Repository navigation
v5.18.0
Added
-
session-start lays native liveness over the peer banner (#1052): on Claude Code the coordinator calls
ListAgentsonce in Phase 0.5 and marks each registry peerbusy/idleand reachable bySendMessage, joined through~/.claude/sessions/*.jsonon the session id the host registry also keys on, never oncwd(two peers in one working copy share it) or a name prefix. Display only; harnesses without the tool print that liveness is not available. -
Lead-guided runs for grill, ux-grill, deepen, test-audit and discovery (#1976): a session that nobody answers — lead-guided by a fleet coordinator, or headless (
claude -p,codex exec) — no longer stalls or skips these skills.skills/_shared/lead-guided-mode.mdnames the four detection signals, the three duties (take the own recommendation, park owner-class questions, record oneEntscheidung (autonom, Lead-geführt): …line per answer) and the owner classes (external effect incl. a public tracker, money, accounts, deletion, security, critical product decision). grill runs a "Selbst-Grill" that answers only with cited evidence; deepen executes the rewrite and names the original as the rejected option; ux-grill, test-audit and discovery take their recommended dispositions; a lead-guided run ends with a report whose open points are statements, never a question. Interactive behaviour is unchanged. grill and deepen carry the text themselves so the OpenAI skills-only export stays self-contained (export check: 0 violations). Measured with headless probes on 2026-11-08: deepen and ux-grill made 0AskUserQuestioncalls, wrote their decision lines and ran to the end. -
Viewport pair per page and
/test --target-kind(#1962): the playwright-driver capture recipe writesax-snapshots/<page-slug>-w1440|w390.yamlplus screenshots for every page (rubric-v2 convention,locator.ariaSnapshot()) and avisited-urls.jsonwith the URL actually driven per page and viewport./test --target-kind live|production-build|dev-server --target-url <url>requests a kind;LIVE RUN/PRODUCTION-BUILD RUNis granted only when the visited URLs have the target's origin, otherwise the report says dev-server and names the mismatch.checkReportCompletereceives routes, never slugs; the evaluator is the single source of the report header line. -
/plan newapplies the research budget with a fixed time limit (#1922): each research agent checks a 4-minute wall-clock limit itself, because/planwaits in the foreground and the coordinator cannot stop a running agent; the call limits of the/planbudget stay. Headless and autonomous runs cannot raise the limits, every research prompt carries them, and PRD Section 2 gets a "Research Budget" subsection with allowed and used calls per agent.skills/plan/mode-new.mdis the third place that states thecross-reposrule, and the test holds all three equal. A headless/plan newrun under the budget is not yet measured, so #1922 stays open. -
/evolvereads owner decisions (lead-session-orchestrator#948):evolve.extra-sourcesacceptskind: owner-decisionsand turns a recurring deviation pattern from the export into anowner-preferencecandidate, only above the thresholds inskills/evolve/references/evolve-analyze-mode.md. The candidate passes the same confirmation as every learning, and such a learning never grants an approval. -
Agent caps are enforced, not just written down (#1959): the new PreToolUse hook
pre-agent-cap.mjsonAgentcounts this session's running agents (subagent sidecars minus finished ones, plus same-message dispatches claimed under a lock), classifies them as reading or process-starting from the agent's tool list, and denies a dispatch above the cap from the accepted Auflagen file, else the Standard-Auflagen (2 + 2) oncefallback_due_athas passed; without fleet integration there is no cap. Below 15 % free memory (released above 20 %) no new process-starting agent starts; unmeasurable memory says „nicht messbar“ and keeps the file cap. It only blocks starts, never kills, and fails open on its own errors.- The transcript is read only when the count can change the verdict (#1965): an engaged brake denies a process-starting dispatch before any read, so a slow parse under memory pressure can no longer run into the 5 s hook timeout (which lets the dispatch through), and the "below the cap" shortcut counts only the dispatch's own class. A dispatch from a subagent counts like the coordinator's: nested agents land in the same
subagents/dir (16 of 4734 subagent transcripts in 14 days started agents, every nested sidecar was there, measured 2026-11-06), so the cap counts the whole agent tree of a session. A nested agent is released by its terminal notification in the main transcript (44 of 45 in 14 days; the rest still waits for the 30-min TTL); this also frees its file scope inpre-task-scope-disjoint.
- The transcript is read only when the count can change the verdict (#1965): an engaged brake denies a process-starting dispatch before any read, so a slow parse under memory pressure can no longer run into the 5 s hook timeout (which lets the dispatch through), and the "below the cap" shortcut counts only the dispatch's own class. A dispatch from a subagent counts like the coordinator's: nested agents land in the same
-
Context threshold for every session (#1960):
post-tool-context-threshold.mjs(PostToolUse) reads the last main-chain assistant line of the session's own transcript and adds one hint at 700k („Übergabe-Datei vorbereiten, keine neue Welle mehr beginnen“) and one at 800k („am nächsten MR-Schnitt übergeben“). Session Configcontext-threshold:(enabled,prepare-k,handover-k; defaults true/700/800). Subagent tool calls never spend the hint; an unmeasurable context writes no marker. The fleet check-in accepts an optionalkontext_k. -
ux-evaluator rubric-v2: a person's judgment before the counters (#1961): six human checks per page and viewport (five-second, primary-action, dead-or-locked, empty-state, density, decoration) run before the mechanical checks, which stay only as a footnote and never as the overall verdict. A report without human-check verdicts for every page × {1440, 390} is incomplete (
BLOCKED); the target is labelled live / production-build / dev-server.scripts/lib/test-runner/human-checks.mjsgives a deterministic floor from the AX snapshot (placeholder areas, dead or locked controls, competing actions)./testtakes the rubric from the profile: web-gate uses v2, the mac profiles keep v1. -
Issue balance per session (#1947): session-end prints „Issues dieser Session: +N angelegt, −M geschlossen, Saldo …“ with population and measurement time (
scripts/issue-balance.mjs); anything unmeasured reads „nicht messbar“, never 0. LOW findings that get filed go into the session's one[Backlog-Sammel]collector; MEDIUM and above stay single issues. -
„Entscheidungen für dich“ and „Nicht geprüft“ (#1950): mandatory sections (up to five points, „keine“ allowed) in the close report and the handover template; plan tasks may name an „Erwartung“ the report compares against.
-
The Stop notification is a per-user switch (#1936):
owner.yamlnotifications.stop: false(or envSO_STOP_NOTIFICATION=off) stopson-stopfrom sending the desktop notification (OSC 9 + OSC 777, "Session stopped — your turn"). The rest of the Stop path runs unchanged; before, the only way out was disabling the whole hook. Default stays on.on/off/1/0are accepted as well as booleans, because YAML reads a bareoffas a string. Measured 2026-11-03: with 10+ parallel sessions in Ghostty, which renders both sequences, every Stop raised two banners. Documented indocs/owner-config-schema.mdandscripts/lib/owner-config.example.yaml. -
Pre-push reuses a Full Gate receipt for the pushed tree (#1897, SO side of N4): before every route,
.husky/pre-pushrunsrun-quality-gate.mjs --receipt-check --pushed-sha <sha>in its temp clone, a read-only probe that passes the push on a green receipt for exactly this clean commit tree (exit 0) and otherwise exits 10 without running, emitting or writing anything, so the remote, partial and full routes run unchanged.- Only
context: 'pushed-tree'records count: the hook's local fallback mints them with the new write-only--receipt-record(always runs the suite, never reads the store or navigator, writes only on a clean tree whose store is another checkout), the remote router after a verified green run (with the gate host's alias and Node). A Quality-wave or session-end receipt from the working copy never answers pre-push. - The key neutralises Node (
gateDefinitionAnyNode), because the router already accepts any gate host and the hosts differ (measured 2026-11-03: m5 v26.13.0, win-wsl v24.21.0, M4 v24.22.1); the Quality wave and session-end keep the full definition. The store is the real repo's (--ledger-root); the partial gate and the mirror smoke never read or write receipts, and pre-push never touches navigator. - A hit carries
reused_from: {tree, at, host, key: 'any-node'}, and every hook-run Full Gate event carriespushed_sha. Applies tomainand tags too; CI stays authoritative. The hook feature-detects--receipt-recordin the pushed tree's runner, so an older commit keeps the old invocation. Measured before: 195 of 200 Full Gate runs since 2026-10-20 were pre-push runs, 9 re-ran a tree already green within 12 h, 0 reused a receipt.
- Only
-
Trusted publishing (ADR-0018 route B) is built, opt-in (#1906):
release.mjs --publish --trustedpushes the tag and the curated snapshot to GitHub, where.github/workflows/publish.ymlpublishes with provenance after the owner approves the run in the GitHub environmentnpm-publish; no npm token is involved. The workflow runs on a pushed tagv*only, and itsguardjob ends the run without publishing when the version is already on npm.scripts/lib/release-trusted.mjsverifies the receipt from public registry data (trusted-publisher marker, provenance workflow, ref, commit and tarball digest, per-file content comparison);--reconcileand--resume-tailuse it for a trusted proof, andrelease --check --trustedreports the trusted-publisher evidence in thepublish-routerow (a plain--checkshows only whether the workflow is armed). The receipt also requires the downloaded tarball's sha512 to equaldist.integrity; undermirror: curated,--resume-tailand--reconcilecheck the snapshot commit'sSource-Commit:trailer before pushing the GitHub tag or confirming a receipt. Without approval in time the script endsawaiting-workflow(exit 1); approve, then--resume-tail. The default--publishstays route C. Pending: the owner's Trusted Publisher setup on npmjs.com, the first rehearsal and the later default switch. -
Codex
SessionEndis wired (#1919):hooks/hooks-codex.jsonregisterson-session-end.mjsforSessionEndwith the upstream maximumtimeout: 3;SessionEndis native and required inscripts/lib/codex/plugin-contract.mjsandcheck-hooks-symmetry; underSO_PLATFORM=codexthe telemetry flush is capped at 700 ms so the hook stays under the 3 s cap. Measured live twice on codex-cli 0.160.1 (2026-11-02, plugin installed from the branch into a throwawayCODEX_HOME): the hook fired, released the lock on its own (the model turn ended in a 401, so Stop never ran) and left no registry entry; the second run raised no peer question (#1874 point 2).Interruptstays unwired and unmeasured. -
Codex version check (#1919):
scripts/lib/codex/version.mjs(MIN_CODEX_VERSION0.160.0,checkCodexVersionwith the three statesok/warn/unmeasurable— a binary that cannot be run is reported asCodex-Version nicht messbar (…), never as a warning);codex-install.mjsimports it, andruntime-readiness --codexadds an advisorycodexkey and line without touching the exit code. -
/planresearch budget (#1922): every Explore agent of the research phase gets a stated budget (web agents at most 6 WebSearch and 6 WebFetch calls, 15 local calls, about 4 minutes; local agents the local figure), fixed in headless and autonomous runs; at the budget the agent stops and reportsNot checked: …, which becomes the finding'sNext check; the recap and the Epic comment carry aResearch:line. Measured basis: one fresh-install/plan newrun made 217 web calls in 40 minutes without reaching a PRD (2026-11-02, #1811). -
One
cross-reposresolution rule (#1922):/,~/and Windows drive paths are paths and used as written, any other entry is a bare name under~/Projects/<name>, and a missing folder is a hint, never an abort — stated verbatim indocs/session-config-reference.mdand session-start Phase 5 (pinned bytests/skills/cross-repos-rule.test.mjs), implemented byresolveCrossRepoEntry()inscripts/lib/config/cross-repo.mjs; the dispatcher (enumerate.mjs) resolves declaredcross-repo.projectsentries by the same rule under its confinement root and skips missing folders. -
The Full Gate receipt also lives in the navigator gate-run ledger (#1897): on a clean tree (the content tree equals
HEAD^{tree}),run-quality-gate.mjs --variant full-gate --receiptreads the navigator ledger as gateso:full-gatewhen the SO-local store has no receipt, and writes every real run there withquelle: so.- A green entry is written only when typecheck, test and lint all passed; a run with a skipped check writes nothing, so a ledger receipt always stands for three real runs. A reuse writes nothing either, because the ledger has no time window; SO checks the 12-hour window itself.
- A reuse from the ledger is visible as
reused_from: {tree, at, source: 'navigator', entry_sha256}. A dirty tree never reaches the ledger. - Without navigator (not installed, or
SO_NAVIGATOR_BIN=off) nothing changes. Any other failure prints one stderr line and never changes the gate's result.
-
/evalnames reused gate runs (#1897, L8): the verification-evidence and gate-health lines add how many passing gate events reused a receipt instead of running. The rubric-v3 formulas are unchanged. -
release.mjs --publishstages the release and waits for the owner's approval (ADR-0018 route C, owner decision queue #7906, #1906): the checked archive goes throughnpm stage publishwith the token from.env.local; the staged version is proven by the+ pkg@versionline ornpm stage list, the publish proof carriesstaged: {at, stageId}, and the script stops with the next steps (owner approval on npmjs.com with 2FA, filed as anowner-handlungqueue item; then--resume-tail --target X.Y.Z, which reconciles the registry before it tags or pushes).--publish --directkeeps the pre-2027 route. Thenpm-token-publishrow no longer demands Bypass-2FA for the staged route (stage-capable), and thepublish-routerow has no January-2027 sunset for it; both stay as before under--direct. A new preflight rowstaged-proof-absentrefuses a second--publishwhile the proof carries astagedblock; an exit 0 ofnpm stage publishthat matches neither receipt signal still writes the proof withstaged.confirmed: falsebefore it reports, so--resume-tailcan finish after the approval. ADR-0018 is Accepted: route B (Trusted Publishing from the curated mirror) is due by 2026-12-15. -
Behaviour test for the site's command rewrite (#1879):
tests/site/commands-rewrite.test.mjsruns the realsite/assets/commands.jsinsidenode:vmwith a stub document and checks theMENTIONrewrite through an inertwindow.__soTestseam; removing the lookahead boundary, the lead group or the,/)members turns it red; no DOM library added, the script's cache-bust bumped to?v=20261102. -
The owner-config probe names a loose
owner.yaml(#1636 point 8): when the file carries any group/world bit or its directory is group/world-readable or -writable, the session-start banner prints the measured modes and the exactchmod 700 '<dir>'; chmod 600 '<file>'fix (paths shell-quoted); silent when tight, on win32, when the file is absent or on a stat error; the probe's telemetry carriesloose_mode: 0/1. -
Three security-delta contract tests (#1912): an already-expired deadline sends no OSV request and reports
degraded/timeoutwith the unchecked count; a detail-fetch timeout is classifieddeadlineand a connection resetnetwork, the finding kept with severityunknown; the CLI text line printsunchecked=<n>for a degraded check. The remaining acceptance bullets of #1912 were already covered (182 tests, measured 2026-11-02). -
Codex
apply_patchwrites reach the Edit/Write guards (#1913):enforce-scope,config-protection, the sessions-ledger guard and the STATE.md guard now run on every file a Codexapply_patchtouches, through the fail-closed adapterhooks/codex-apply-patch-guard.mjs(patch parsing inscripts/lib/codex/apply-patch-parse.mjs). It is wired as theapply_patchPreToolUse group inhooks/hooks-codex.jsonand listed inSAFETY_HOOKS. The payload shape was derived from the codex-cli 0.160.1 binary and an external report, not from a live-logged payload; logging one is the follow-up #1914. Proposal and report by GitHub user dviolante (GH#83); the implementation was written here. -
ADR-0017 fixes the destructive guard's target and non-goals (#1854, owner decision queue #5975 option 1): the guard protects against irreversible loss of uncommitted or untracked work and shared working-copy state (the rule ids in
.orchestrator/policy/blocked-commands.json); lexical variants the tokenizer can read and frequently measured forms are in goal, payloads it cannot see (variable payloads, other interpreters,xargs,find -exec, script files, nesting past the caps) are non-goals. The open guard follow-ups are classified against it. -
check-direct-renamevalidator (#1865):validate-pluginfails when a file underscripts/orhooks/gains a directrename/renameSynccall outside the shared atomic writers; a per-file allowlist (24 calls in 19 files) can only shrink, and an intentional non-write rename carries// direct-rename:ok <reason>on the same line or a comment-only line above. -
so-hud
outwardHeadless: deny(#1817): an opt-in userConfig switch (defaultallow) that refuses outward actions (push to main, force, tags, merge, release, publish, deploy) in runs with no surface instead of letting them through with a receipt; the refusal writes receiptohne-zeichenflaeche-abgelehntand names the way to allow the call. A hurdle, not a guarantee: protected branches and CI stay authoritative. Claude Code only. -
release --checkrowpublish-route(#1906): reports which publish route is configured (token,trusted-publishing,none); the token route turns red from 2027-01-01, when npm ends direct publishing with Bypass-2FA tokens, with an injected clock (SO_RELEASE_NOW) for rehearsals. The OIDC workflow.github/workflows/publish.yml(tag guard,npm audit signatures) was added as an inert draft here and is armed on tag push by the route B entry above; the owner decision is queue #7873. -
.claude/rules/bash-harness-pitfalls.md§ 7 (#1890):writer | grep -qunderpipefailis randomly red (SIGPIPE 141); usegrep -q PATTERN <<< "$var"or a file.