Conversation
…EGISTRY.json sweep after main registry, local wins collisions (todo #132, DPLAN-0290 night item 0). Live failure was watchdog agent @baud = agent-not-found post-admin-lane; now resolves. 2 tests red-first w/ world pinned to tmp_path (todays CI lesson applied), suite 488 green, ruff clean, live-proven
Codecov Report❌ Patch coverage is 📢 Thoughts on this report? Let us know! |
…pos (FPLAN-0403), trigger reload sentinel + on() idempotent (FPLAN-0404), memory date_guard snapshot drain (DPLAN-0283), flow+prax quick_status ownership merge (FPLAN-0405, Patricks zero-todo card). All owner suites green (1019/957/1063/787/1240), all audits 100%, every fix red-first, sequential one-at-a-time per Patricks ruling. Full record in DPLAN-0290
… 100% (Patricks directive, waves of 2). Living APLANs planted fleet-wide (standing health records, never recreated). Rule E fixed everywhere red-first: help flag anywhere explains, never executes - live detonations incl memory 17-branch reset, backup real snapshot, drone file delete on --help probes. seedgo help_flag_safety standard born + recalibrated 3x same day on fleet evidence (router exemption removed, value-slot gap closed, log_structure de-scored: it read runtime state - a score that moves on its own is worse than a wrong one). Bypass registries measured control-first (cli 21-8, drone 48-29, flow 74-58, ai_mail 51-17). 6 branches had tests pinning broken behaviour, fixed. Commit gate itself fixed in-train: template .trinity skeleton made the gate run pytest INSIDE spawns template, planting the cache spawns hygiene tests reject - outermost-citizen-wins mapping, 3 tests red-first, drone 1073 green. Final proof: 17/17 at 100%, 45 standards 100%, 0 type errors. Full record in DPLAN-0291
…caught it. Root .gitignore ignores .spawn/ everywhere; only aipass_framework had the template exception, so project_agent/.spawn/.template_registry.json existed locally but never shipped - fresh clones could not stamp project_agent citizens, and spawns new hygiene test FileNotFoundError'd on every CI python lane (green locally, red in clean checkout - the S230 species). Two whitelist lines matching the existing aipass_framework idiom; registry is a machine-written file manifest, no secrets. Hygiene suite 6 green
… Patrick-present + live-verified. Real cause found by @hooks in CC transcripts (205k lines to 07-09): auto_process timeout 120 lived ONLY in deployed settings, never the manifest - a reconcile ~08-02 wiped it silently, and its legit 78-120s first-prompt run (memory vectorize+rollover) died at the 30s default = silent context loss. BAUD acquitted: timeouts start 07-15, worst day 07-28, zero in BAUD cwds. Fix stamps timeout 90 (auto_process 120) in ALL layers same breath: hooks.json inner + provider manifest + init template (init update union-merge wipes live-only values), 3 config-pinning tests + engine test proving >30 reaches worker.join uncapped, trust re-enrolled with the edit (hooks.json edit breaks trust hash - @hooks 8-min probe outage, reverted+reported). README single-dispatch claims fixed (3 sentences, false 73+ days). Suite 1454 green. Deployed ~/.claude/settings.json stamped to match (not in repo). Follow-up: DPLAN-0294 relocates auto_process off the prompt lane (Patricks ruling, compass #272)
…ep, deletion record, fragments park. NIGHT (all 10 roster items, sequential, watchdogged, zero commits till review): auto_process off the prompt lane (memory detached child + hooks shell swap, 0.7s handler vs 78-120s block, live-proven on real rollover); UPS sweep 11 injectors/2 workers, user_message_relay = candidate 1; 17/15 class WARNING to INFO at emitter w/ guard (16 pct of all escalation mail ever); system_logs double-watch root-caused as ONE reader TWO FILES (prax dual-write) not two watchers - twin detection + live-tree names, UNKNOWN mystery solved (11-name list vs 17 branches); seedgo log_structure NEVER RAN in any commit (TypeError under bare except, mocks accept any kwarg) - crashes now surface, observe-mode ships w/ negative verdict; drone default 30 to 60s one constant three layers, --drone-timeout placement tested; prax Mission Control resolves projects/* from registries+passports (BAUD scopes; misattribution + CWD-relative bugs fixed beneath); prax measured own prompt-lane species: first log call 0.60s on every drone command - proposals filed. Four self-corrections amendment-before-hardening (compass 276). MORNING (Patrick rulings): deletion record everywhere - drone rm all outcomes to .ai_central/deletions.jsonl + prax line, passport caller, env cannot silence audit; hooks rm_gate records every raw rm seen (~10ms rm-lane only); spawn could never delete own templates/ - outermost-citizen-wins fix; symbolic fragments tier PARKED revivable per Atlas external review + unused (Compass is the active piece, stubs + comments say so; unwired-is-not-unloaded import side-effect cut); CPLAN capture-plan type born (records wearing plan lifecycle, CPLAN-0001 = Atlas capture). Suites all green, audits 100 pct. Full record: DPLAN-0295 run log, CPLAN-0001
…allowed the parked symbolic tier, so 8bd53d6 recorded the park as pure deletion and fresh clones could not revive it (same species as 292d815 template registry). Selective whitelist: only parked_symbolic_20260814 tracked, rest of memory/.archive stays local. 9 files, byte-identical copies of the deleted tier + revival README
…s (Patricks word, live-verified end to end today). @api FPLAN-0411 phases 1-2 + fleet: loopback-only FastAPI behind sha256-hashed bearer tokens (compare_digest, read|operate scopes, revocation next-request); /v1/feed ts-cursor clamped both ends w/ gap flag + at-least-once boundary; /v1/files + /v1/diff names-never-paths, 512KB refuses-not-trims, diff through drone @git; /v1/fleet + /v1/rooms consume baud --snapshot so which-agents-are-alive has ONE implementation - binary resolution prefers the deployed release build over any PATH copy, SNAPSHOT_READY kept as operational kill switch (closed = 503 with a reason, never a synthesized fleet). 713 tests, seedgo 100 pct, live-proven on a running server: 200 in 0.27s on 17 real branches, post-revoke 401, wrong-case exit-1 envelope path carrying bauds own sentence. Self-caught in build: layering (ai_mail/drone internals -> drone public surface) and the corrupt-registry-looks-like-typo regression split 400 caller / 503 ours. pyproject [host] extra (fastapi, uvicorn). @ai_mail: fleet->project refusal now tells the truth - outer wall said Unknown branch (address exists; sent @api phantom-hunting), now names the replies-only ruling (DPLAN-0288); wall STAYS SHUT, load-bearing test pins no-inbox-written; 8 tests 2 red-first, 1083 green. Morning riders committed on the same word: hooks TG relay cursor clamp + tests, @memory vector embedder parked to .archive/parked_symbolic (Atlas arc, revivable), skills TG start/kill control-verb fixes + tests, README/APLAN truth-ups (ai_mail, api, hooks, memory, skills, trigger). Verification chain all Patricks own hands: dev-window five-lane click-through (identical = pass), release rebuild 2m14s, his own --snapshot run (envelope, no window), /v1/fleet 200 on the binary his launcher execs
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
| root = resolve_branch_root(branch) | ||
| target = _fence(root, file) | ||
|
|
||
| if not target.is_file(): |
| if not target.is_file(): | ||
| raise ReadRefused(f"Not a file: {file!r}") | ||
|
|
||
| size = target.stat().st_size |
| raise ReadRefused(f"File is {size} bytes, over the {MAX_READ_BYTES} byte read cap") | ||
|
|
||
| try: | ||
| content = target.read_text(encoding="utf-8") |
| if ".." in parts: | ||
| raise ReadRefused("File name may not contain '..'") | ||
|
|
||
| resolved = (root / candidate).resolve() |
| logger.warning("[host_api] fence blocked a read outside %s: %r", root, file) | ||
| raise ReadRefused("File resolves outside the branch") | ||
|
|
||
| if not resolved.exists(): |
|
|
||
| try: | ||
| with socket.socket(family, socket.SOCK_STREAM) as probe: | ||
| probe.bind((host, 0)) |
…from the local venv (by @api; .github + setup.sh lanes by devpulse on their verdict). The red: bind-refusal test reached its patched serve only when fastapi was importable - on runners the CLI took the install-hint early-return first, Expected serve called once / called 0 times, identical on 4 python lanes + Windows. @api reproduced before fixing: meta-path blocker for fastapi/uvicorn/starlette - and their FIRST blocker used find_module, removed in 3.12, which silently blocked nothing and nearly read as not-reproducible; a green suite from an unverified harness is not evidence (their words, banked as a learning). Fixed BOTH availability-dependent tests to pin is_available, and the repro dragged out a worse neighbor: the non-numeric-port test was PASSING on CI for the wrong reason (same early return fired before the port check it claims to exercise) - new regression guard asserts a VALID port actually REACHES serve, so that false green cannot re-form. 719 green with the extra, 694 + 25 skip without, verified both ways, seedgo 100 pct. Verdict half two, my lanes: ci.yml installs .[dev,host] on both jobs and setup.sh adds host to the operator extras - the 25 skips were TestPing/TestAuth/TestScopeEnforcement/TestReadLane/TestFleetRoutes, i.e. CI had never made ONE HTTP request against the first network-listening service in AIPass; now every lane does, linux via ci.yml, windows/macos via setup.sh. Extra stays optional in pyproject; the guarded-import path keeps coverage via the tests that patch is_available False. Riding along from the same @api sitting: aliveness semantics pinned per @baud contract reply - has_room, outside_room, live_agent_sessions answer three DIFFERENT questions, five tests incl. one asserting this server invents no alive/online/active key anywhere (the green-circle-over-nobody lie stays dead)
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
2 participants
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.
No description provided.