CREDIBILITY="gpt" lever to scale; FMP has no options data; ran N=90 clean
  • thesis-dd skill + Atlas tabbed reports — stress-test a pasted thesis vs primary sources (5 refute-not-confirm lanes → classify claims → claim-check verdict, not a price target); ship one-pager+DD as tabs via Atlas report_tabs table at /read/<goal> + /reports; AAOI=goal 64
  • "The reports page" — = https://atlas.204.168.210.83.nip.io/reports (Atlas company-report index); a goal lands on it when goals.onepager>1200 chars; deduped by ticker; /read/<id> auto-tabs by ## sections; onepager-synthesis-verify workflow = the verified-synthesis upgrade of the one-pager skill
  • onepager-synthesis-verify workflow — the Claude Code workflow that upgraded /company-onepager: Atlas gathers, then draft→honesty lenses→revise→re-verify; publish from <dir>/final.md; MDT=goal 67 first run; v2 variant in company-onepager-v2/ adds provenance-ledger + profit-discipline lenses (9 total) + evidence-aware prep
  • gpt-5.6 family + model routing — sol/terra/luna via codex CLI (cursor-agent's gpt-5.6-* burn the CURSOR window — not a composer fallback); operator directive: bulk→composer→grok→gpt-5.6-terra, sol for high-IQ; proven grok/codex headless invocations; wired into Atlas run_model fallback chains
  • systemd manager env ZAI hijack — systemctl show-environment exports ANTHROPIC_BASE_URL→api.z.ai to EVERY service: plain claude -p under systemd errors 400 [1211] on real Anthropic model ids while shell repro works; scrub env per-service (Atlas does), don't unset globally
  • Micron-hunt pipeline — /home/dev/micron-hunt gather→extract→predict(Fable×Sol)→compute→verdict; four-factor test (demand step × FULL cost curve × pricing transmits × consensus hasn't done math); run-1 verdicts (VPG option, PRS lane, LD/HDS/THK priced past physics); PRS batch goals 119-124
  • SpaceX one-pager + custom IPO workflow — SpaceX is mid-IPO (SPCX, ~$135/$1.77T, xAI merged in Feb-2026); /read/70; hybrid Atlas+4 primary-source lanes; verify Atlas "disclosed" claims vs EDGAR; Starlink 63% margin bears no arm's-length launch cost
  • Datacenter buildout research — /home/dev/datacenter-buildout-research → datacenters.204.168.210.83.nip.io (deploy=copy to /var/www/datacenters); north-star = leading indicator for the semiconductor cycle (operational GW LAGS; watch leading-edge pipeline flows); engine.py + _render_charts.py/_render_fva.py; roadmap.html tab shell
  • AI buildout signal stack — COMPLETE — run done 2026-06-11, memo atlas /read/76; re-run recipe + harvest gotchas inside; YE-2026 due-dates to grade
  • ClickUp migrations chat recipe — workspace 14378329 / channel 6-901508819544-8; the one authorized ClickUp write; full recipe + bg-script worktrees on dev@65.109.84.219
  • Top Bar v4 prototype + integration plan — OneSheet topbar redesign (ClickUp 86ca3mk7f); fast loop at /home/dev/topbar-v4-proto (live http://topbarv4.204.168.210.83.nip.io); shared CSS at wwwroot/css/topbar-v4.css; integration = new SiteLayout (hooks mapped)
  • X/Twitter social API keys + status — keys in ~/.config/orchestrator-tree/secrets.env (bearer valid + FUNDED 2026-06-14, pay-per-use so may have re-depleted; recent-search works, FULL-ARCHIVE hit 402 CreditsDepleted 2026-06-23; use cashtag $TICKER, filter pump-spam); alt token dead (401); distinct from xAI/Grok; x-longform-posts skill curates thesis posts (fetch.py)
  • YouTube research + NotebookLM rig — this box's IP is YouTube-blocked (no yt-dlp/transcript-api/browser) → Gemini-watches-the-URL for video content; youtube-video-research + notebooklm skills; NEW youtube-transcripts skill = Gemini near-verbatim transcript via vid_transcript.py; NotebookLM driven via VNC login (vnc.204…nip.io) + CDP at /home/dev/tools/notebooklm/
  • ChatGPT Deep Research rig (from this box) — chatgpt-deep-research skill; run ~/.claude/skills/chatgpt-deep-research/run_remote_deep_research.sh "Q"; drives logged-in ChatGPT Pro in a VNC rig on 65.109.84.219 /opt/chatgpt-vnc over SSH (noVNC :6081 + PASSWORD.txt to fix login); ~10–30 min; remote-only
  • Memory browser dashboard — https://memories.204.168.210.83.nip.io/ organizes all injected memories (per-type, link graph, search, prune flags); code /home/dev/memory-browser/app.py + memory-browser.service (:8790)
  • AAOI research pipeline — /home/dev/aaoi-research/pipeline self-resuming source-discovery→one-pager→3-grader(Claude/Codex5.5/Gemini)→harder-rubrics loop; hourly cron+flock; tolerates codex rate-limit windows
  • autonomous-equity-research skill — /autonomous-equity-research <TICKER> generalizes the AAOI pipeline to any ticker (config-driven driver.py); globs box for seeds, validates, launches detached + wires cron
  • LingoClaw reteach root cause (corrected) — reteach_after_wrong HIGH is NOT buffer depth (buffer is 1-deep); real cause = selection patterns-before-vocab + free-play wrong turns matching only on hiddenWord; prefer_component_id queue-jump tried+reverted (fired 0× on free-play)
  • NT-BDC loan-level extraction run — 15-fund SOI pull done 2026-07-13; pipeline+deliverables in NonTraded_BDC_Extraction_Run/LoanLevel; SEC-HTML parsing gotchas
  • Powermode-lab acceptance contract — 99-criterion 9/10 dogfooding contract at /opt/powermode-lab/shared/acceptance/ on 65.109.84.219 (NB: came from run #673 being MIS-grounded; still valid for powermode-lab itself)
  • Automations-wizard acceptance contract — the CORRECT run-#673 deliverable: 47-criterion 9/10 dogfooding contract for the OneSheet automations wizard (base+MVP) at /home/dev/automations-wizard-acceptance/; MVP = OnDocumentCreate→conditions→agent(emit_fields)→propose-first fill; W13 = e2e money-shot journey
  • Autowiz build rig — autonomous executor of that contract: /home/dev/automations-wizard-build/, instance onesheet-autowiz (:8086 / autowiz.65.109.84.219.nip.io), 4 codex tmux loops + hourly Fable overseer cron (10 * * * *); STOP/DONE flags control it
  • Kunal identification-redesign run — full A–F sprint done 2026-07-14 in REVISED_GRANULAR_APPROACHES/FINAL_IDENTIFICATION_REDESIGN_V2; verdicts + locked-estimator/pair-reconstruction gotchas; new tests need approved addendum
  • LingoClaw test-runs-fable audit rig — ./run.sh → no-media per-turn before/after screenshot slideshow at lingo-fable.204…nip.io; media drop-in no-media-test.conf (remove to restore TTS/images)
  • Sun City Subs autofix loop — Carolyn email → sol-xhigh diagnose → auto-fix ≥80% or ntfy suncitysubs-bc7f3a; cron 6-56/10
  • Sun City Subs — notification channel policy — one message per notification (SMS only for classes ≤5 days out); shadowed datetime import silently killed the 2-day reminder
  • Sun City Subs — watchdog, backups, GitHub — private repo + deploy-key alias, local-only nightly DB backups, hourly Python tripwire + 4x/day Claude reviewer to ntfy suncitysubs-bc7f3a
  • Grok CLI privacy opt-out breaks video — bogus "ZDR must provide output.upload_url" 400 kills subscription image_to_video → silent still-image fallback; fix = /privacy opt-in in the grok TUI (not a team setting, not fixable in auth.json)
  • DLB local release/deploy recipe — board runs from /home/dev checkout+runtime (:5065/:5066); push to origin main is rejected; deploy via release.py build --deploy-testing + promote with DLB_TESTING_PROBE_URL=5066; frozen-asset digest sync + legacy JS twin
  • NT55 common-amount successor run — COMMON_DEBT_AMOUNTS_R3 NT55 release NT55R_3f171455d0a4b3632ba7 (2026-09-29); gotchas: root-owned project dir (sudo install -d), PYTHONDONTWRITEBYTECODE on baselines, gzip mtime breaks byte replay, iXBRL footnote DOM
  • Kunal multi-account sessions — kunal-session launcher: 18 tmux sessions pinned to ~/.claude-k{,2,3} / ~/.codex-k{,2,3}; never use default ~/.claude/~/.codex for Kunal
Show all ↓

Pruning heuristics: orphan = not in MEMORY.md and not referenced by any [[link]]. stale = not modified in 45+ days. These are hints, not verdicts.