Review And Improve

Review and Improve

Coding-agent workflow — a /slash-command your coding agent (Claude Code, Codex, …) runs while developing this repo. Not a runtime skill the deployed @context agent runs; those live in skills/.

You are sweeping the whole repo for public-consumption readiness — docs accuracy, the context agent reachable end to end, 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, a provider missing from the architecture diagram) and you fix it in place. The rest is a punch list you surface to the user.

This is a recurring sweep — meant to be re-run regularly. On a clean repo it ends with "no diffs"; on a dirty one it brings everything back to coherent.

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 missing a wired provider or runtime skill (or listing one that's gone).
• An agent file on disk not imported in app/main.py (rare — single-agent product, but the AGENTS.md recipe allows more; add the import + append to agents=[...]).
• Missing quick_prompts block for a registered agent (draft three from the agent's INSTRUCTIONS; flag the new entries so the user can refine).
• Missing or wrong cross-links between skill files in .agents/skills/.
• Single-line factual claim in one doc contradicted by another doc or by code (e.g. one doc says "hot-reload picks up new agents" while another says a restart is required) — 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).
• Dependency edits in pyproject.toml.
• Anything in db/, compose.yaml, or Dockerfile.
• 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.)
• If multiple worktrees of this repo exist on disk, only one container can bind to localhost:8000 — Step 4 will reflect whichever repo last brought the container up, not necessarily this worktree's app/main.py. Step 4 has a cross-check for this.
• 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:

• Top-level docs: README.md, AGENTS.md, example.env, and the workflow skills in .agents/skills/.
• Code: app/, agents/, db/, evals/, scripts/.
• Configs: compose.yaml, Dockerfile, pyproject.toml, railway.json.

Skip: .venv/, *_cache/, .git/, anything generated.

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. These reads are independent — fan them out across subagents (or parallel Read/grep calls) and collect the results rather than walking the repo serially. Build a mental model of:

• Registered agents — what's imported in app/main.py's agents=[...]? (Expected: context.)
• Agent files on disk — what's in agents/?
• Wired providers — what's in create_context_providers() in agents/sources.py, and which are env-gated?
• Runtime skills — what folders are in skills/, and does each SKILL.md's name match its directory?
• Env vars actually read — grep os.environ, os.getenv, getenv(, plus settings/config modules.
• Quick prompts — what's in app/config.yaml under chat.quick_prompts?
• Eval cases — what's in evals/cases.py?
• 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. The diffing is read-only, so parallelize it freely — but apply the fixes serially so two edits never race on the same file.

4. Live container smoke

First, confirm the live container is serving this repo's agents — not a stale clone or a different worktree. Compare the API's registered agents against what you parsed from app/main.py:

curl -s http://localhost:8000/agents | jq -r '.[].id' | sort

If the list doesn't match the slugs in agents=[...], flag it — Step 4 will be testing the wrong code. Common causes: the container is bound to a different repo path, or docker compose restart is needed. Stop and surface to the user.

For each agent registered in app/main.py, hit it with one of its quick_prompts. Identity decides the surface on context: compose treats owner@example.com as the owner, so owner-path probes must send that id — any other user_id (including the anon sentinel) hits the capture-only guest surface instead. (An OWNER_ID set in .env replaces the compose default entirely — check .env first and probe with that id instead.) Probe both: the owner path with a quick prompt, and the boundary with any other id (expect a polite capture-only reply, no data readback). The probes don't share state, so fire them concurrently (background the cURLs) and collect the results:

curl -sS -X POST http://localhost:8000/agents/<slug>/runs \
  -F "message=<one of the quick_prompts for this slug>" \
  -F "user_id=owner@example.com" \
  -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

Pass = HTTP 200, non-empty content, no errors in the container logs:

docker logs context-api --since 30s 2>&1 | grep -E "Running: \w+\(" | head -40

(Running: <tool>( is the tool-call line shape agno emits when AGNO_DEBUG=True, which compose sets for dev. Without AGNO_DEBUG expect no matches — HTTP 200 and a non-empty body are then your only signal.)

Quality issues (response is plausible but wrong, missing citations, wrong tool fired) are out of scope — note them and recommend the improve-agent skill (autonomous) or the extend-agent skill (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 reports. If validate fails, surface the errors verbatim — type and lint errors usually point at real bugs introduced since the last sweep, not noise to suppress.

6. Evals (ask before running)

python -m evals hits OpenAI — costs money and takes a few minutes. Ask before running:

Run python -m evals to confirm no agent regressed? (Hits OpenAI; takes 1-3 minutes.)

If yes, run python -m evals. If any case fails, add it to "Needs your call" with the eval-and-improve skill as the recommended follow-up. If the user declines, skip this step entirely — it does not affect the rest of the report.

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.
• Recommended follow-up — usually the improve-agent skill (if a live agent looked off) or the eval-and-improve skill (if evals failed).

A clean sweep takes 3-5 minutes (10+ if the venv needs to be created). A dirty one is 15-30, mostly because live smoke surfaces agent regressions you have to triage.