agno-agi Review and Improve
Coding-agent workflow: run as
/review-and-improveor by describing the task.
Sweep the repo for docs accuracy, reachability, and consistency. Fix mechanical drift in place; surface the rest. AGENTS.md is the source of truth (CLAUDE.md symlinks to it).
Auto-fix: stale doc paths; example.env entries missing for vars the code reads, or stale ones nothing reads (keep those a comment marks optional/future); a registered agent or workflow missing from an architecture diagram; an agent file on disk not imported in app/main.py; a missing app/config.yaml manifest entry (draft one and flag it); broken links between docs and skills; a single-line claim in one doc contradicted by another doc or by code (fix the doc).
Flag, don't fix: section-level rewrites; code beyond imports; pyproject.toml; anything in db/, compose.yaml, Dockerfile; failing evals or live agents.
0. Preconditions
curl -sSf http://localhost:8000/health returns 200 (Step 4 needs it). Suggest git checkout -b review/$(date +%Y%m%d).
1. Scope
Restate in a few lines: README.md, AGENTS.md, example.env; .agents/skills/*/SKILL.md; app/, agents/, teams/, workflows/, db/, evals/, scripts/; compose.yaml, compose.prod.yaml, Dockerfile, pyproject.toml. Skip .venv/, caches, .git/. Fold in any concern the user names.
2. Inventory
Read first, fix once. Build the picture: registered agents/teams/workflows in app/main.py; files on disk; what app/registry.py declares and imports; env vars read (getenv(, os.environ, plus env_flag/_int_env helpers); the manifest; eval cases; register_schedules() targets; what each script does.
3. Consistency pass
| Check | Where |
|---|---|
| Every agent/team/workflow file is registered | agents/, teams/, workflows/ ↔ app/main.py |
| Every registered component has a manifest entry | app/main.py ↔ app/config.yaml |
| Every env var in code is documented, and every documented var is read | code ↔ AGENTS.md table ↔ example.env (RUNTIME_ENV is deliberately absent from example.env) |
| Every path and script named in docs exists and does what's claimed | README.md, AGENTS.md, skills ↔ filesystem |
| Architecture diagram matches registrations | AGENTS.md |
| Eval cases reference real agents and tools | evals/cases.py |
| Every schedule hits a real workflow id | app/schedules.py ↔ workflows/ |
Registry inventory prose matches app/registry.py |
AGENTS.md |
| Key Files table matches the tree | AGENTS.md |
Skill name: equals its folder; ../../../ links resolve; .claude/skills symlink resolves |
.agents/skills/ |
.mcp.json servers agree with the docs |
.mcp.json |
MCP endpoint claims match mcp= in app/main.py; scripts/mcp_check.sh calls a registered agent |
docs ↔ code |
4. Live smoke
Confirm the container serves this repo's components (code components are is_component=false):
curl -s http://localhost:8000/agents | jq -r '.[] | select(.is_component == false) | .id' | sort
curl -s http://localhost:8000/teams | jq -r '.[] | select(.is_component == false) | .id' | sort
A mismatch with app/main.py means another checkout owns port 8000 — stop and say so.
Smoke every code agent and the team with one quick prompt each (/agents/<slug>/runs, /teams/agno/runs; user_id=claude-review, stream=false). Pass = 200, non-empty content, no errors in docker logs agentos-api --since 30s. Bracket the whole step with the learning snapshot pair, and platform-builder's run with the builder pair (improve-agent Step 2); use only the builder's first quick prompt (the others build real components). Workflows stay out — deployment-check runs on its cron, run-evals spends model budget.
Then ./scripts/mcp_check.sh → MCP OK — <n> tools and a real answer.
Quality issues are out of scope — note them and point at improve-agent.
5. Format + validate
source .venv/bin/activate
./scripts/format.sh && ./scripts/validate.sh
Surface validate errors verbatim.
6. Evals
Ask first (hits OpenAI, 1–3 minutes): python -m evals --tag release. Failures go to "Needs your call" with eval-and-improve as the follow-up.
7. Report
Nothing fixed or flagged → "Repo is consistent and the live container is healthy. No follow-up needed." Otherwise: Fixed automatically (one line per change, with path); Needs your call (ranked, with recommended action); git diff --stat and a suggested commit (chore: review-and-improve sweep + a bullet per bucket).