Resolution: rules, limits, line/file counts and not-covered judge the aliased tree path medusa scans and name the repository path; findings, cache key (tree path, repository path, not covered) and accept rows keep the real path.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Requirement 001 S8: the hub can use a shared OpenAI-style speech service (KVB_ORCHESTRATOR_SPEECH_URL, key from a mounted file) instead of the local piper and whisper engines, and the image takes its pinned artefacts through an optional INFINITO_ARTEFACT_BASE cache. A new transport serves speech and the LLM: one wall-clock deadline over DNS, connect, TLS handshake and body, certificate and hostname always verified, no redirect followed. A key travels only over https or to loopback; both the speech and the LLM start refuse it otherwise. Key files are capped at 4096 bytes. Streamed WAV answers with placeholder sizes are accepted only when the data chunk runs to the end, and are rewritten before caching. Accepted MEDUSA false positives (41, independent review, operator pick): raw client socket patterns in transport.py, AIC-EXFIL-005 on the sealed key path, JB-SCAN-2237 on the hub's own system prompt, STG009 on the RIFF pad byte and LLM01 at language.py:111, where agent-written titles reach the NL prompt but the model only proposes a plan the operator confirms.
🎯 Fix: 93 %
💥 Break: 4 %
💸 Cost: –
⏱️ Time: –
🛡️ Security: 100 %
🏆 Score: 89
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Operator order 2026-11-13: once the operator ran a command, the agent that queued it must hear it directly, without the operator relaying it. Every row close now sends one radio call to the stored initiator (party name plus the session id add pinned; legacy rows by the A name): run and hub ran (done, exit 0), check still open, the check-pass sweep of list and run, done, cancel, failure and hub failed. The call names row, outcome and exit code and never carries command output; shared stderr stays in failed.output of the register. radio.py send takes --sid for a call to exactly one address. Accepted MEDUSA false positives: CGS-SCAN-029 and CGS-SCAN-030 at operator.py:930, which runs the row command only after the operator's per-row answer.
🎯 Fix: 93 %
💥 Break: 4 %
💸 Cost: 105 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 85
🚦 Gate: ⚠️ cost 105 % > 100: ask the operator before the commit
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
security.py cached any run that exited cleanly, so a run MEDUSA cut short by its file size, line or path limits, or one that skipped an unscannable type, was stored and reported as 100 %. It now predicts which changed files MEDUSA takes, from MEDUSA's own limits and exclusion defaults read from its source, and caches a result only when MEDUSA reports exactly that count. Files it does not scan are listed as not covered; a diff with nothing scannable, an unreadable limit or a count mismatch exits 4 instead of scoring 100. Coverage details move to coverage.md, keeping SKILL.md under the word cap.
🎯 Fix: 95 %
💥 Break: 2 %
💸 Cost: 93 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 100
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Requirement 006 grew SKILL.md to 853 words, over the 600-word cap test_lock enforces. The tables and message block details move to tables.md and the score journal with its backfill to journal.md; SKILL.md keeps the commit-block rule and points to both. The journal reader set in test_score_journal follows the text to journal.md.
🎯 Fix: 95 %
💥 Break: 2 %
💸 Cost: 110 %
⏱️ Time: –
🛡️ Security: ~100 %
🏆 Score: 85
🚦 Gate: ⚠️ cost 110 % > 100: ask the operator before the commit
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The stop-stage cleanup ran the configured docker image prune, which also removed dangling images still pinned by digest, so the next deploy pulled them again. A guard shim now sits first on PATH for every configured cleanup line: a plain docker image prune (also -a) becomes one pass that removes only images whose repository and digest are both <none>, without force, on the daemon the call names; an image a container uses is kept. system prune, rmi, image rm and compose down --rmi are refused and reported. Worktree cleanup commands run as argv lists; their nested image prune gets the same leftover pass inside the worktree daemon. The guard protects operator-written lines against accidents, not against a script that bypasses PATH.
🎯 Fix: 95 %
💥 Break: 3 %
💸 Cost: 106 %
⏱️ Time: –
🛡️ Security: 100 %
🏆 Score: 87
🚦 Gate: ⚠️ cost 106 % > 100: ask the operator before the commit
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
006's label test scanned docs/requirements too, but the 006 requirement has to name Cost and Time to describe their rename, so the test failed once the requirement and the code met on the unify branch. The test now skips docs/requirements, and the 006 acceptance line says so.
🎯 Fix: 96 %
💥 Break: 1 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 95
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Home tab left of Stream shows every panel and item the operator starred: a star on every panel and item pins it, tiles sit on a 12-column grid, resize in width and height, reorder by drag, render live, and a pinned item that no longer exists shows a gone tile with an unpin button. Home becomes the start tab once something is pinned. The layout lives in 007's view store, changed only through update() so two tabs never overwrite each other's pins. Committed below the 90/5 bar on the operator's word so the unify e2e proves it before master moves.
🎯 Fix: 88 %
💥 Break: 6 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 83
🚦 Gate: ⛔ fix 88 % < 90, break 6 % > 5
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Resources tab draws one hardware chart and one tokens chart with an accumulated switch, plus a combined overlay of hardware, tokens and the burnup, every series switched by a legend chip and all shown by default. The hub keeps every toggle in a validated view store: PATCH /api/view takes a set and the expected old values and answers 409 on a stale expectation, update() re-runs a change on the stored value, list paths change only through update(), and a page shown again from the back-forward cache fetches the stored view anew.
🎯 Fix: 93 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 89
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
score.py names the Cost and Time metrics Tokens and Velocity from one label file and appends every scored block to a score journal in the council root. backfill.py rebuilds that journal from commit messages with its own label mapping, deterministic and idempotent, so another host can import its history. The hub serves the journal, and the Gantt burndown draws each metric as its own overlay on a right 0-100 % axis, every overlay off by default and switched by legend chips whose state the view store keeps.
🎯 Fix: 90 %
💥 Break: 5 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 86
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
README.md conflict with the queue-cleanup timer sentence resolved by keeping both: the unify paragraph plus the council deny sentence. install.sh auto-merged (queue-cleanup state dir, pycache prune, council-deny call), shellcheck clean.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
New scripts/council-deny.js holds the deny rules (SPOT) for ~/.claude/agents-council/** and, new, ~/.local/state/agents-council/**, where the skills' scripts keep state only they may change. enable-plugins.js takes them from it for <TARGET>/.claude/settings.local.json as before. The global install now also merges them into ~/.claude/settings.json, the only user-wide settings file, because a settings.local.json binds only sessions rooted in its own repository; the operator no longer edits settings by hand. The writer follows a settings symlink (dangling included), keeps the file mode, writes a 0600 wx temp file and renames it, and refuses invalid JSON or a non-object with exit 1 and the file untouched.
Tests: tests/test_dream.py 26 passed (denies once beside existing rules, user settings kept through a symlink with mode, dangling link kept, invalid JSON untouched with exit 1); two independent reviews, round 2 holds.
🎯 Fix: 93 %
💥 Break: 3 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 90
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
install.sh copies the rendered skills with cp -a, which carried __pycache__, *.pyc and *.pyo along; Python loads a lone .pyc before the module of the same name, so a stray file in a skill could shadow its script. After each copy, find -H prunes them in .agents/skills and .claude/skills, also when .claude/skills is a symlink to .agents/skills; the delete before the copy uses -H too. orchestrator.sh drops .pyc and .pyo besides __pycache__ in the hub tree.
Tests: test_no_python_bytecode_reaches_the_installed_skills (linked and unlinked) and test_install_ships_no_bytecode_from_the_tree fail on base 3418738 and pass; shellcheck and ruff clean; two independent reviews, round 2 holds.
🎯 Fix: 95 %
💥 Break: 2 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 93
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
radio.py conflict resolved by the unify side plus the prepared resolution (seen() on top of the feedback-channel radio.py); receipts.py and its tests adapted to the unify helpers. Merged tree: test_hub_receipts 15, test_radio_seen 8, test_radio 45 passed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
receipts.py builds, for every radio row in a window, the sessions that took it (from their prompt and radio-seen marks) and those that acknowledged a call, with names taken only from the host agent-sessions.json snapshot, so a forged name in a row never reaches the readers; one snapshot build per call keeps a 90k-entry history at 13 ms instead of 8385 ms. radio.py gains seen(), which pending() now uses, so the hooks and the hub decide taken the same way. gating.py and receipts.py share one script loader (skill_scripts.script_module). Not wired into the hub yet: phase B.
Tests: tests/test_hub_receipts.py and tests/test_radio_seen.py; both import unify-side helpers and run green merged into unify/skills-20261112-night with the prepared merge resolution.
🎯 Fix: 92 %
💥 Break: 3 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 89
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hub gates uncached /api/tts renders on the latest W4 host CPU sample (503 busy or unknown, cached lines always served, signed speech.threshold 30 to 90 in steps of 10, speech.gate in /api/state). The page plays a WebAudio squelch-tail burst when no speech plays, offers Auto, Quiet and Silence per browser, and always sounds the MAYDAY, PAN-PAN and SECURITE procedure word. The radio panel becomes a minimalist VHF faceplate with display line, mode buttons, knobs, PTT send and the internet station; the settings dialog loses its radio section. The log gets Sent, Received and All tabs with class, sender, recipient and agent chips plus text search. Decisions 3 and 10 of 004 now state that the procedure word still sounds in Silence.
Security: MEDUSA-MEX-SCAN-1167 at orchestrator/kvb_orchestrator/hub.py:126 accepted as a false positive after the round 5 skeptic review and the operator pick of 2026-11-13 (regex hit on rce inside RESOURCES; CpuGate reads host CPU samples, no model).
Break risk 6 % is above the 5 % bar; committed and merged on the operator pick of 2026-11-13 for this bundle.
🎯 Fix: 92 %
💥 Break: 6 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 86
🚦 Gate: ⛔ break 6 % > 5
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
New skill kvb-org-queue-cleanup: each round judges every queue job as dead, held or unclear from /proc/locks and a non-blocking flock, removes only the dead ones pinned to the same id and lock tag, and reports the round on the radio; radio silence or a stall ends a round early. install.sh installs the systemd timer, service and path units and creates ~/.local/state/agents-council with mode 0700 (install -d, SC2174 fixed). The silence holder is read with .get, so a silence row without a sender no longer raises KeyError.
Security: 19 MEDUSA findings in skills/kvb-org-queue-cleanup/scripts/cleanup.py accepted as false positives after the round 9 skeptic review and the operator pick of 2026-11-13: MM-SCAN-058 (:153), JB-SCAN-641 and JB-SCAN-616 (:185, :191, :356, :361, radio.clip display sanitizer), JB-SCAN-2702/2399/2633 (:319), JB-SCAN-1866 and JB-SCAN-5115 (:264, :291, :337), CGS-SCAN-040 (:406, timed poll loop by design), GENAI-SCAN-1270 (:425, JSON register), RT-SCAN-194 (:577, sandbox refusal message).
Tests: tests/test_queue_cleanup.py 35 passed, 37 subtests; shellcheck and ruff clean.
🎯 Fix: 92 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 88
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ticks in 001 and 003 now name a test that passes on the final code: a 21-screen sweep in Chromium and Firefox at 1920, 1366 and 390 (no page scroll, every scroller reaches its last child, nothing clipped), the sessions timer check, the pkgmgr branch tree file, the forecast scope, the resources charts, bubble schedule, assets, stop and reading time, and the kanban and cleanup executors. Scrolling screens move into a pane so only main scrolls. The cleanup sheet keeps its dry-run note as visible text, now pinned by a test. The W6 fade check stamps removal from a MutationObserver against a timer started on the same mutation. Boxes whose final browser run is still owed stay unticked.
🎯 Fix: 95 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 91
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The offices and cleanup executors drained jobs and results through paths the hub can write, so a symlinked jobs, results or spool directory, a linked job file or a linked subdirectory let a root or host unit read, write or delete outside the spool. verifier.drain() now opens every directory from / one name at a time with O_NOFOLLOW: a link is followed only by a non-root unit and only while every directory entered so far is root's and nobody else can write it; the final directory must be the unit's own and not group or world writable, else the cycle refuses and changes nothing. Reads, writes and removals run relative to the open directory fds; a removed subdirectory goes through shutil.rmtree with dir_fd, which refuses a link swapped in. A result that cannot be written keeps its job as results/.<id>.unreported so it never runs again, and a job that can be neither kept nor removed prints one line and leaves the cycle going.
MEDUSA: 2 findings accepted after an independent review, 0 true positives: WEB-PATH-002 at verifier.py:146 (OSError.filename in a refusal message, no file access uses it), B105 at cleanup.py:53 (TOKEN_NEEDED is a refusal sentence, accepted before on 2026-11-12).
🎯 Fix: 92 %
💥 Break: 5 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 87
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
W1 clicks each checkbox through a fresh lookup, since every click redraws the list, and counts feedback on asks as well as reviews. W2 gives work/c the Fibonacci estimate 3, as the hub drops non-Fibonacci values. W3 serves the page on localhost, where the passkey signs, and opens the collapsed repo card before dragging. The orchestrator suite finds the ask card by its command code and reports the page text when none is found. 45 passed, 490 subtests (w1, w2, w3, w6, orchestrator).
🎯 Fix: 95 %
💥 Break: 3 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 92
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
operator.py stores the adding session's name and id with the row. Done,
cancel and failure go to that stored name and pinned session id, never
re-resolved: a radio post by default, a call when that session picked it
with radio.py feedback set call. The caller's own parts are dropped. The
hub counts unanswered calls per session id.
MEDUSA: 34 findings accepted after an independent review, 0 true
positives: 3x CGS-SCAN-029 by design (operator-confirmed runner in
operator.py execute and its call), 2x CGS-SCAN-030 (exec inside the name
execute), CODEGEN-SCAN-003 and CGS-SCAN-028 (open inside Popen),
3x CGS-SCAN-040 (prompt loops end), 4x GENAI-SCAN-1270 (rasci.database()
JSON register), 20x FLA-SCAN-055 (dba inside the word feedback).
🎯 Fix: 93 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 89
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
medusa v2026.7.0 picks files by their last lower-case extension and has no scanner for .mjs or .cjs, so security.py counted them as scanned (100 % in N files) or ended in report unreadable for an .mjs-only change. The scratch tree now gives them a .js alias that cannot collide with a changed path or folder (a hashed name when the alias would pass NAME_MAX), and every finding, fingerprint, accept row and cache key uses the real path. Reported by Avatar 3.0 via Main.
🎯 Fix: 94 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 90
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The bubble whose speaker is talking gets z-index 1 while it is not leaving, so it is always the front one. With look.faces on, the ancestry view draws the face of the Operator, Chair and Session nodes (r13 ring, data-face); subagents keep their circles. anchorOf intersects a bubble's anchor with its clipping ancestors, so a bubble never points at a hidden part of a node. Unit tests, a clip mutant and an e2e run with screenshots cover it.
🎯 Fix: 95 %
💥 Break: 5 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 90
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hub read queue.json alone, so an active entry whose hold had exited kept showing as running. client/queue_live.py probes each hold lock as queue.py's held() does (shared non-blocking flock, hard links probed too) and reads the whole queue-history.jsonl, cached by inode, size and mtime; an active entry with a history row started after its enqueue ends with that row's exit, one whose holder is gone becomes gone. The outcome lives in the hub-owned field hub_outcome; queue.py status is never rewritten. resources.py uses the same settled(), plan medians skip the queue kind, and a guard test keeps every client module in the Dockerfile COPY line. Also closes fd leaks in read_history and records.read_regular.
MEDUSA: 11 findings in resources.py (lines 73, 81, 262, 445 x2, 576, 227, 479, 524, 578) and crons.py:125 accepted as false positives after an independent review, operator pick 2026-11-12.
🎯 Fix: 92 %
💥 Break: 5 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 87
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
W4 added the Resources tab after W1 wrote this assertion; the merge kept both and the test failed on master.
🎯 Fix: 99 %
💥 Break: 1 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 98
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A row with a lone surrogate in gate-parents.jsonl, which sandboxed agents can write, raised UnicodeEncodeError at the TEMP table insert and crashed every commit-msg, pre-push and pre-merge-commit hook and the hub gate view. opened() now keeps only rows whose repo, branch and parent encode; the test fails without the change.
🎯 Fix: 97 %
💥 Break: 2 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 95
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
W5 moved recorded parents into a journal that gates.opened() loads into a TEMP table; the W3 hub opened gates.db raw, so parents_of() raised and the gate view stayed empty. opened() and journal() take the kept directory explicitly (CLI: watch.KEPT, hub: its /kept mount), opened() closes the connection when setup raises, and the W3 tests use opened() plus a test for a parent from the kept journal.
🎯 Fix: 95 %
💥 Break: 3 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 92
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The pkgmgr wrappers replay renamed kvb-git-branch-gc: branch cleanup outside worktrees is now kvb-git-branch-cleanup, and removing merged branches with their worktrees is kvb-git-gc. The hub's held-by-worktree hint, the cleanup dialog and its test now name kvb-git-gc.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Conflicts combined W1 (Inbox leaves the Stream for Tasks, Waits on Operator) with W4 (Resources tab): five tabs with Resources, Stream keeps only the Firewall, the inbox hash keeps its redirect; index.html.j2 includes both diffs.js and resources.js.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>