agent-pipeline
Skill0 starsChanged 21 days agoMITView on GitHub
Deploys a Claude Code agent pipeline into any repository (todo.md queue, a task-runner in a worktree per task, review by tiers R0–R3 decided by a script, Codex as the primary reviewer, a single bookkeeping writer, no auto-push), adapted to the project through a mandatory interview; upgrade mode updates an already deployed or hand-built .claude/** to the current templates while preserving project-specific work; audit mode checks an existing .claude/** against the principles. Use on "set up the agent pipeline", "deploy the do-all pipeline in this repo", "update/reinstall/rebuild the pipeline", "/agent-pipeline init", "/agent-pipeline upgrade", "/agent-pipeline audit", "audit our pipeline against the principles"; also on the Russian phrases «разверни агентный flow», «настрой конвейер do-all в этом репо», «обнови/переустанови/пересобери конвейер», «проверь наш пайплайн по принципам».
- Reads credentials
- Commits and pushes
2 other projects use a version of this.
Adopted by
Most recently taken up by
- nicksonthc/claude-skilltheir own version, first seen Nov 1, 2026
- JS-mark/skillstheir own version, first seen Jun 4, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- JS-mark/skills★ 2same name
- nicksonthc/claude-skill★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 2 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 13
Before starting, read SKILL_DIR/reference/principles.md (why it is this way) and SKILL_DIR/reference/knobs.md (the interview matrix — the …
- Very specific: line 76
Tools — right after the table, before the interview: No Codex CLI or no login → offer to install per tools.md §1: the commands and where …
- Distinctive: line 90
After the answer — one line "accepted: … → changes …" (what exactly in which files).
- Very specific: line 168
Keep a "file → substitutions" list for the report.
- Very specific: line 286
Do not touch R∖T files under any circumstances.
- Distinctive: line 345
Every question — with the pros/cons of all options and a recommendation for this project with a reason.
- Very specific: line 354
Do not run /do-next, do not commit, do not push, do not change .env*, .git/config.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
agent-pipeline — init | upgrade | audit
Below, SKILL_DIR = the directory of this file (symlink ~/.claude/skills/agent-pipeline → the skill's
git repo). Everything is done in the main session with the owner; the skill itself does not
run /do-next, does not commit and does not push.
Before starting, read SKILL_DIR/reference/principles.md (why it is this way) and
SKILL_DIR/reference/knobs.md (the interview matrix — the single source of options, pros/cons
and recommendation rules) and SKILL_DIR/reference/tools.md (detection and installation of the
owner's tools by consent); for upgrade — also SKILL_DIR/reference/upgrade.md (file classes,
merge rules, U questions). Argument empty — ask init, upgrade or audit. This skill is written
in English; talk to the owner in the language they write in (knobs.md D).
init — five phases
Phase 1. Reconnaissance (read-only, ask no questions)
Collect and show the owner ONE "found → proposed" table before any questions:
- Stack and package roots:
package.json(+pnpm-lock/yarn.lock/package-lock),pyproject.toml/requirements*.txt/uv.lock,go.mod,Cargo.toml,Makefile; monorepo — several roots (workspaces,apps/*,packages/*). - Commands: tests, types, lint, build, formatter — from
scripts,Makefile, CI yaml (.github/workflows,.gitlab-ci.yml). Dev server and port (scripts.dev,vite.config,PORT=in code/.env.example,docker-compose). - Environment: gitignored env files (
git check-ignore .env .env.local .envrc), dependency directories (node_modules,.venv,target,vendor). - Git:
git remote -v;git rev-parse main origin/main(they diverge →baseRef: head);git log --shortstat -50→ median diff lines; presence of.claude/**,AGENTS.md,CLAUDE.md,.worktreeinclude,todo.md/claude-progress.md/blockers.md. Non-empty.claude/agentsor a.claude/settings.jsonwith hooks → stop: offerupgrade(update while preserving existing work) oraudit(report only), do not touch files.git rev-parse --verify HEADfails (not a single commit) → ask the owner to make the first commit before generation: the worktree and the tier script require HEAD, and the skill does not commit. Instruction files:test -L CLAUDE.md(symlink toAGENTS.md→ "⚠ will ask",knobs.mdB);grep -nE '^@(\./)?AGENTS\.md' CLAUDE.md(import already present →{{AGENTS_IMPORT}}is empty); nested ones —git ls-files '*AGENTS.md' '*CLAUDE.md';AGENTS.override.md,AGENTS.local.md,.agents/(Claude Code does not read them — into the table as a fact);claude --version(AGENTS.mdis read natively since 2.1.277 and only when there is noCLAUDE.md— that is why the pipeline keeps the import,principles.md§6). - Candidate docs for
CANON_DOCS:AGENTS.md,ARCHITECTURE.md,README.md,DESIGN.md,docs/index.md, design guide (design_guidelines.md,docs/ui/*), API doc, plans directory; for each — the date of the last change (git log -1 --format=%cs -- <file>). Rules of other tools —.cursorrules,.cursor/rules/,.github/copilot-instructions.md,.windsurfrules,.clinerules— read as a source of invariants and commands, do not include inCANON_DOCS. Language signals forPIPELINE_LANG(knobs.mdD): the prose language of these docs, of existing state files and ofgit log -30 --format=%s; they agree with the owner's conversation language → a fact in the table, otherwise "⚠ will ask". - Critical path candidates — by the heuristics of
knobs.mdA1 (git ls-files | grep -iE 'auth|session|payment|billing|schema|migration|docker-compose|Dockerfile|workflows|secret'), plus the domain core from the invariants inAGENTS.md/README. - UI:
client/,src/components,.tsx/.vue/.svelte; Playwright (playwright.config*, MCPplaywrightinclaude mcp list). UI present → the number of raster assets (git ls-files | grep -cE '\.(png|jpe?g|webp)$') and the "Assets" knob (knobs.mdB) — "⚠ will ask". - Owner's environment:
command -v codex python3 jq; the Codex CLI and its login, theimage_generationfeature, MCPrembg— with the detection commands oftools.md§2 (the last two — only with UI from item 7);~/.codex/config.toml(model,model_reasoning_effort), the catalog~/.codex/models_cache.json;pwdcontainsDropbox. Model freshness perknobs.mdE: the Codex catalog and the wrapper's effort set (§2, local) and the Claude snapshot against Anthropic's official pages (§3, network) → a "Models" row in the table: "configured → status (current / flagged: why / snapshot stale) → proposed". - Signs of parallel agent sessions (
AGENT_ID, several DBs in compose) and of an external tracker (Jira keys in commits) — they enable the optional questions ofknobs.mdB.
Table format: one row per item — "what was found (file/value) → what will be substituted (placeholder)". Mark anything ambiguous "⚠ will ask". Do not move on to the interview without this table (rule: explain first, then options).
Tools — right after the table, before the interview. No Codex CLI or no login → offer to install per tools.md §1: the commands and where they write — in the chat,
consent — one AskUserQuestion, install only what was checked, afterwards — repeat detection and
an updated table row (recommendation A2 depends on it). Refusal or installation failure — an
[OWNER] line in the report, the interview goes on. rembg and image_generation are offered
later — after the answer to the "Assets" knob (knobs.md B).
Phase 2. Interview — mandatory, must not be skipped
Five questions knobs.md A1–A5, one at a time, each strictly in this format:
- In the chat: the mechanics of the fork (2–3 lines); options with "+" and "−"; the line
"Recommended for this project:
<option>— because<fact from reconnaissance>". AskUserQuestion: the recommended option first, labeled "(Recommended for this project)";multiSelectfor lists (critical paths, doc triggers).- After the answer — one line "accepted: … → changes …" (what exactly in which files).
Order and content:
- A1 critical paths →
CRITICAL_GLOBS. - A2 review models — the review backend (Codex + Claude reader / Codex + Fable reader / Claude
only / Codex only), then the Codex model and effort by tier (
CODEX_MODEL_R12,CODEX_MODEL_R3,CODEX_EFFORT_R12,CODEX_EFFORT_R3;gpt-6.1-solorgpt-6-astraon R3 and the effort — by theknobs.mdA2 rule from the A1 domains, the R3 share and the A3 profile; flagged models never), the model ofreviewerandreview-consolidator. - A3 strictness profile —
standard/strict/light→REVIEW_PROFILE. - A4 executor models and effort —
task-runner,coder,architect,researcher,tester;effortLevelglobally and per-roleeffort:only when non-default; the fallback family ({{FALLBACK_MODEL}}). One screen with 4 options: "default (Opus for runner/coder/architect/reviewer, high; fallback Fable)" / "Fable for architect and reviewer" / "everything on Fable (fallback Opus)" / "default +xhigh/maxfor reviewer/architect" — plus Other. The family names follow theknobs.mdE leader rule: when the snapshot is stale, the screen is built from the current leader. - A5 docs and UI —
committertriggers, design guide, canonical docs, one line about the project.
Optional questions (knobs.md B) — only those that reconnaissance marked "⚠ will ask". Everything
else — the defaults of knobs.md C, without questions. The outcome of the phase — a
"setting → value → file" table in the chat; the owner confirms with one word, then generation.
Phase 3. Generation
From SKILL_DIR/templates/ into the target repo, substitutions — per the phase 2 table. Mechanics:
- Language. Templates are English.
PIPELINE_LANG = en→ copy as is. Otherwise translate the prose of every generated.md(agents, skills, both sections, state-file headers) intoPIPELINE_LANGwhile substituting, and write the text values ofknobs.mdC in that language; the contract strings listed inknobs.mdD stay English (script and JSON content, frontmatter exceptdescription:, markers, verdicts, report headings and field labels, section headings)..sh,.jsonanddocs/agent-pipeline-principles.mdare never translated.{{PIPELINE_LANG}}→lang=in both section markers;{{LANG_RULE}}→AGENTS.mdsection (empty foren). .claude/agents/*.md,.claude/skills/{do-all,do-next,status}/SKILL.md,.claude/scripts/*,.claude/hooks/*— copy + replace{{NAME}}(in.md) and edit the PROJECT blocks (# === PROJECT-SPECIFIC (agent-pipeline init) === … # === /PROJECT-SPECIFIC ===in.sh):CRITICAL_GLOBS,CODEX_*,REVIEW_PROFILE, thresholds,PROTECTED_FILES,FORMAT_CMD/FORMAT_EXT_RE/FORMAT_BIN_CHECK,STATE_RE. Replace the marker lines# {{CRITICAL_GLOBS}}/# {{PROTECTED_FILES}}with array elements, one per line, in quotes.model:/effort:in frontmatter — per A2/A4. UI + Playwright MCP —{{UI_MCP_TOOLS}}in the tester'stools:(the visual check is done by the tester, not the runner). The "Assets" knob —{{ASSET_TOOLS}}and{{ASSET_RULE}}incoder.md,{{ASSET_SECTION}}in theCLAUDE.mdsection (values —knobs.mdC;codex-image.shis always copied). Multi-line placeholders ({{DOMAIN_CHECKLIST}},{{UI_CHECKLIST}},{{DOC_TRIGGERS}}) — as ready text;_QUOTED— the same lines with the prefix>; empty value — remove the placeholder line and the empty line after it..claude/settings.json:__STACK_ALLOW__→ the stack's allow lines;__FALLBACK_MODEL__→ the fallback family from A4;__ASSET_ALLOW__→mcp__rembg__*lines whenrembgis in the "Assets" knob, otherwise the line is removed; if the file already exists — merge with apython3script (notjq '*': it overwrites scalars and arrays with the right-hand document): scalars — the owner's value stays, template keys are added only when missing;permissions.allow/denyandenv— union (no duplicates, owner first);hooks.<event>— a template entry is added if no existing entry has the samecommand, otherwise it is skipped. Show the owner the diff before writing..gitignore— append the lines fromgitignore.appendthat are missing..worktreeinclude— from the template, PROJECT block per reconnaissance; if the file exists — append what is missing.AGENTS.md/CLAUDE.md: exists — append the section fromAGENTS.section.md/CLAUDE.section.mdto the end (do not create duplicate headings); absent — create from the template.{{STACK_SUMMARY}},{{COMMANDS_TABLE}},{{INVARIANTS}},{{CRITICAL_SUMMARY}},{{PROTECTED_SUMMARY}}— from reconnaissance;{{SKILL_REV}}—git -C SKILL_DIR rev-parse --short HEAD. Keep the markers<!-- agent-pipeline:begin … -->/<!-- agent-pipeline:end -->from the section templates:upgradeuses them to find and replace its own section.{{AGENTS_IMPORT}}in theCLAUDE.mdsection — the bare line@AGENTS.md(no backticks and no indent — otherwise Claude Code will not expand the import); empty if the owner already importsAGENTS.mdoutside the section.CLAUDE.mdis a symlink toAGENTS.md→ per the answer to theknobs.mdB question: replace the symlink with a regular file (import + section) or stop; do not write both sections into one file.todo.md,claude-progress.md,blockers.md— fromtemplates/state/only if they do not exist; they exist — show the owner the instruction header from the template and offer to insert it andSMOKE-1.SKILL_DIR/reference/principles.md→docs/agent-pipeline-principles.mdof the target repo (CLAUDE.mdrefers to it); there is a docs directory with a different name — put it there.chmod +x .claude/scripts/*.sh .claude/hooks/*.sh;mkdir -p .claude/state .claude/worktrees(+{{EXEC_PLANS_DIR}}/.gitkeep, if the directory does not exist); Dropbox guard perverify.md§6.- Keep a "file → substitutions" list for the report. Do not touch files outside
.claude/**,.worktreeinclude,.gitignore,AGENTS.md,CLAUDE.md,docs/agent-pipeline-principles.md,{{EXEC_PLANS_DIR}}/.gitkeepand the three state files.
Phase 4. Verification
Strictly per SKILL_DIR/reference/verify.md §1–§7: nine test matrices green, review-tier.sh
--task-id SMOKE-1 yields TIER=R1 and PROFILE=, hooks block/print, worktree add/remove,
git check-ignore, codex-review.sh check, Dropbox xattr, grep '{{[A-Z_]*}}' empty.
Red — fix it (the template or the
substitution) until green; the foreign tree is not changed in the process.
Phase 5. Report and handoff
In the chat:
- A table of created/changed files with substitutions.
- An "accepted by default → where to switch" table (from
knobs.mdC). [OWNER]actions: fromtools.md— what the owner declined to install or what failed to install (with commands),codex login;/autocompact 400kin user scope if desired, a commit of.claude/**+AGENTS.md/CLAUDE.md+ the state files (done by the owner or, at the owner's word, bycommitterwith an explicit list of paths; subjectchore(pipeline): set up agent pipeline (agent-pipeline init)), a restart of the Claude Code session (hooks and skills are read at startup).- The first run —
verify.md§8:/do-next SMOKE-1and what to check. - Where to tune things later: the Codex profile and model —
review-tier.sh; role models — frontmatter, the fallback —fallbackModel+{{FALLBACK_MODEL}}lines; revisiting the profile and the new models — byfound=/rounds=/codex=in.claude/state/runs.log; an aged Codex model shows up asMODEL_WARNin the runner's report, an aged Claude choice — in/agent-pipeline audit. A staleknobs.mdE snapshot → a line "updateknobs.mdE in the skill repository".
upgrade — update a deployed or foreign pipeline
For a repo where .claude/agents is non-empty or settings.json has hooks. The pipeline there is
either "own" (installed by a previous version of the skill: PROJECT blocks in .sh, the
agent-pipeline:begin marker in CLAUDE.md) or "foreign" (built by hand or by another tool). The
mechanics of classification, merging and migration — SKILL_DIR/reference/upgrade.md; here — the
procedure. The same five phases as in init; only the differences are listed below.
Phase 1. Reconnaissance
Everything from init items 1–9 (the "stop" rule from item 4 does not apply) and the "Tools" step after the table, plus:
- Tree cleanliness:
git status --porcelain -- .claude CLAUDE.md AGENTS.md .worktreeinclude .gitignore todo.md progress.md claude-progress.md blockers.mdis empty. Not empty → stop: the owner commits or sets the changes aside (git stash) themselves; upgrade does not start until the old versions are in git. - Origin: the marker
<!-- agent-pipeline:begin skill=… date=… lang=… -->inCLAUDE.mdand PROJECT blocks in.claude/{hooks,scripts}/*.sh→ "own" (name the version, date and language); none of this → "foreign". The current skill version —git -C SKILL_DIR rev-parse --short HEAD.lang=is the pipeline'sPIPELINE_LANG— keep it; nolang=→ detect perknobs.mdD (upgrade.md§0). - Inventory of
.claude/**,CLAUDE.md,AGENTS.md,.worktreeinclude, the state files — each path into a class ofupgrade.md§1:T∩R =(matches),T∩R ≠(differs; for own ones — "PROJECT block present/absent"),T∖R(will appear),R∖T(project file, not touched),R∖T ✗(name conflictcommands/<x>.md↔skills/<x>/). - Current setting values — as interview recommendations, per the table in
upgrade.md§2:CRITICAL_GLOBS,REVIEW_*,CODEX_*, thresholds from the oldreview-tier.sh;model:/effort:from the frontmatter of same-named agents;PROTECTED_FILESand secret patterns from the oldsafety-check.sh; doc triggers from the oldcommitter.md; the actual state file names (progress.md≠claude-progress.md→ "⚠ will ask"). - State files against the norms of
principles.md§6: size,[x]lines intodo.md, lines1500 characters, the progress header not on one line → "⚠ will ask" (U4).
CLAUDE.md/AGENTS.mdprose that contradicts the pipeline: direct calls of@coder/@reviewerfrom the main session, a custom/do-allloop,git add ., a different review backend, other state file names; rules duplicated in both files (after the@AGENTS.mdimport Claude sees them twice) — a list of lines with numbers. Into the report as[MANUAL], do not edit.
The "found → proposed" table — as in init, plus an inventory table "class → path → action".
Phase 2. Interview
A1–A5 from knobs.md are mandatory and in the same format, but "Recommended for this project" =
the current value from item 13, if there is one ("because that is what <file>:<line> has now");
the current value contradicts a knobs.md rule → recommend per knobs.md, give the current value
as the second option labeled "current". Then the U questions from upgrade.md §3 — only those
that reconnaissance marked "⚠ will ask":
- U1 differing files
T∩R ≠—multiSelect: which to keep as old; empty = replace everything (recommendation), the old goes to legacy. - U2 state file names — rename with
git mv(recommendation) / keep and substituteSTATE_RE. - U3 project agents from
R∖T— do not add to the runner'sAgent(…)allow (recommendation) / add the listed ones. - U4 state archive — the skill prepares the archives verbatim (recommendation) / the owner does it themselves.
The outcome — a "setting → value → file" table and the inventory table with final actions; the owner confirms with one word.
Phase 3. Generation
Strictly in the order of upgrade.md §4:
- Legacy copy
.claude-legacy/<YYYY-MM-DD>/of everything that will be replaced or moved (withoutsettings.local.json,state/,worktrees/,.env*). Before this step — not a single write. T∖R— as in init phase 3.T∩R ≠— replace: own → the PROJECT blocks of the old file are carried over verbatim, show the diff "old block ↔ interview value" and ask once for all files; foreign → the values from item 13 into the PROJECT blocks of the new file, the body of the old one — into the migration table.R∖T ✗— the old file goes to legacy.settings.json— merge perupgrade.md§4.5: the template's hook entry wins if itscommandpoints to a replaced or new file; entries for files that moved to legacy are removed, shown to the owner; interview scalars (effortLevel,worktree.baseRef,autoCompactWindow,fallbackModel) — from the interview, the owner's other scalars stay;allow/deny/env— union. Diff before writing.CLAUDE.md/AGENTS.md— replace between theagent-pipeline:begin/endmarkers; no markers, but there is a heading## Agent pipeline (agent-pipeline)(legacy:## Агентный конвейер (agent-pipeline)) — replace from it to the next##; nothing at all — append as in init. Do not touch the rest of the text.- State files — only per U2 (
git mv) and U4 (archives);SMOKE-1— if there is no such ID. docs/agent-pipeline-principles.md— replace entirely.chmod, directories, Dropbox guard,.gitignore,.worktreeinclude— as in init.
Do not touch R∖T files under any circumstances. Keep a "file → class → action → substitutions"
list.
Phase 4. Verification
verify.md §1–§7 plus §9 (upgrade only): legacy is outside .claude/, no
commands↔skills conflicts, every hook command points to an existing file, R∖T files are
byte-identical to HEAD, exactly one section marker, no old state file names in .claude/**.
Phase 5. Report and handoff
As in init, plus:
- The migration table of
upgrade.md§5: "legacy file → what it has that the new one lacks → where to move it". Moving the content is done by the owner or by the main session at the owner's word; upgrade itself does not move it. - The
R∖Tlist with a frontmatter lint (model:present; noallowedTools;disallowedTools: Bash(…)does not disable Bash) — report only. - The item 15 lines from
CLAUDE.md/AGENTS.md— as one[MANUAL]task (writing totodo.md— at the owner's word). [OWNER]: migrate per the table, delete.claude-legacy/, a commitchore(pipeline): update agent pipeline (agent-pipeline upgrade)with an explicit list of paths, a session restart,/do-next SMOKE-1(verify.md§8).
audit — report only, no edits
A check of the target repo's .claude/** against the principles. The outcome — a table "item →
present / absent / differs → principles.md section". Items:
| # | Check | How | |
| 1 | A runner per task: an agent with isolation: worktree, tools: Agent(<allow list>), report ≤15 lines | frontmatter of .claude/agents/*.md | |
| 2 | The orchestrator is thin: /do-all//do-next//status (skills or commands), does not write state files, ff-merge | command texts | |
| 3 | The tier is decided by a script (review-tier.*), not an agent; it prints the Codex model/effort; the profile/lineup come from the script | grep -n 'CODEX_MODEL|ROUND1|PLAN' .claude/scripts/*; hardcoded model in agents | |
| 4 | Codex is the primary reviewer via the Codex CLI (wrapper codex-review.sh, codex exec read-only); round 2 verifies in the same thread (threadId); UNAVAILABLE ≠ APPROVE | codex-reviewer*.md, bash .claude/scripts/codex-review.sh check | |
| 5 | Double review in a single message, the consolidator checks only the disagreements; two rounds, the second is final | task-runner, review-consolidator (in a hand-built setup they may be named differently — look for the consolidating role and the double-review skill) | |
| 6 | Fate of S3/S4: collector task/won't fix/⏸️; only S1/S2 spawns a task | runner prompts, todo.md header | |
| 7 | Hooks: safety-check (Bash), agent-sync-check (Agent), codex-prompt-write-guard (Write), post-edit, pre-compact, session-start, notify; registered in settings.json; tests green | jq .hooks, a run of *.test.sh | |
| 8 | settings.json: worktree.baseRef, fallbackModel, autoCompactWindow, MCP_TOOL_TIMEOUT, git push in deny | jq | |
| 9 | Frontmatter: no dead allowedTools; disallowedTools: Bash(…) does not accidentally disable Bash; model: on every agent | grep | |
| 10 | run_in_background appears in prompts only as a prohibition; the item 7 hook is present and passes a call without the key (Claude Code 2.1.309+ has no such parameter) | grep | |
| 11 | State files: only open items in todo.md, lines ≤1500 characters, the progress header is one line, archives | wc -c, awk 'length>1500' | |
| 12 | .claude/state, .claude/worktrees, /artifacts/codex-prompts in .gitignore; .worktreeinclude; Dropbox xattr | git check-ignore, xattr -p | |
| 13 | A single bookkeeping writer; ≤1 commit per runner exit; ID in the subject; no auto-push | prompts, git log --format=%s -30 | |
| 14 | Deliberate divergences (a task board instead of todo.md lines, task tiers, a value filter, telemetry in the repo, a context checkpoint, port slots) | mark "deliberately absent", not as a defect (principles.md §10) | |
| 15 | Instruction files: CLAUDE.md imports @AGENTS.md (or is a symlink to it) — otherwise Claude Code does not see AGENTS.md: it is read natively since 2.1.277 and only without a CLAUDE.md; shared rules — in AGENTS.md (Codex reads the same file from cwd), in CLAUDE.md — only Claude specifics, no duplicates; the section marker — one per file; prompts do not require a separate Read of AGENTS.md; a nested AGENTS.md next to a CLAUDE.md without an import, AGENTS.override.md, AGENTS.local.md, .agents/ — Claude does not read them; the AGENTS.md chain from the root to cwd ≤ 32 KiB (the Codex limit project_doc_max_bytes: beyond it the rules silently do not reach the reviewer); pluginConfigs."agents-md@builtin" in the project settings.json — a dead setting (it works only in user/managed) | grep -cE '^@(\./)?AGENTS\.md' CLAUDE.md, test -L CLAUDE.md, grep -c 'agent-pipeline:begin' CLAUDE.md AGENTS.md, git ls-files '*AGENTS*.md' '*CLAUDE.md' .agents, wc -c AGENTS.md, jq .pluginConfigs .claude/settings.json, `grep -rniE 'read .AGENTS | прочитай .AGENTS' .claude/ (principles.md` §6) |
| 16 | The owner's tools against what the config expects: REVIEW_BACKEND≠claude → the Codex CLI and a login; codex-image.sh in the tools: of coder → image_generation; mcp__rembg__* in tools:/allow → MCP rembg. "Differs" — only when the config expects something absent; an extra installed tool is not a defect | the detection of tools.md §2, bash .claude/scripts/codex-review.sh check; into the report — the installation commands, audit itself installs nothing | |
| 17 | Model freshness (knobs.md E): CODEX_MODEL_*/CODEX_EFFORT_* of review-tier.sh against the Codex catalog and the wrapper's effort set; model: of runner/coder/architect/reviewer against the leader rule, and the owner's Claude Code versions against the one the alias needs (opus = Opus 5.5 from 2.1.280); fallbackModel ≠ the runner's model and matches the fallback in the prompts. "Differs" — a flagged model or effort, a non-leader family without a recorded reason, fallback = primary | knobs.md E §2 commands; claude --version; bash .claude/scripts/review-tier.sh (a MODEL_WARN= line); grep -n '^model:' .claude/agents/*.md; jq .fallbackModel .claude/settings.json; grep -rn 'model: "' .claude/agents/task-runner.md .claude/skills/do-all/SKILL.md |
For every "absent/differs" — one line: what, where, which principles.md section it refers
to, what to change (no edits — report only). At the end — the three most costly divergences and
a proposal: file them as [MANUAL] tasks in todo.md (writing — only at the owner's word).
Rules
- Reconnaissance → table → interview → confirmation → generation → verification → report. No phase is skipped; the interview is not replaced by defaults even on a request like "do as you see fit" — in that case all five screens are shown with the recommended option, and the owner picks the first item.
- Every question — with the pros/cons of all options and a recommendation for this project with a reason.
- The owner's existing files are not overwritten: merge (
settings.json), append (.gitignore,AGENTS.md,CLAUDE.md), skip (state files). There is already.claude/agents/*or hooks insettings.json→upgradeoraudit, notinit(the Phase 1 rule). - In
upgrade, only files that have a template counterpart (T∩R ≠) or a name conflict (R∖T ✗) are replaced, and only after the legacy copy with a clean tree; project files (R∖T) are not rewritten;settings.local.jsonis neither read nor copied; the only git write isgit mvof the state files per U2. - Do not run
/do-next, do not commit, do not push, do not change.env*,.git/config. - Outside the target repo the skill writes only by installing the owner's tools (the
Codex CLI,
image_generation, rembg MCP) and only pertools.md: the commands are shown in advance, consent — viaAskUserQuestion, only what was checked is installed;auditinstalls nothing. - Language (
knobs.mdD): talk to the owner in the language they write in — tables, questions, options, reports; never ask about it. The generated files follow the repository (PIPELINE_LANG, defaulten); ask only when the signals disagree or are absent. Contract strings and scripts are English in every language. - Long material goes in
reference/; this file is the procedure and the checklist.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

