@event4u/agent-config 16.4.0 → 17.0.0

package/src/config/skill-fork-pilot.json

@@ -0,0 +1,17 @@
+{
+  "_comment": "The fork pilot's skill set, road-to-skills-that-arrive-when-called 3.2. `context: fork` is the FIRST host's own frontmatter key, so this is a projection-time render for that host only and never a package-owned key on an authored skill — 1.2 forbids the second thing, and the whole reason the package key was renamed to `context_window` in 1.1 was to give this key back. Every name here must be `report` in skill-output-shape.json (D2: only a report skill may fork, because a forked run sees no conversation and leaves nothing for later turns), and `tests/install/skill_fork_pilot.test.ts` fails if one is not. The cap of five is the step's own: a pilot large enough to read and small enough that a harm finding costs five skills, not three hundred.",
+  "status": "pilot",
+  "max_names": 5,
+  "names": [
+    "deep-reading-analyst",
+    "project-analysis-core",
+    "project-analysis-hypothesis-driven",
+    "project-analyzer",
+    "repomix-packer"
+  ],
+  "why_these": "The five of the six D2 skills that carried the retired `context: large` key. They are the population the defect was found in, and they are all `report`. They were chosen for carrying that key, not for size: in skill-body-budget.json three of the five sit under 5,500 bytes against a catalogue maximum near 22,000, so a pilot result is not the best case for forking. `universal-project-analysis` is the sixth and is deliberately left out to keep the pilot at five and leave one untouched sibling to compare against.",
+  "hosts": [
+    "claude-code"
+  ],
+  "widening_is_posed_not_taken": "Step 3.3. Extending this set is an owner question recorded beside 3.2's result, never an agent decision."
+}

package/src/config/skill-lifecycle-probes-codex.json

@@ -0,0 +1,57 @@
+{
+  "_comment": "Second-host probes for road-to-skills-that-arrive-when-called 0.3, in the row shape of skill-lifecycle-probes.json (validated by recordProblems in src/scripts/_lib/injection_effect.ts). Every probe ran `codex exec` in a scratch project whose `.agents/skills` held only zz-*-probe skills; the user roots were the maintainer's real ones, so the omitted counts include them. Transcripts are machine-local rollout files under ~/.codex/sessions and are named, not committed. Phase 4 of the roadmap is gated on these rows for this host.",
+  "live-pickup-agents-skills": {
+    "host": "codex",
+    "state": "unobserved",
+    "host_version": null,
+    "transcript": null,
+    "date": null,
+    "expiry": "2027-01-17",
+    "reason": "Probe: a session writes a skill under the repo's `.agents/skills` mid-turn and is asked whether its system-supplied listing now carries it. Four attempts on 2026-10-19 under codex-cli 0.148.0 did not reach an answer: the first stopped on a coordination question the package's own global instructions raised, the next three ended `429 Too Many Requests` after the session start — the subscription's rate limit, so a later run can take it. The host page says changes are detected automatically and to restart if one does not appear; that is documentation, not an observation. Re-tried 2026-10-20 under codex-cli 0.148.0 with the probe reshaped so the write comes from OUTSIDE the session mid-turn (the session sleeps, a second process writes `.agents/skills/zz-late-probe`), because a probe that writes its own skill cannot run: the sandbox reports `.agents` read-only under `-s workspace-write` (session 01a156f2-4cee-74e1-ae05-c3ff3d558dd6), so the host protects the very directory the probe must write. That run stopped on the same coordination question the package's own global instructions raise; a third run with the question pre-answered (session 01a156f2-e055-7161-aebb-94ba36ab63fb) ended `429 Too Many Requests` at the subscription's rate limit before the sleep, the second rate-limited day in a row. Two things were observed on the way and are not the cell: a skill present only under the repo's `.agents/skills` IS listed (`BEFORE: zz-early-probe`), and the omission warning fired at 426 omitted skills on this machine's real user roots. The cell stays `unobserved`: a host-side rate limit is a wait, not a reading."
+  },
+  "omission-order-and-warning": {
+    "host": "codex",
+    "state": "observed-true",
+    "host_version": "0.148.0",
+    "transcript": "~/.codex/sessions/2026/10/19/rollout-2026-10-19T21-38-11-01a155ab-e506-7ee1-bd01-e5bdb7da0f40.jsonl",
+    "date": "2026-10-19",
+    "expiry": "2027-01-17",
+    "reason": "Verbatim warning on stderr: `warning: Exceeded skills context budget. All skill descriptions were removed and 554 additional skills were not included in the model-visible skills list.` With 120 repo-scope probe skills the injected list held 347 entries: the 5 system skills first, then all 120 repo-scope skills, then the user roots (`~/.codex/skills`, `~/.agents/skills`) in name order, cut off at `guideline-writing`. So descriptions go first and as a block, then entries are dropped from the alphabetical tail of the user scope while system and repo scope survive whole."
+  },
+  "implicit-disabled-counts-against-budget": {
+    "host": "codex",
+    "state": "observed-false",
+    "host_version": "0.148.0",
+    "transcript": "~/.codex/sessions/2026/10/19/rollout-2026-10-19T21-39-15-01a155ac-e1af-7643-bb1f-e47def5bf033.jsonl",
+    "date": "2026-10-19",
+    "expiry": "2027-01-17",
+    "reason": "Same 120 repo-scope skills, 60 of them with `agents/openai.yaml` `policy: allow_implicit_invocation: false`. The 60 disabled ones left the model-visible list entirely, the omitted count fell from 554 to 481 and the list grew from 347 to 355 entries. The totals do not reconcile exactly: 347 + 554 = 901 and 355 + 481 = 836, five fewer than the 841 that removing 60 predicts, and these two runs do not explain the five. A skill with implicit invocation disabled is not listed and does not occupy the budget — a resident set can be enforced by disabling implicit invocation for everything outside it."
+  },
+  "double-listing": {
+    "host": "codex",
+    "state": "observed-true",
+    "host_version": "0.148.0",
+    "transcript": "~/.codex/sessions/2026/10/19/rollout-2026-10-19T21-38-11-01a155ab-e506-7ee1-bd01-e5bdb7da0f40.jsonl; ~/.codex/sessions/2026/10/19/rollout-2026-10-19T21-39-15-01a155ac-e1af-7643-bb1f-e47def5bf033.jsonl",
+    "date": "2026-10-19",
+    "expiry": "2027-01-17",
+    "reason": "The host does not read `.claude/skills`: a probe skill placed only there in the repo was reported not available. It does read `~/.codex/skills` beside the documented `~/.agents/skills`, and a skill present under both is listed twice (every user-scope entry in both sessions appears once per root, 111 and 145 pairs) — the host never merges same-name skills, so a package that writes both roots spends the budget twice."
+  },
+  "unlisted-skill-not-in-model-list": {
+    "host": "codex",
+    "state": "observed-true",
+    "host_version": "0.148.0",
+    "transcript": "~/.codex/sessions/2026/10/20/ rollout for session 01a15751-db98-7523-a93d-c9fc3605aead; paired with 01a1573b-c7bb-7450-afca-6091f998cbe4 and 01a15755-42d0-7752-9404-8d3f27447275",
+    "date": "2026-10-20",
+    "expiry": "2027-01-18",
+    "reason": "Two-armed, taken for 5.2's arm S3 and then for the blocker `second-host-unlisted-reachability`. Under a scratch `HOME` and `CODEX_HOME` carrying the resident set of 4.1, asked which of two names are in its skills list, the session answered `RESIDENT=YES` for a resident skill and `NONRESIDENT=NO` for `docker`, which is installed on disk and carries `agents/openai.yaml` with `policy.allow_implicit_invocation: false`. Asked to INVOKE the non-resident one it answered `UNAVAILABLE` and made no tool call; asked to invoke the resident one it did not use the skill mechanism either, it `sed`-read the file it knew the path to. So what is established is narrower than `the host refuses an invocation` and wider than `nothing changed`: a policy-disabled skill is NOT in the model-visible skills list, and a model that cannot see it does not reach for it. Its body stays readable by a caller that already knows the path, which is a filesystem read and not the skill mechanism. Consequence: `unlisted` remains the factually correct bucket (installed, not listed) and must not be read as `reachable`."
+  },
+  "unlisted-skill-reached-through-finder": {
+    "host": "codex",
+    "state": "observed-true",
+    "host_version": "0.148.0",
+    "transcript": "scratch CODEX_HOME rollouts for sessions 01a15902-c78d-7aa3-a5cf-8af3541c2c52 (named), 01a15909-e404-7620-a795-25a15790880f (search-read), 01a1590a-35ac-7772-acc6-a015728db02e (unknown name), with 01a15902-2356-7a70-bc4f-86c28634249b (listing) and 01a15909-9e93-7181-8cc0-2450cb989e82 (unprompted)",
+    "date": "2026-10-20",
+    "expiry": "2027-01-18",
+    "reason": "Remedy (a) of `second-host-unlisted-reachability`. With the resident set of 4.1 plus the installer-written `agent-config-skill-finder` entry, the finder is in the model-visible list and `docker` is not, with no omission warning. Asked to use `docker` by name, the session read the finder and then `cat` the non-resident SKILL.md and followed an instruction on its last line; asked to find and follow the fitting package skill for a task, it ran `agent-config route:explain --skills`, read `sql-writing` whole and followed its last-line instruction; an unknown name was refused with the nearest real names. A first named run that did a ranged read missed the last line, which is why the finder now prescribes a whole-file read. Not observed: a model given an ordinary task with nothing pointing at skills consulted no skill at all, finder included — the `model-suggested` path 4.2 does not count. Full record: agents/evidence/analysis/second-host-finder-probe-20261020.md."
+  }
+}

package/src/config/skill-lifecycle-probes.json

@@ -0,0 +1,57 @@
+{
+  "_comment": "First-host probes for road-to-skills-that-arrive-when-called 0.2, in the row shape of host-injection-effect.json (validated by recordProblems in src/scripts/_lib/injection_effect.ts) plus a host and an expiry. An `observed-*` row needs host_version, transcript and date; the transcripts are machine-local session files under ~/.claude/projects and are named, not committed. Every probe skill was named zz-arrival-probe-*, lived under the user skills root only for the probe, and was removed afterwards. Phase 4 of the roadmap is gated on rows (a) and (b) here being observed-true.",
+  "a-same-turn-pickup-symlink": {
+    "host": "claude",
+    "state": "observed-true",
+    "host_version": "2.1.302",
+    "transcript": "~/.claude/projects/-private-tmp-claude-501-sp-probe-proj/069ab1f1-9f18-4ca0-9046-0cba80ee0c50.jsonl; ~/.claude/projects/-private-tmp-claude-501-sp-probe-proj/2eff55cb-a222-47b5-9a1e-e63e37c3fc4a.jsonl",
+    "date": "2026-10-18",
+    "expiry": "2027-01-16",
+    "reason": "Probe (a), symlink. A `claude -p` session ran `ln -s` into ~/.claude/skills and then called the Skill tool by name in the same turn. In the first session the call returned `Launching skill: zz-arrival-probe-a2` on the first attempt. In the second (2eff55cb) the first call returned `Unknown skill: zz-arrival-probe-d1`, the skill appeared in the listing right after, and one retry launched it. So pickup is same-turn but races the directory watcher: a materialiser must expect one miss and retry once, never assume the first call lands. A second, unrelated live session on the same machine also received both new skills in its listing mid-turn."
+  },
+  "a-same-turn-pickup-copy": {
+    "host": "claude",
+    "state": "unobserved",
+    "host_version": null,
+    "transcript": null,
+    "date": null,
+    "expiry": "2027-01-16",
+    "reason": "Probe (a), copy. Both attempts to `cp -R` a skill folder into ~/.claude/skills were refused by the host's auto-mode permission classifier (once inside the probe session, once from the lane that ran the probes), so the copy case was never written. Needs a run with a permission rule for that copy, or the owner running it."
+  },
+  "b-removed-skill-second-session": {
+    "host": "claude",
+    "state": "observed-true",
+    "host_version": "2.1.302",
+    "transcript": "~/.claude/projects/-private-tmp-claude-501-sp-probe-proj/069ab1f1-9f18-4ca0-9046-0cba80ee0c50.jsonl; ~/.claude/projects/<package checkout>/a0033114-4e57-4f4f-a961-0924a2dad4a5.jsonl",
+    "date": "2026-10-18",
+    "expiry": "2027-01-16",
+    "reason": "Probe (b). Listing: an entry is never withdrawn by a removal — both a resumed session and a second live session kept listing the removed skill. Already-loaded body: it stays in the conversation (the resumed session restated the token without a tool call). Invocation differs by process: a RESUMED session (new process) answered `Unknown skill: zz-arrival-probe-a2` after the symlink was removed although its target still existed; a session that stayed LIVE throughout still launched a skill whose symlink had been removed while the target folder existed, and answered `Unknown skill` only once the target was gone too. So removing the entry does not stop a live session that already resolved it, and removing the content does — the shape step 4.4's lease must hold for."
+  },
+  "c-session-start-reload-flag": {
+    "host": "claude",
+    "state": "unobserved",
+    "host_version": null,
+    "transcript": null,
+    "date": null,
+    "expiry": "2027-01-16",
+    "reason": "Probe (c). The flag is documented: `hookSpecificOutput.reloadSkills: true` on SessionStart makes the host re-scan the skill and command directories after SessionStart hooks complete, because discovery otherwise runs before they finish (code.claude.com/docs/en/hooks, read 2026-10-18). Not observed: preparing a SessionStart hook that installs a skill was refused by the host's auto-mode permission classifier in the lane that ran (a), (b) and (d)."
+  },
+  "d-context-large": {
+    "host": "claude",
+    "state": "observed-true",
+    "host_version": "2.1.302",
+    "transcript": "~/.claude/projects/-private-tmp-claude-501-sp-probe-proj/2eff55cb-a222-47b5-9a1e-e63e37c3fc4a.jsonl",
+    "date": "2026-10-18",
+    "expiry": "2027-01-16",
+    "reason": "Probe (d). A skill carrying the retired package value `context: large` launched inline: no subagent or forked context was started, the body arrived as a message in the invoking conversation, and the model could see earlier turns including a canary word. The host neither errors on the unknown value nor forks on it, so before 1.1 the key was silently inert, not harmful — and no skill could have forked while it held the key."
+  },
+  "e-name-only-listing": {
+    "host": "claude",
+    "state": "observed-true",
+    "host_version": "2.1.296",
+    "transcript": "agents/evidence/analysis/host-skill-listing-override.md",
+    "date": "2026-10-13",
+    "expiry": "2027-01-16",
+    "reason": "Probe (e), read from the search roadmap's 0.1 evidence file rather than re-probed, as the step asks: `skillOverrides: name-only` keeps the listing entry without its description while invocation by the Skill tool and by `/name` both still work. The transcript pointer is that file, which carries the session pointers of the original runs."
+  }
+}

package/src/config/skill-output-shape.json

@@ -0,0 +1,316 @@
+{
+  "_comment": "Output shape per authored skill, road-to-skills-that-arrive-when-called 3.1. Only a `report` skill is a fork candidate (decision D2): a forked run sees no conversation and leaves nothing for later turns. Classified once by hand against `criterion`, conservatively: a skill that edits, iterates, co-authors interactively or shapes how later work is done is `standing`, and so is every case that was not clear. The shape of this file was set by an AI council on 2026-10-18 (both seats): every skill listed explicitly, set-equal to src/skills (tests/scripts/skill_output_shape.test.ts), and `default` is what a reader returns for a name missing here at runtime — never a substitute for listing it.",
+  "criterion": "report: the invocation terminates in a self-contained result — a review, verdict, audit, inventory, triage or converted document — and needs no later turn of the task to be complete. standing: guidance, a workflow or an authoring loop that shapes the rest of the task.",
+  "default": "standing",
+  "first": [
+    "deep-reading-analyst",
+    "project-analysis-core",
+    "project-analysis-hypothesis-driven",
+    "project-analyzer",
+    "repomix-packer",
+    "universal-project-analysis"
+  ],
+  "skills": {
+    "accessibility-auditor": "report",
+    "activation-design": "standing",
+    "adr-create": "standing",
+    "adversarial-review": "report",
+    "agent-docs-writing": "standing",
+    "agent-security-review": "report",
+    "agents-md-thin-root": "standing",
+    "ai-code-blindspots": "standing",
+    "ai-council": "standing",
+    "alerting-doctrine": "standing",
+    "analysis-autonomous-mode": "standing",
+    "analysis-skill-router": "standing",
+    "api-design": "standing",
+    "api-endpoint": "standing",
+    "api-testing": "standing",
+    "architecture-review-lens": "report",
+    "artisan-commands": "standing",
+    "async-python-patterns": "standing",
+    "authz-review": "report",
+    "aws-infrastructure": "standing",
+    "blade-ui": "standing",
+    "blameless-post-mortem": "standing",
+    "blast-radius-analyzer": "report",
+    "brand": "standing",
+    "brand-asset-generation": "standing",
+    "brand-audit": "report",
+    "brand-identity": "standing",
+    "brand-strategy": "standing",
+    "brand-to-tokens": "standing",
+    "bug-analyzer": "report",
+    "build-buy-partner": "standing",
+    "canvas-design": "standing",
+    "character-consistency": "standing",
+    "check-refs": "report",
+    "churn-prevention": "standing",
+    "code-intelligence": "standing",
+    "code-refactoring": "standing",
+    "code-review": "report",
+    "command-routing": "standing",
+    "command-writing": "standing",
+    "comp-banding": "standing",
+    "competitive-moat-analysis": "standing",
+    "competitive-positioning": "report",
+    "complexity-first-planning": "standing",
+    "composer-packages": "standing",
+    "condense-memory": "standing",
+    "content-funnel-design": "standing",
+    "context-authoring": "standing",
+    "context-document": "standing",
+    "contract-review": "report",
+    "contracts-cognition": "standing",
+    "conventional-commits-writing": "standing",
+    "copilot-agents-optimization": "standing",
+    "copilot-config": "standing",
+    "corpus-grounding": "standing",
+    "customer-research": "standing",
+    "dashboard-design": "standing",
+    "data-flow-mapper": "report",
+    "data-handling-judgment": "standing",
+    "database": "standing",
+    "dcf-modeling": "standing",
+    "deal-qualification-meddic": "report",
+    "decision-record": "standing",
+    "decision-review": "report",
+    "deep-reading-analyst": "report",
+    "defense-in-depth": "standing",
+    "dependency-upgrade": "standing",
+    "description-assist": "standing",
+    "design-intelligence": "standing",
+    "design-review": "report",
+    "design-system-capture": "standing",
+    "design-tokens": "standing",
+    "design-variations": "standing",
+    "devcontainer": "standing",
+    "developer-like-execution": "standing",
+    "discovery-interview": "standing",
+    "doc-coauthoring": "standing",
+    "docker": "standing",
+    "docx-authoring": "standing",
+    "dpa-review": "report",
+    "editorial-calendar": "standing",
+    "eloquent": "standing",
+    "emit-tickets": "standing",
+    "error-handling-patterns": "standing",
+    "estimate-ticket": "report",
+    "evaluate-llm-feature": "report",
+    "existing-ui-audit": "standing",
+    "expansion-playbook": "standing",
+    "experiment-loop": "standing",
+    "fe-design": "standing",
+    "feature-planning": "standing",
+    "file-editor": "standing",
+    "finishing-a-development-branch": "standing",
+    "flux": "standing",
+    "forecast-accuracy": "standing",
+    "forecasting": "standing",
+    "forensics-report": "report",
+    "form-handler": "standing",
+    "frontend-render-security": "standing",
+    "fundraising-narrative": "standing",
+    "funnel-analysis": "report",
+    "gated-reach": "report",
+    "git-workflow": "standing",
+    "github-ci": "standing",
+    "grafana": "standing",
+    "gtm-launch": "standing",
+    "guideline-writing": "standing",
+    "hiring-loop-design": "standing",
+    "history-design": "standing",
+    "html-deck": "standing",
+    "humanizer": "standing",
+    "iconography": "standing",
+    "image-analyser": "report",
+    "image-creator": "standing",
+    "image-editing": "standing",
+    "image-generation": "standing",
+    "image-provider-routing": "standing",
+    "incident-commander": "standing",
+    "jira-integration": "standing",
+    "jobs-events": "standing",
+    "js-library-packaging": "standing",
+    "judge-artifact-completeness": "report",
+    "judge-bug-hunter": "report",
+    "judge-code-quality": "report",
+    "judge-injection-defense": "report",
+    "judge-security-auditor": "report",
+    "judge-spec-compliance": "report",
+    "judge-synthesis": "report",
+    "judge-test-coverage": "report",
+    "laravel": "standing",
+    "laravel-api-endpoint": "standing",
+    "laravel-dto": "standing",
+    "laravel-horizon": "standing",
+    "laravel-mail": "standing",
+    "laravel-middleware": "standing",
+    "laravel-migration": "standing",
+    "laravel-notifications": "standing",
+    "laravel-pennant": "standing",
+    "laravel-pulse": "standing",
+    "laravel-reverb": "standing",
+    "laravel-scheduling": "standing",
+    "laravel-validation": "standing",
+    "laravel-websocket": "standing",
+    "launch-readiness": "standing",
+    "learning-to-rule-or-skill": "standing",
+    "learning-tutor": "standing",
+    "legal-intake-triage": "report",
+    "legal-practice-profile": "standing",
+    "license-compliance-audit": "report",
+    "license-compliance-borrow-check": "standing",
+    "license-compliance-credits": "standing",
+    "lint-skills": "report",
+    "livewire": "standing",
+    "livewire-architect": "standing",
+    "llm-provider-knowledge": "standing",
+    "logging-monitoring": "standing",
+    "logo-generation": "standing",
+    "market-entry-analysis": "standing",
+    "markitdown": "report",
+    "mcp": "standing",
+    "mcp-builder": "standing",
+    "md-language-check": "report",
+    "memory-consolidation": "standing",
+    "merge-conflicts": "standing",
+    "messaging-architecture": "standing",
+    "migration-architect": "standing",
+    "mobile-e2e-strategy": "standing",
+    "module-detect-on-the-fly": "standing",
+    "module-management": "standing",
+    "monorepo-workspace": "standing",
+    "motion-choreographer": "standing",
+    "multi-tenancy": "standing",
+    "nda-triage": "report",
+    "nextjs-patterns": "standing",
+    "okr-tree-modeling": "standing",
+    "onboarding-design": "standing",
+    "onboarding-program": "standing",
+    "one-on-one-cadence": "standing",
+    "openapi": "standing",
+    "operational-readiness": "report",
+    "org-design": "standing",
+    "overbuild-review-lens": "report",
+    "override-management": "standing",
+    "pdf-tools": "standing",
+    "perf-feedback-craft": "standing",
+    "performance": "standing",
+    "performance-analysis": "report",
+    "persona-improvement": "standing",
+    "persona-writing": "standing",
+    "pest-testing": "standing",
+    "php-coder": "standing",
+    "php-debugging": "standing",
+    "php-service": "standing",
+    "pipeline-strategy": "standing",
+    "pixar-storyteller": "standing",
+    "playbook-authoring": "standing",
+    "playwright-architect": "standing",
+    "playwright-testing": "standing",
+    "po-discovery": "standing",
+    "positioning-strategy": "standing",
+    "prediction-pool-optimizer": "standing",
+    "premortem": "report",
+    "privacy-review": "standing",
+    "project-analysis-core": "report",
+    "project-analysis-hypothesis-driven": "report",
+    "project-analysis-laravel": "report",
+    "project-analysis-nextjs": "report",
+    "project-analysis-node-express": "report",
+    "project-analysis-react": "report",
+    "project-analysis-symfony": "report",
+    "project-analysis-zend-laminas": "report",
+    "project-analyzer": "report",
+    "project-docs": "standing",
+    "prompt-engineering-image": "standing",
+    "prompt-engineering-patterns": "standing",
+    "prompt-optimizer": "standing",
+    "prompt-validator": "report",
+    "quality-tools": "standing",
+    "react-native-setup": "standing",
+    "react-shadcn-ui": "standing",
+    "readme-reviewer": "report",
+    "readme-writing": "standing",
+    "readme-writing-package": "standing",
+    "reasoning-orchestrator": "standing",
+    "receiving-code-review": "standing",
+    "recursive-verification": "standing",
+    "refine-prompt": "standing",
+    "refine-ticket": "standing",
+    "release-comms": "standing",
+    "repomix-packer": "report",
+    "requesting-code-review": "standing",
+    "retention-loops": "standing",
+    "review-routing": "report",
+    "rice-prioritization": "report",
+    "risk-officer": "report",
+    "roadmap-management": "standing",
+    "roadmap-writing": "standing",
+    "root-cause-frameworks": "report",
+    "rtk-output-filtering": "standing",
+    "rule-refactor": "standing",
+    "rule-writing": "standing",
+    "runway-cognition": "standing",
+    "scenario-modeling": "standing",
+    "scene-expander": "standing",
+    "schema-review": "report",
+    "screenshot-hygiene": "standing",
+    "script-writing": "standing",
+    "secrets-management": "standing",
+    "security": "standing",
+    "security-audit": "report",
+    "security-maturity-assessment": "report",
+    "sentry-integration": "standing",
+    "sequential-thinking": "standing",
+    "server-hardening": "standing",
+    "skill-improvement-pipeline": "standing",
+    "skill-management": "standing",
+    "skill-reviewer": "report",
+    "skill-writing": "standing",
+    "song-to-script": "standing",
+    "source-discovery": "report",
+    "spreadsheet-authoring": "standing",
+    "sql-writing": "standing",
+    "stakeholder-tradeoff": "standing",
+    "standards-from-config": "standing",
+    "storybook-workshop": "standing",
+    "subagent-orchestration": "standing",
+    "supply-chain-intake": "standing",
+    "svg-authoring": "standing",
+    "symfony-workflow": "standing",
+    "systematic-debugging": "standing",
+    "tailwind-engineer": "standing",
+    "tech-debt-tracker": "standing",
+    "technical-specification": "standing",
+    "terraform": "standing",
+    "terragrunt": "standing",
+    "test-case-discovery": "report",
+    "test-driven-development": "standing",
+    "test-performance": "standing",
+    "testing-anti-patterns": "standing",
+    "threat-modeling": "standing",
+    "throughput-vs-morale-tradeoff": "standing",
+    "token-optimizer": "standing",
+    "traefik": "standing",
+    "type-design": "standing",
+    "typography-system": "standing",
+    "ui-apply-generic": "standing",
+    "ui-component-architect": "standing",
+    "unit-economics-modeling": "standing",
+    "universal-project-analysis": "report",
+    "upstream-contribute": "standing",
+    "using-git-worktrees": "standing",
+    "validate-feature-fit": "report",
+    "verify-completion-evidence": "standing",
+    "verify-repair-loop": "standing",
+    "video-director": "standing",
+    "vision-articulation": "standing",
+    "voc-extract": "report",
+    "voice-and-tone-design": "standing",
+    "wireframe": "standing",
+    "workspace-link": "standing",
+    "worktree-lifecycle": "standing"
+  }
+}

package/src/config/step-depth-table.json

@@ -0,0 +1,58 @@
+{
+    "_comment": "Ceremony depth per change class — the one table road-to-a-step-that-costs-what-it-changes 2.2 names. The class comes from src/scripts/classify_change_risk.ts, raise-only. `mode: shadow` lowers nothing; `step_receipt` reports what a lowered depth WOULD skip. The owner decided on 2026-10-21 (agents/evidence/owner-directives/2026-10-21-review-cap-and-depth-flip.md, option (d)) that a lower depth may apply only inside `flip`: R0/R1 steps whose every path is structurally safe (src/scripts/_lib/depth_flip.ts), and only once a committed reading (`flip.reading`) shows `flip.requires` met per class. Setting `mode: active` without that reading is refused by tests/scripts/step_receipt.test.ts. A class whose sampled full-depth audits (src/config/depth-audit-samples.json) break the same threshold returns to full depth automatically.",
+    "schema_version": 1,
+    "registered_at": "2026-10-21",
+    "owner": "maintainer",
+    "review_by": "2027-04-21",
+    "mode": "shadow",
+    "flip": {
+        "scope": "structurally-safe-paths",
+        "classes": [
+            "R0",
+            "R1"
+        ],
+        "requires": {
+            "min_steps": 30,
+            "max_escaped": 1,
+            "max_critical": 0
+        },
+        "reading": null
+    },
+    "full": {
+        "review": "rounds until the completion review passes (plan-review-gates § 2.5)",
+        "council": "on any open question the run cannot settle from the tree (roadmap-process-loop § 5 step 4)",
+        "evidence": "an analysis or review artefact under agents/evidence/",
+        "retro": "per merge",
+        "revalidation": "council on a stale or drifted plan"
+    },
+    "classes": {
+        "R0": {
+            "review": "deterministic gates only",
+            "council": "owner-reserved questions only",
+            "evidence": "commit message and PR body",
+            "retro": "batched per drain",
+            "revalidation": "deterministic stale check"
+        },
+        "R1": {
+            "review": "one round",
+            "council": "owner-reserved questions only",
+            "evidence": "commit message and PR body",
+            "retro": "batched per drain",
+            "revalidation": "deterministic stale check"
+        },
+        "R2": {
+            "review": "full, under road-to-a-review-that-ends' round cap once it lands",
+            "council": "full",
+            "evidence": "full",
+            "retro": "full",
+            "revalidation": "deterministic stale check; council only on detected drift"
+        },
+        "R3": {
+            "review": "full, plus a parallel independent seat",
+            "council": "full",
+            "evidence": "full",
+            "retro": "full",
+            "revalidation": "full"
+        }
+    }
+}

package/src/config/turnaround-budget.json

@@ -26,6 +26,12 @@
         "context_floor_max": "DOWN"
     },
     "empty_corpus": "fail",
+    "regression_reader": {
+        "added_at": "2026-10-17",
+        "reason": "road-to-agent-throughput-recovery step 1.3 — `blocking_share` and `context_floor_max` read red from 2026-10-01 to 2026-10-15 with nobody reading the exit 1. `--queue-root` turns that exit 1 into one self-repair record (one per verdict shape; a later red reading increments it and stamps `last_seen`), which the prompt-time hook injects as the queue line.",
+        "invocation": "./scripts-run src/scripts/probe_turnaround --store \"<the main checkout's transcript store under ~/.claude/projects/>\" --limit 10 --against-baseline --queue-root \"$(git rev-parse --path-format=absolute --git-common-dir)/..\"",
+        "note": "Both paths are explicit because the run may start in a lane worktree: there the default store is the worktree's own empty project directory, which `empty_corpus: \"fail\"` reports as an empty corpus and queues as one, never as green. The schedule that runs this line daily is machine-local (cron, launchd or a routine on the measuring machine) and is not a tracked file; the CI exclusion in `_comment` is unchanged."
+    },
     "instrument_vs_source": {
         "context_floor": "AGREES. Instrument 217,385–230,705; source analysis 218,705–230,705. The upper bound is identical and the lower differs by 0.6 %, which is the strongest available evidence that the two readers parse the same store the same way.",
         "blocking_share": "AGREES within the window difference: 60 % here, 64 % in the source.",
@@ -61,6 +67,90 @@
                 "calls_per_request": "72.67 → 93.74. REPORTED ONLY, and the movement is the denominator again exactly as `directions` records: `api_calls` is comparable (3,052 → 3,281) while `user_requests` fell 42 → 35. Evidence of nothing."
             },
             "gate_verdict": "exit 1, on `blocking_share` alone. `mean_batch_size` passed."
+        },
+        {
+            "read_at": "2026-10-16",
+            "reason": "road-to-agent-throughput-recovery step 0.2 — the two regressions `--against-baseline` has reported since 2026-10-01 (`agents/evidence/analysis/turnaround-reading-2026-10-01.md:100` left them 'for whoever owns that number next') get an owner. This entry records the reading and the ownership; it raises and lowers no baseline.",
+            "owned_by": "road-to-agent-throughput-recovery",
+            "owned_figures": ["blocking_share", "context_floor_max"],
+            "invocation": "./scripts-run src/scripts/probe_turnaround --store ~/.claude/projects/-Users-mathiasberg-projects-galawork-galawork-packages-event4u-agent-config --limit 10 --against-baseline",
+            "invocation_note": "Run from a lane worktree, so `--store` points at the main checkout's store; the bare command would read the worktree's own (empty) project directory and exit 1 on `empty_corpus: \"fail\"`.",
+            "corpus": {
+                "window": "the 10 most recently modified sessions in the main checkout's transcript store, read 2026-10-16; the measuring session wrote to a worktree-suffixed store and is not in this window",
+                "sessions": 10,
+                "user_requests": 267,
+                "api_calls": 2912,
+                "tool_calls": 2962,
+                "tool_using_requests": 2577,
+                "blocking_tail_calls": 130,
+                "blocking_tail_minutes": 555,
+                "tool_minutes": 686,
+                "note": "Ratios recompute from this block: 2912 / 267 = 10.91; 2962 / 2577 = 1.15. `blocking_share` is computed from unrounded seconds, so the rounded minutes here give 555 / 686 = 0.8090 against the recorded 0.8091; the minutes are for reading, the share is the probe's. The cause split of the blocking tail, in minutes: sleep-poll 212, ci-wait 186, network 115, unknown 38, test 3, build 1, subagent-wait 0, mcp 0, user-wait 0 (sum 555)."
+            },
+            "reading": {
+                "calls_per_request": 10.91,
+                "mean_batch_size": 1.15,
+                "blocking_share": 0.8091,
+                "context_floor_min": 212193,
+                "context_floor_max": 244472
+            },
+            "delta_vs_baseline": {
+                "blocking_share": "+0.1889 — 0.6202 → 0.8091. REGRESSION, owned by road-to-agent-throughput-recovery from this entry on. Not re-baselined.",
+                "context_floor_max": "+13,767 — 230,705 → 244,472. REGRESSION, owned by road-to-agent-throughput-recovery from this entry on. Not re-baselined.",
+                "mean_batch_size": "1.01 → 1.15, in the gated direction (UP). Not raised: a ratchet is never tightened on a single local reading.",
+                "calls_per_request": "72.67 → 10.91. REPORTED ONLY — `user_requests` rose 42 → 267, the denominator effect `directions` names."
+            },
+            "gate_verdict": "exit 1, on `blocking_share` and `context_floor_max`."
+        },
+        {
+            "read_at": "2026-10-16",
+            "reason": "road-to-agent-throughput-recovery step 1.2 — the throughput baseline the roadmap's goal is judged against, registered here rather than in a new baseline file. It is a git reading, not a transcript reading, so it carries its own window shape and does not feed `--against-baseline`.",
+            "kind": "throughput-git",
+            "invocation": "./scripts-run src/scripts/probe_turnaround --ref 085ec41ff --git 2026-09-28..2026-10-13 — first-parent commits of `main` @ 085ec41ff, dated in Europe/Berlin; a night is 22:00–08:00, a commit before 08:00 belongs to the previous evening's night. The denominator is EVERY night in the window, so a night on which runs stalled and nothing landed stays in it. Net ticked steps = added minus removed `- [x]` lines under `agents/roadmaps/` in each commit's first-parent diff, excluding files the commit creates (a roadmap authored with checked steps is authoring, not work done). A merged change is a commit whose subject ends in `(#N)` or is GitHub's merge subject.",
+            "corpus": {
+                "window": "nights 2026-09-28 .. 2026-10-13",
+                "ref": "085ec41ff",
+                "calendar_nights": 16,
+                "nights_with_a_commit": 14,
+                "commits": 108,
+                "commits_naming_a_pr": 108,
+                "true_merge_commits": 13,
+                "net_ticked_steps": 282,
+                "net_ticked_steps_including_created_files": 306,
+                "note": "Ratios recompute from this block: 108 / 16 = 6.75 merged changes per night; 282 / 16 = 17.625 → 17.63 net ticked steps per night. Every commit in the window names a PR, so commits and merged changes coincide here; `true_merge_commits` is small because PRs land as squash commits. Same method over 2026-08-10..23: 14 nights (13 with a commit), 148 commits all naming a PR, 383 net ticked (10.6 / 27.4); over 2026-08-24..09-06: 14 nights (13), 98 commits all naming a PR, 311 net ticked (7.0 / 22.2). Counting created files too gives 794 / 420 / 306, i.e. 56.7 / 30.0 / 19.1 per calendar night — the series the roadmap's Context printed."
+            },
+            "reading": {
+                "merged_changes_per_night": 6.75,
+                "net_ticked_steps_per_night": 17.63
+            },
+            "directions": {
+                "merged_changes_per_night": "UP",
+                "net_ticked_steps_per_night": "UP"
+            }
+        },
+        {
+            "read_at": "2026-10-21",
+            "reason": "road-to-a-step-that-costs-what-it-changes step 0.2 — the per-STEP baseline its Goal is read against. Thresholds were pre-registered (that roadmap's D14) in a commit before this reading.",
+            "owned_by": "road-to-a-step-that-costs-what-it-changes",
+            "invocation": "./scripts-run src/scripts/probe_turnaround --steps --limit 316 --store <every ~/.claude/projects store of this repository, comma-separated: the main checkout's and each worktree's>",
+            "invocation_note": "316 is the number of transcripts across those stores modified in the 7 days before the reading, subagent transcripts under <session>/subagents/ included. 75 of the 79 transcripts that close a roadmap step are subagent files, which is why the session files alone read 13 steps.",
+            "corpus": {
+                "window": "the 7 days to 2026-10-21, every store of this repository, measuring session excluded",
+                "transcripts": 316,
+                "closed_steps": 75
+            },
+            "reading": {
+                "step_p50_minutes": 4.2,
+                "step_p90_minutes": 33.6,
+                "share_by_class": { "model": 0.303, "change": 0.028, "verify": 0.039, "ceremony": 0.095, "wait": 0.535, "refusal": 0.0, "unknown": 0.0 },
+                "steps_by_change_class": { "R0": 4, "R1": 5, "R2": 30, "R3": 36 },
+                "p50_minutes_by_change_class": { "R0": 2.1, "R1": 2.7, "R2": 7.7, "R3": 0.1 }
+            },
+            "limits": "A step runs from the previous close in the same transcript, so a subagent that ticks several boxes in a row gives the work to the first and near-zero intervals with no edited path to the rest; those classify R3 by upward resolution, which is why the R3 p50 is 0.1 min and the R3 count is high. The anatomy measures time inside the transcript that closes the step; the orchestrating session's review rounds, PR and CI wait around a subagent's step are not in it.",
+            "directions": {
+                "step_p50_minutes": "DOWN",
+                "ceremony_share": "DOWN"
+            }
         }
     ]
 }