queue.py hold printed nothing when a waiting hold got its grant, so a quiet command such as node --check looked like a silent exit (Quality, test-324: history shows it ran 5 s, exit 0). A hold that printed a wait line now prints 'granted after waiting; running <cmd>'; holds granted at once stay quiet. watch.py counted every make target starting with test, quality or verify- as heavy, which queued make test-runner-down behind the gate it was meant to relieve (Stage 2.0); a target ending in -down is never heavy.
🎯 Fix: 94 %
💥 Break: 3 %
💸 Tokens: 100 %
⏱️ Velocity: 100 %
🛡️ Security: 100 %
🏆 Score: 91
🚦 Gate: ✅ 90/5, tokens and velocity <= 100, security 100 %
Created-with: https://s.veen.world/skills
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
watch.py treats node --check FILE and node -c FILE as a parse-only call: launched() no longer names it a program launch and light_program() admits it. Any other option, a missing file or a --require hook keeps the old heavy verdict; test_watch.py pins both sides.
🎯 Fix: 93 %
💥 Break: 2 %
💸 Tokens: –
⏱️ Velocity: –
🛡️ Security: 100 %
🏆 Score: 91
🚦 Gate: ✅ 90/5, tokens and velocity <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
queue.py hold probes the command's process tree from /proc every poll
(standstill.py): the descendants by parent link plus every process still
in the command's session, their CPU ticks, rchar/wchar/read_bytes/
write_bytes and the process set. Unchanged for HUNG_SECONDS (3600) it
records the stuck processes (pid, comm, state, wchan; never a command
line), SIGTERMs the tree and the sessions its members lead, waits
KILL_GRACE_SECONDS while recording what appears, SIGKILLs the rest on
every pass and exits 124. A tree running docker, podman or ssh, or with a
stopped member, is never judged. The history row carries the probe as
`hung`, learn() skips such rows, and list marks an active job past three
times its longest learned run (at least 30 min) `suspect hung`.
Residual: a process that setsids out of the tree and every recorded
session between two passes escapes; only a subreaper or a cgroup per job
would hold it.
🎯 Fix: 90 %
💥 Break: 5 %
💸 Tokens: 104 %
⏱️ Velocity: 100 %
🛡️ Security: 100 %
🏆 Score: 82
🚦 Gate: ⚠️ tokens 104 % > 100: operator approved the commit (2026-11-15)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Probe sandboxes gain the channels, quietBubbles and known.now the merged branches added; the hub view keeps queue_jobs apart from the job list; the image ships sandbox_paths.py; kvb-match writes pages into the council root pages folder; the install test scaffolds the council parts and agent roots; skill texts move detail into their side files to stay under the word caps.
🎯 Fix: 88 %
💥 Break: 4 %
💸 Tokens: 98 %
⏱️ Velocity: 100 %
🛡️ Security: 100 %
🏆 Score: 86
🚦 Gate: ⛔ fix 88 % < 90
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
councilroot.py resolves one durable council root for every script and migrates the old /tmp root under the held locks: a preflight conflict moves nothing, each legacy file is renamed to a staging name before it is copied so a late writer is merged by the catch-up, id-less journals merge by line multiset, a differing whole file is kept beside its namesake, and a crash marker makes install copy the skills and report. One settings-file list, managed drop-ins included, serves councilroot.py and queue.py. Tests run once on unify/skills-20261114 after the merge (operator order 2026-11-14).
🎯 Fix: 88 %
💥 Break: 7 %
💸 Tokens: 100 %
⏱️ Velocity: 100 %
🛡️ Security: 100 %
🏆 Score: 82
🚦 Gate: ⛔ fix 88 % < 90, break 7 % > 5
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
watch.py admits heavy work only while PSI, swap-out rate and RAM forecast stay under the configured limits; config.default.json is the single default the install writes once (mode 600, no clobber) and every hook, admit and queue command falls back to. Browser runners and package-manager script forms count as heavy, preflight derives its probes from the light and heavy tool lists, and the shipped default carries no prune command, buildx path or host-specific key. Tests run once on unify/skills-20261114 after the merge (operator order 2026-11-14).
🎯 Fix: 85 %
💥 Break: 8 %
💸 Tokens: 99 %
⏱️ Velocity: –
🛡️ Security: 100 %
🏆 Score: 79
🚦 Gate: ⛔ fix 85 % < 90, break 8 % > 5
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
937c2c0 added the line prefix KEPT to watch.py, which shadowed the state directory path watch.KEPT that gates.py and the queue cleanup use; the prefix is now KEPT_LINE. The receipts page test gets an identity revealed() since page-dom.js wraps names through it after the firewall merge, the operator tasks test expects the Home tab of 008, and the hub view id fixture uses a non-ASCII word the English check accepts. Accepted MEDUSA false positives, re-keyed by the rename: MM-SCAN-058 at watch.py:1417 and WM-SCAN-354 at watch.py:1418.
🎯 Fix: 92 %
💥 Break: 2 %
💸 Cost: –
⏱️ Time: –
🛡️ Security: 100 %
🏆 Score: 90
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
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>
Conflict in queue.py cmd_hold: kept the light-footprint aliases and per-kind stats, added W5's gated() call; W5's qualified() signature is gone and was unused below.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Light single-file runs booked about 17 % RAM and 5 GB learned from heavy runs under a polluted signature. Each entry now stores its aliases, freezes its footprint at grant and learns per signature and kind; heavy kinds book at least the old pooled share, and a job keeps its own learned disk need when the old pool is empty. Hooks take the payload cwd and fall back to the old rule when it is unknown or malformed. Six skill docs now agree on what self-admits. test_queue 70, test_watch 46 and test_operator 30 green with negative controls; two independent reviews.
🎯 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>
gates.py installs hooks that check the merge level before a push without writing git config, writes receipts from green queue gates, and the queue gains the after-wait and rank. Requirement 003 strikes the hook-enforcement criterion: agents keep --no-verify, hooks bind humans and CI. Committed below the bar on the operator's word 2026-11-12 (one gate after the merge): test_gates 15, test_queue 63 and test_watch 46 green, security 100 %, break risk 6 % for the live queue and hook change.
🎯 Fix: 90 %
💥 Break: 6 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 85
🚦 Gate: ⛔ break 6 % > 5
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An empty /tmp/.git (2026-11-11 19:47, owner unknown) made council.py gate treat all of /tmp as a git worktree: every unclaimed Edit or Write under /tmp was blocked and test_council turned every skills gate red. A parent now counts only when its .git is a directory holding HEAD or a regular file starting with gitdir:, as git itself requires, so normal repos, linked worktrees and submodules still block unclaimed edits. A FIFO named .git is never opened, so it cannot hang the hook. Root fix proposed by the chair Main.
🎯 Fix: 94 %
💥 Break: 3 %
💸 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>
cache_files walked every cache_dirs entry with os.fwalk, which raises FileNotFoundError for a folder that does not exist (~/.cache/ms-playwright on this host), so cleanup --caches crashed and trim_caches never ran at disk stage warn. A root that cannot be walked is now skipped and the other roots are still trimmed.
🎯 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>
kvb-org-swarm keeps agent work in its own worktree until it meets 90/5; kvb-org-unify brings the swarm's branches back together. Every skill that named kvb-org-split now names kvb-org-swarm. Cost 111 % accepted by the operator on 2026-11-11.
🎯 Fix: 92 %
💥 Break: 3 %
💸 Cost: 111 %
⏱️ Time: 100 %
🛡️ Security: 100 %
🏆 Score: 80
🚦 Gate: ⚠️ cost 111 % > 100: ask the operator before the commit
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Requirement 001 phase 6, S4 (decision 29):
- Skills renamed: kvb-org-evening -> kvb-org-drain, kvb-org-sunset ->
kvb-org-suspend, kvb-org-sunrise -> kvb-org-resume, kvb-agent-relax
-> kvb-agent-drain, kvb-agent-sleep -> kvb-agent-suspend,
kvb-agent-wakeup -> kvb-agent-resume; their text uses the new verbs and
every reference in skills, README, orchestrator and tests follows.
- night.py takes drain, suspend, resume, resume --alone and status. A
state file written before the rename keeps working: mode evening reads
as drain, sunset as suspend, the awake list as resumed and is rewritten
on the next save; the hub's council.night_mode maps the same values.
mode: day and handover_night_<session>.md keep their names.
- Orchestrator: the Day/Night menu shows Drain, Suspend and Resume, each
confirmed, signed through /api/board/sign with the passkey and run as
the existing argv job; the old action names are refused.
- kvb-skill-cleanup: git rates two renames at 44 and 50 % similarity,
under its default threshold, so legacy.py detects renames from 40 %.
- tests/test_night_names.py fails on any old name in the tree outside a
rename record "old ->".
BREAKING CHANGE: the six old skill names and night.py's clear and wake
are gone. Run make install (it reinstalls the orchestrator too), then
kvb-skill-cleanup to remove the old installed copies; until the install,
the running hub calls the old night.py path.
Evidence: make test 679 OK, ruff and markdownlint clean
(/tmp/s4-make-test-1.log); e2e 22 OK in Chromium and Firefox
(/tmp/s4-e2e-3.log); MEDUSA 23 files, the 2 findings are the
byte-identical atomic state write that moved with night.py, already
accepted after a skeptic at its old path, rescan 0 open.
🎯 Fix: 93 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: –
🛡️ Security: 100 %
🏆 Score: 89
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Requirement 001 phase 6, S0:
- Speech: the security class shows as "Sécurité" and is spoken as in
maritime radio, "say-cure-ee-tay" three times. A hex run of 7 or more
characters with a letter and a digit (commit, session, message ids) is
spelled in the NATO alphabet; every other digit is spoken on its own
in ICAO pronunciation (42305: four two tree zero fife). The spelling
runs inside clip() before the length cut, so a phrase still fits
SPEECH_LIMIT; the displayed text keeps its digits and hashes.
- English: every orchestrator label, tooltip, message, error, script
output, document and test string is English (Regeln -> Firewall,
Hierarchie -> Hierarchy, Steuerung -> Controls, Ausführen -> Run, ...);
the page declares lang="en" and formats dates as en-GB. Task filter
ids are English. German stays only in the skill trigger phrases that
quote the operator, and as identifiers: the view keys (view=) and the
leitstand id.
- tests/test_english.py scans every tracked text file for umlauts and
German words, with an allowlist of the quoted trigger phrases.
BREAKING CHANGE: the Stream search tokens art: and seit: are now kind:
and since:; an old art: or seit: token counts as a plain word.
Evidence: make test 657 OK, ruff and markdownlint clean
(/tmp/s0-make-test-1.log); e2e 17 OK, none skipped (/tmp/s0-e2e-1.log);
MEDUSA 86 files, 24 findings on reworded strings accepted after an
independent skeptic (21 path:line, 0 rejected, no behaviour change in
signing, approving, rules, denying, offices, events), rescan 0 open.
🎯 Fix: 93 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: –
🛡️ Security: 100 %
🏆 Score: 89
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Root cause of the 2026-11-07 20:14 OOM: the queue granted jobs per kind
only, so several heavy kinds and direct heavy commands outside the queue
ran side by side until the kernel killed the editor.
- queue.py: every grant checks the sum of the learned RAM and CPU
footprints of active and granted jobs against budget_percent; an
unlearned kind takes the whole budget and runs alone. RAM used plus
what young jobs (RAMP_SECONDS) may still take must stay below
load_limits. Memory stage stop blocks every grant.
- watch.py: memory stage from MemAvailable, swap, PSI full and the swap
of the container holding most of it; the block hook classifies
builds, make install/verify-*, and pytest or unittest beyond one file
as heavy through wrappers, sh -c, command substitutions and heredocs
a shell reads.
- history.py: footprints per kind and signature, seed RAM for pending
shares.
- operator.py: [h]old queues a heavy row through queue.py hold and asks
again on its grant, with a flushed tty and a timeout; sudo rows, which
need the terminal, stay direct.
Residuals: a shell reading code from a pipe or here-string, its own
script, a program named by a substitution, quotes inside an expanding
heredoc, an escaped backslash before a substitution; a lone job waits
while the host is over load_limits.
🎯 Fix: 92 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: –
🛡️ Security: 100 %
🏆 Score: 88
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Agents views gain a Hierarchie view: an org chart of the chair and
the line offices read from the council claims. The operator appoints,
hands over and recalls an office there; the page signs each change with
the passkey and a host unit (client/offices.py, offices path unit)
verifies the signature, removes a live holder itself and runs a plain
council.py claim for the successor. council.py loses the --operator
flag, so no agent can override a live holder.
The agent order is now separate from task priority: rasci.py agents
order stores it in agent-order.json beside the register through the
nofollow helpers and never writes a --prio pin. It only breaks ties of
equal task rank, also among the queue grants. The Hierarchie view edits
it as a drag list.
Host units run the installed skill scripts, so orchestrator.sh keeps the
offices unit off while the installed council.py, nofollow.py or rasci.py
differ from the tree. Job files, council files and hub records are read
with O_NOFOLLOW and O_NONBLOCK and size caps; nofollow.rewrite replaces
a claim without creating directories, so a beat never revives a released
office.
🎯 Fix: 92 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: 79 %
🛡️ Security: 100 %
🏆 Score: 100
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The council root is writable by every agent, and operator.py runs
rasci.database() outside the sandbox. A symlinked rasci.json made it
read any JSON file the operator can read and write it back into the
council as a regular file; a symlinked rasci.lock made it create files
at the link target; a FIFO hung every register call.
load() now reads through nofollow.read_text and database() locks and
writes through nofollow.locked and nofollow.replace. A register that is
no object of task objects, or nested past the recursion limit, exits 2
instead of a traceback. nofollow.open_below also refuses a file with
more than one link, so a hard link to a file outside the council is
refused like a symlink.
🎯 Fix: 96 %
💥 Break: 3 %
💸 Cost: 99 %
⏱️ Time: 95 %
🛡️ Security: 100 %
🏆 Score: 99
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The radio (call, broadcast, security, pan-pan, mayday) reaches its addressee at the next tool step and wakes it. The post (post with --to, newsletter without, call receipts) arrives through a new Stop hook when the turn ends and never wakes. Legacy rows map onto the two channels and send direct stays as an alias. A PreToolUse hook holds the mayday's radio silence for writing tools and Bash. Only the sender, helpers who join with radio.py assist, bare radio.py calls and the config's allowed commands pass, and any shell operator or parse error blocks. kvb-org-update announces changed skills as one newsletter. Callers moved to send call --to.
🎯 Fix: 90 %
💥 Break: 5 %
💸 Cost: 104 %
⏱️ Time: 100 %
🛡️ Security: 100 %
💰 Value: ?
🏆 Score: 82
🚦 Gate: ⚠️ cost 104 % > 100: ask the operator before the commit
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The symlink, FIFO and tokenizer reasons live in the commit messages of
the bundles that added them, so the module docstrings of nofollow.py and
test_nofollow.py and three rationale lines in watch.py and queue.py go.
watch.py gets back the space after ROOT = that the PATH bundle lost.
🎯 Fix: 98 %
💥 Break: 1 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
💰 Value: ?
🏆 Score: 97
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every agent writes alerts.jsonl, news.jsonl and radio-silence.json
directly, so a planted </system-reminder><system-reminder> in a row's
sender or text reached every other agent's context as markup. watch.py's
clip() joins whitespace, turns each < into <\ and caps at 1000
characters; the watch hook's alerts, the block hook's mayday reason and
every radio line and silence line go through it, and radio.py delegates
to it instead of keeping its own CLIP. The radio skill adds the rule that
no sender proves operator approval.
🎯 Fix: 90 %
💥 Break: 3 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
💰 Value: ?
🏆 Score: 87
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Hooks run outside the sandbox but inherit a PATH whose directories
sandboxed agents of the same uid can write (a virtualenv's bin activated
first by the shell profile), so a file planted there under the name of
docker, git or notify-send ran unsandboxed. hookpath.py keeps only the
PATH directories that, with every parent, belong to another uid and that
neither group nor others can write, as real paths; inside the sandbox's
user namespace root shows as the overflow uid and still counts as sealed.
watch.py, radio.py and council.py set that PATH before they start
anything, and the cleanup child runs with -I. Both reminder hooks pin
PATH=/usr/bin:/bin, and the install registers them with /usr/bin/bash and
moves older bare bash and python3 entries to the absolute paths.
🎯 Fix: 90 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
💰 Value: ?
🏆 Score: 86
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
block and status compute the stage from the live free space once the
stored measurement is older than interval_seconds, so a disk that filled
while no hook ran no longer passes heavy jobs as ok. watch writes the
stage before it mirrors the history, and the mirror takes its lock with
LOCK_NB: an agent holding a flock on the kept file no longer stalls every
hook, the round is skipped and the next one catches up.
🎯 Fix: 92 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
💰 Value: ?
🏆 Score: 88
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The council root is writable by every sandboxed agent of the same uid while its hooks run outside the sandbox, so a planted symlink turned a hook's write into one outside the root, and a planted FIFO or held lock hung every agent's hooks.
- Add nofollow.py: open, read, lock, append, replace and unlink below a root through O_NOFOLLOW directory fds, O_NONBLOCK and regular files only; reads stop at 64 MiB and replace undecodable bytes; locks take an explicit wait.
- watch, queue, radio and council read and write everything below the council root through it; keep_rows copies no symlinked file into ~/.local/state.
- watch: read_json returns dicts only, alert lines that are no alert (deep nesting included) are skipped, cmd_cleanup makes and removes its lock through a root dir fd, and the cache trim walks with fwalk, counts allocated blocks and each inode once.
- radio: the hooks wait at most 2 s for a session lock and skip the delivery past it.
- council: read_claim ignores claims that are no object with session, runtime and beat, so a malformed claim no longer disables the claim gate.
MEDUSA acceptances (kvb-think-criticize 2026-11-06, skeptics A and B): MM-SCAN-058 on nofollow.replace callers, B105 and HPT006 in history.py, RT-SCAN-054 in council.py (name matches).
🎯 Fix: 92 %
💥 Break: 4 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
💰 Value: ?
🏆 Score: 88
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Move the council config to ~/.claude/agents-council/config.json, which the sandbox cannot write; the installer adds Edit(~/.claude/agents-council/**) to permissions.deny.
- Run every kvb hook as /usr/bin/python3 -I and rewrite old bare python3 entries.
- Refuse a council root that is a symlink, another uid's, or group or other writable, in watch, queue and radio.
- Run worktree_cleanup entries as argv without a shell in /, only for compose project names.
- queue block: tokenize every Bash command like bash (quotes, word-start comments, line continuations), and hold each held command to the Bash deny and ask rules of the managed, user and project settings; refuse a hold it cannot parse.
- radio: deliver fields on one line of at most 1000 characters, at most 10 rows.
Operator step after install: move ~/.config/agents-council/config.json to ~/.claude/agents-council/config.json.
MEDUSA acceptances (kvb-think-criticize 2026-11-06, skeptics A and B): ACP-SCAN-028 load_config (config outside the sandbox, Docker residual 093); CODEGEN-SCAN-004 and LLM06 run_cleanup and worktrees_command shell strings (operator config only); SAST-PATH-001 queue hold Popen (the caller's own command in its sandbox); PA-133 radio hook context (peer rows, clipped); JB-SCAN-641/616/2979 clip and MM-SCAN-058 replace (name matches).
🎯 Fix: 90 %
💥 Break: 5 %
💸 Cost: 100 %
⏱️ Time: 100 %
🛡️ Security: 100 %
💰 Value: ?
🏆 Score: 86
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
security.py runs medusa with HOME inside its temporary directory, since medusa creates ~/.medusa on start and the sandboxed home is read-only. security.py accept records a reviewed finding (rule, path, sha1 of the stripped line, reason) in <council root>/security-accepted.jsonl; the watch hook keeps it in ~/.local/state/agents-council/ next to the queue history, and a scan drops matching findings from the score as accepted until the line changes. Accepted so far: MEDUSA-MM-SCAN-058 x3 and MEDUSA-FTSCAN-231 x2 (false positives in tokens.py and score.py), WEB-CMDI-001 and SAST-EXEC-003 on watch.py shell=True (by design, operator config).
🎯 Fix: 93 %
💥 Break: 3 %
💸 Cost: 119 %
⏱️ Time: –
🛡️ Security: 100 %
💰 Value: ?
🏆 Score: 76
🚦 Gate: ⚠️ cost 119 % > 100: ask the operator before the commit
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every skill that posted to news.py now uses radio.py post, every SendMessage to a peer who must act becomes radio.py send direct --wake (which prints the SendMessage fallback for sessions without a listener), and every request for the chair's grant becomes queue.py hold. kvb-check-gate wraps gate.py run in queue.py hold; the operator sweep is one broadcast without a wake.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
kvb-org-radio is the default channel to other agents: broadcast and direct reach the receiver on its next prompt, security and pan-pan mid-turn through a PostToolBatch hook, and only mayday or direct --wake wake a session, through the receiver's own asyncRewake Stop hook (a SendMessage fallback is printed for sessions without a listener). It keeps the news.jsonl store and lock, so news.py stays a forwarding shim. A mayday holds radio silence for two hours: watch.py blocks new heavy commands and the operator gets a desktop notice. hotfix.py announces merged hotfixes as security items.
kvb-org-queue grants heavy jobs by rule instead of chair messages: queue.py hold waits under a flock with disk reservation, hotfix-first and register rank order, slot limits, stage and silence checks, then runs the command and releases. Every job is recorded; history.py learns disk need per command and slot limits per kind (p90, five runs), watch.py mirrors the history to ~/.local/state. enable-plugins.js registers the new hooks behind a file check so a missing script never blocks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Maintainer messaged every agent about every hotfix, waking each for a full turn. Hotfixes are now register rows hotfix/<slug> (repository identity by git common dir and normalized origin URL, branch, sha, PR, failing-line signature, targets, stage), written with hotfix.py add/update and queried only when an agent acts, never by cron: pending lists merged hotfixes missing from HEAD (filtered by targets) and open ones; match finds hotfixes whose signature fits a failing line, evaluated in one worker under a shared deadline; void tells gate.py run, before every target, when a merged hotfix missing from HEAD makes the rest of the run worthless (exit 3), the operator's one exception to never cutting a job short; sync fetches first, then under the worktree locks merges origin/<default> and every base up the funnel chain recorded with rasci.py set --base, refusing a detached or foreign branch, a diverged base or a busy worktree, aborting cleanly on a conflict.
gate.py gains hold, which runs any heavy command, deploys included, under the worktree lock, forwards signals, returns the command's code and exits 75 when busy; run holds the same lock and takes the sync lock shared. kvb-check-gate, kvb-ci-triage, kvb-org-organize (Peer protocol, capacity), kvb-org-coordinate and kvb-org-split follow; rasci gaps, wsjf and artifacts skip hotfix and operator rows. Three independent skeptic rounds found 34 defects (three of them high), all fixed with tests; negative controls fail against the earlier code.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every loaded SKILL.md and every listed description is paid on every turn of every session, and every delivered message costs its recipient a full turn. scripts/budget.py gains SKILL_WORDS = 600, enforced by test_lock, and test_naming forbids a skill naming its own slash command; 53 descriptions and 10 body lines drop it. The 18 longest skills move details into 31 side files read only in their situation (organize's capacity, succession, peers, offices, relay and lineage; cleanup's docker, worktrees, execute and ask; code-tune, changelog, triage, dream, split and others), so all SKILL.md files shrink from 33866 to 26291 words without dropping a rule; an independent review of HEAD against the new text found no loss in 60 of 64 skills and three weakenings, now restored (WSJF re-estimates after an objective change, cleanup's volume checklist, essence's absolute dates).
The kvb-org-organize Peer protocol sends status that asks nothing to the news board, answers nothing that asks nothing, keeps messages terse, and lets light runs admit themselves in the operator's wording; the skills follow it (operator reads the register, relax and wakeup reply through the news, evening collects them there, wsjf asks only for missing estimates, cleanup and triage batch their questions). Skills that explore code ask the codebase-memory-mcp graph first, subagent briefs scale with the stakes and cap their answers, and the skills that need the full register or council tables call show --wide or list --long.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agents read these tables on every status, organize and matrix step. rasci.py show now prints Prio, WSJF, Task, Status, Worker and A (task slug never cut, title cell 40 characters, worker and A 24), down from 1817 to 1003 words on the live register; show --wide prints the old 14-column table byte for byte. council.py list prints office, holder, state and a 40-character work cell with whitespace collapsed, 1236 to 276 words; list --long prints the old form. The old assertions now run against --wide and --long; new tests pin the compact defaults, cutting, --all, --mine and pipes in titles, and fail against the previous scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every delivered SendMessage costs its recipient a full turn of re-reading its context. news.py keeps status that asks nothing in <council root>/news.jsonl: post adds an item for every session or --to named ones (at most 500 kept, under a lock), and its UserPromptSubmit hook prints the unseen items of the last 24 hours (at most 10, own posts left out) with the session's next turn and marks them seen per session; the hook never fails a prompt. show prints without marking. enable-plugins.js registers the hook; test_news covers single delivery, addressing, own posts and broken input.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
kvb-org-evening and kvb-org-sunset narrow the organisation to the session the operator types them in; kvb-org-sunrise undoes it. night.py keeps the state in <council root>/night.json: set makes the caller active and lists the other live sessions, status tells each session its part (active, awake, relax, sleep), wake lifts one session alone, clear ends the night; a missing or invalid state means day. Evening messages every other session to run kvb-agent-relax (finish the task, end its loops, start nothing); sunset messages them in waves of three to run kvb-agent-sleep (park at the next safe point, never cut a running job short, end loops, write a ten-line handover into the project's memory directory, reply asleep), and the operator then sets inbound messages to hold. At night the active session admits its own runs through watch.py and leaves the sleepers' offices alone. Sunrise switches messages back to accept, clears the state and sends kvb-agent-wakeup, which restores loops, jobs and roles from the handover; the operator can also wake one session alone by typing it there.
A peer's request to do less now needs no operator approval in the kvb-org-organize Peer protocol, and kvb-org-robot stops a run at its next safe point when the night reaches it. The README gains the agent category's state meaning and the new nodes; test_night covers set, status, wake, clear and a bad state.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A claim made with --session <name> stored the name, but the PreToolUse gate compares the hook payload's session_id, so the gate refused every edit of the claiming session and a later release without the flag exited 3. caller() now takes CLAUDE_CODE_SESSION_ID whenever it is set and falls back to --session only for other runtimes. test_council covers a claim and release with --session under a Claude Code session (negative control against the old caller fails). Claims made with --session before this fix go stale after an hour without a heartbeat and can then be taken over.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
branches.py lists every open branch of the remote with its lag behind the base and whether it belongs to the operator: a pull request by the operator's login, or a branch whose own commits all carry the operator's email; merged branches drop out and fork pull requests are marked. The skill merges the base into the operator's own branches directly and into others only after consent through kvb-think-listen, never rebases, aborts on a conflict for kvb-git-merge-fix, and hands every push and gh pr update-branch to kvb-org-operator. tests/test_update_branches.py builds a bare origin with own, foreign, merged, PR and fork cases. Operator order 2026-11-03; three descriptions trimmed for the listing budget.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds kvb-transcend (collector), kvb-skills-transcend (into the skills repo) and kvb-code-transcend (into the current repo). transcend.py scans the own scratchpad, peer scratchpads and the agent homes for scripts, tests and skills the skills repo does not ship, skips secret-named and oversized files, and remembers carried/dropped decisions by content digest in the gitignored .state/transcend.json, so a rescan offers only new or changed candidates. Wired into kvb-skills-tune (own scripts), kvb-code-tune step 6 (measurements) and kvb-skills-dream step 7. Seven descriptions are tightened to stay inside the 1300-word listing budget.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Operator rule: a hotfix gate (a branch from the default branch that fixes it) takes the next free gate or test slot before every other line; everything else keeps its WSJF rank, and a running gate is never cut short for a hotfix. Written into the chair's queue in kvb-org-organize and the grant step of kvb-org-wsjf.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
kvb-org-wsjf: the chair asks its subordinates for Fibonacci estimates of business value and complexity, rasci.py estimate records them per party, and WSJF = median value / median complexity ranks the tasks. rasci.py wsjf prints the WSJF table with OKR and the task id; the register's show takes its Prio from that rank (an operator --prio pin first) and adds a WSJF column. Four descriptions were tightened to keep the listing budget.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
rasci.py keeps the organisation's tasks in <council root>/rasci.json: RASCI roles, objective with key results, priority, worker, start, end, workdir, branch, tools and artifacts; done is refused while a key result is open. council.py writes every claim, takeover and release through, and show lists the live claims as one row per task with R, A, S, C and I columns. kvb-org-artefacts manages the artifacts in the same table for kvb-cleanup. The org, plan, triage, tune, cleanup, branch-gc and dream skills now record into and read from the register; five descriptions were tightened to keep the listing budget.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
All 38 first-party skills move to kvb-<category>-<name> with the categories think, create, code, ci, git, check, org and skills; kvb-cleanup and kvb-help are the only uncategorized names. tests/test_naming.py pins the schema, the Agent Skills name rules and the README organigram. enable-plugins.js repoints hook commands of renamed skills in settings.local.json so a stale script path cannot block tools.
BREAKING CHANGE: the old skill names and slash commands (/commit-next, /dream, /autotune, ...) are gone; use their kvb- successors listed in the README organigram.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>