Repo-wide drift sweep for public-readiness — diff docs against code, confirm every agent is registered and reachable, every env var documented, every doc path exists, and scripts behave as advertised; auto-fix mechanical drift and flag the rest. Use before a public release or after a refactor.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Repo-wide drift sweep for public-readiness — diff docs against code, confirm every agent is registered and reachable, every env var documented, every doc path exists, and scripts behave as advertised; auto-fix mechanical drift and flag the rest. Use before a public release or after a refactor.
Review and Improve
Coding-agent workflow — a /slash-command your coding agent (Claude Code, Codex, others) runs while developing this repo. Invoke it by name (e.g. /review-and-improve) or describe the task and it triggers automatically.
You are sweeping the whole repo for public-consumption readiness — docs accuracy, every agent reachable, scripts that actually do what the docs claim, no stale env vars, format + validate clean. Most drift is mechanical (renamed file, missing entry in example.env, new agent not in the architecture diagram) and you fix it in place. The rest is a punch list you surface to the user.
AGENTS.md is the source of truth for repo conventions; CLAUDE.md is a symlink to it — edit once, both update.
What you auto-fix vs. what you flag
Auto-fix in place (no asking):
Stale file paths in any doc.
Missing entries in example.env for env vars the code actually reads.
Stale entries in example.env for vars nothing reads — delete unless the surrounding comment block describes them as optional/future ("alternate model providers", "future feature"). Flag instead of fixing if intent is unclear.
Architecture diagram in AGENTS.md / README.md missing a registered agent or workflow.
New agent file on disk not yet imported in (add the import + append to ).
Missing manifest entry in app/config.yaml for a component registered in app/main.py (draft a one-line description and three quick prompts from its INSTRUCTIONS; flag the new entries so the user can refine). Only code components — a Studio-built one carries its own description.
Missing or wrong cross-links between docs and the coding-agent skills in .agents/skills/*/SKILL.md (and between skills).
Single-line factual claim in one doc contradicted by another doc or by code — auto-fix the doc, not the code.
Flag, don't fix (surface for the user):
Section-level doc rewrites (a premise is now wrong).
Code changes beyond imports (instructions, tools, model swaps).
Failing eval cases or failing live agents — recommend the right follow-up prompt; don't fix here.
0. Preconditions
Live container reachable: curl -sSf http://localhost:8000/health returns 200. If not, ask the user to docker compose up -d --build first — Step 4 needs a live container. (docker compose ps is unreliable from worktrees or alternate clones — trust the health probe.)
Recommend a feature branch so auto-fixes are easy to revert: git checkout -b review/$(date +%Y%m%d).
1. Scope check
Restate the surface area in 4-5 lines so the user can redirect before you read everything:
If the user has a specific concern (recent refactor, prepping a public release, a doc they think is stale), fold it in now.
2. Inventory
Read every file in scope. Build a mental model of:
Registered agents + team — what's imported in app/main.py's agents=[...] and teams=[...]? This is the whole population this sweep is responsible for. Components the Studio built at runtime have no source file, no manifest entry, and no place in a docs diff.
Agent and team files on disk — what's in agents/ and teams/?
Registered workflows — what's imported into app/main.py and passed to AgentOS(workflows=[...])? Workflow files on disk in workflows/?
Schedules — what register_schedules() (app/schedules.py) registers: the env gate where one exists (ENABLE_DEPLOY_CHECK), and run-evals' ships-disabled posture (its enabled bit is user-owned via the AgentOS UI after first creation).
Scripts — for each file in scripts/, what does it actually do? (Headers and the first few lines are usually enough.)
Don't write anything yet — read first, fix once.
3. Consistency pass
The bulk of the work. Diff each pair below; auto-fix per the rules at the top.
Agent renamed/removed; the MCP smoke check breaks or tests the wrong agent
4. Live container smoke
First, confirm the live container is serving this repo's components — not a stale clone or a different worktree. The listing endpoints return two populations: code components registered in app/main.py (is_component=false) and components Platform Builder built at runtime (is_component=true). Only the first is this repo's, so compare on that filter:
If that list doesn't match the slugs in agents=[...] / teams=[...], stop and surface it to the user — the rest of Step 4 would be testing the wrong code. Common causes: the container was brought up from another worktree or clone of this repo (only one can bind localhost:8000), or docker compose restart is needed.
Then smoke each code component with one of its quick_prompts from app/config.yaml — every agent in agents=[...]and the team in teams=[...]. Agents use /agents/<slug>/runs; the team uses /teams/agno/runs with the same flags:
curl -sS -X POST http://localhost:8000/agents/<slug>/runs \
-F "message=<one of the quick_prompts for this slug>" \
-F "user_id=claude-review" \
-F "stream=false" \
-o /tmp/review-<slug>.json \
-w "HTTP %{http_code} in %{time_total}s\n"
jq -r '.content // .' < /tmp/review-<slug>.json | head -20
Workflows stay out of the live smoke: deployment-check already runs on its daily cron, and run-evals spends real model budget. Step 3 — registered, schedules pointing at real ids — is their coverage.
Two components need more than a manifest prompt:
platform-builder — use its first quick prompt only. "What can you build for me?" is plan-only. The other two are "Build …" asks, and the builder passes publish=true to finish a build, so smoking with one would create and publish a real, dispatchable component and leave it live. If a create fires anyway, delete_new_builder_state(pre) removes it — so take pre = snapshot_builder_state() before the builder's call (both from evals/hooks.py; they sweep new components, schedules, learning rows, and notes). Nest that inside the learning bracket below rather than instead of it — the outer bracket is uncapped, whereas the builder helper refuses past 25 rows.
agno — capture is ungated. Its quick prompts are read-only questions, but the model decides for itself what to file on any run: per-user profile and memory on all four components, plus notes and entities — shared by everyone on the platform — on the team. Bracket the whole smoke step, once around all the components rather than per prompt, with the snapshot_learning_state / delete_new_learning_state pair from evals/hooks.py — the full snippet is in improve-agent Step 2.
The diff works on row identity: it removes what the sweep created, cannot undo an edit inside a row that already existed, and sweeps a note a teammate files while you run. On a platform people are actively using, run Step 4 in a window you own, or tell the user you are skipping the delete and leaving whatever the smoke filed in place.
Pass = HTTP 200, non-empty content, no errors in the container logs:
(Running: <tool>( is agno's tool-call log line under AGNO_DEBUG=True, which compose sets for dev. Without it expect no matches — HTTP 200 and a non-empty body are then your only signal.)
Then smoke the MCP interface (mcp_server=True in app/main.py):
./scripts/mcp_check.sh
Pass = MCP OK — <n> tools plus a non-empty agent response. It runs inside the container and calls platform-manager (read-only), so it is safe to run repeatedly.
Quality issues (response is plausible but wrong, missing citations, wrong tool fired) are out of scope — note them and recommend improve-agent (autonomous) or extend-agent (user-driven) depending on whether the user has a specific fix in mind.
5. Format + validate
source .venv/bin/activate # if not already active
./scripts/format.sh
./scripts/validate.sh
format.sh auto-fixes. validate.sh gates on formatting as well as lint and types, so keep that order — a sweep that skips format.sh fails validate on its own edits. If validate fails, surface the errors verbatim; don't suppress them.
6. Evals (ask before running)
The release tag hits OpenAI, so ask first:
Run python -m evals --tag release to confirm no agent regressed? (Hits OpenAI; takes 1-3 minutes.)
If any case fails, add it to "Needs your call" with eval-and-improve as the recommended follow-up. If the user declines, skip this step.
7. Report
If nothing was fixed and nothing flagged, skip the first two blocks below and just print: "Repo is consistent and the live container is healthy. No follow-up needed." No commit suggested.
Otherwise, wrap up with three blocks, in order:
Fixed automatically — one line per change, with file path. Terse.
Needs your call — flagged items, ranked by severity. For each: one-line description, file (and line if useful), recommended action.
Diff + next step:
git diff --stat
Suggested commit message — chore: review-and-improve sweep plus one short bullet per fix bucket.
