AGENTS.md

1# ECC for Codex CLI
2 
3This supplements the root `AGENTS.md` with Codex-specific guidance.
4 
5For repo navigation, surface ownership, and PR diff packet guidance, read
6`docs/CODEX-NAVIGATION-GUIDE.md` after this supplement.
7 
8## Model Recommendations
9 
10| Task Type | Recommended Model |
11|-----------|------------------|
12| Routine coding, tests, formatting | GPT 5.5 |
13| Complex features, architecture | GPT 5.5 |
14| Debugging, refactoring | GPT 5.5 |
15| Security review | GPT 5.5 |
16 
17## Skills Discovery
18 
19Skills are auto-loaded from `.agents/skills/`. Each skill contains:
20- `SKILL.md` — Detailed instructions and workflow
21- `agents/openai.yaml` — Codex interface metadata
22 
23Available skills:
24- tdd-workflow — Test-driven development with 80%+ coverage
25- security-review — Comprehensive security checklist
26- coding-standards — Universal coding standards
27- frontend-patterns — React/Next.js patterns
28- frontend-slides — Viewport-safe HTML presentations and PPTX-to-web conversion
29- article-writing — Long-form writing from notes and voice references
30- content-engine — Platform-native social content and repurposing
31- market-research — Source-attributed market and competitor research
32- investor-materials — Decks, memos, models, and one-pagers
33- investor-outreach — Personalized investor outreach and follow-ups
34- backend-patterns — API design, database, caching
35- e2e-testing — Playwright E2E tests
36- eval-harness — Eval-driven development
37- strategic-compact — Context management
38- api-design — REST API design patterns
39- verification-loop — Build, test, lint, typecheck, security
40- deep-research — Multi-source research with firecrawl and exa MCPs
41- exa-search — Neural search via Exa MCP for web, code, and companies
42- claude-api — Anthropic Claude API patterns and SDKs
43- x-api — X/Twitter API integration for posting, threads, and analytics
44- crosspost — Multi-platform content distribution
45- fal-ai-media — AI image/video/audio generation via fal.ai
46- dmux-workflows — Multi-agent orchestration with dmux
47 
48## MCP Servers
49 
50Treat the project-local `.codex/config.toml` as the default Codex baseline for ECC. The current ECC baseline enables GitHub, Context7, Exa, Memory, Playwright, and Sequential Thinking; add heavier extras in `~/.codex/config.toml` only when a task actually needs them.
51 
52ECC's canonical Codex section name is `[mcp_servers.context7]`. The launcher package remains `@upstash/context7-mcp`; only the TOML section name is normalized for consistency with `codex mcp list` and the reference config.
53 
54### Automatic config.toml merging
55 
56The sync script (`scripts/sync-ecc-to-codex.sh`) uses a Node-based TOML parser to safely merge ECC MCP servers into `~/.codex/config.toml`:
57 
58- **Add-only by default** — missing ECC servers are appended; existing servers are never modified or removed.
59- **7 managed servers** — Supabase, Playwright, Context7, Exa, GitHub, Memory, Sequential Thinking.
60- **Canonical naming** — ECC manages Context7 as `[mcp_servers.context7]`; legacy `[mcp_servers.context7-mcp]` entries are treated as aliases during updates.
61- **Package-manager aware** — uses the project's configured package manager (npm/pnpm/yarn/bun) instead of hardcoding `pnpm`.
62- **Drift warnings** — if an existing server's config differs from the ECC recommendation, the script logs a warning.
63- **`--update-mcp`** — explicitly replaces all ECC-managed servers with the latest recommended config (safely removes subtables like `[mcp_servers.supabase.env]`).
64- **User config is always preserved** — custom servers, args, env vars, and credentials outside ECC-managed sections are never touched.
65 
66## External Action Boundaries
67 
68Treat networked tools as read-only by default. Search, inspect, and draft freely within the user's requested scope, but require explicit user approval before posting, publishing, pushing, merging, opening paid jobs, dispatching remote agents, changing third-party resources, or modifying credentials.
69 
70When approval is ambiguous, produce a local plan or draft artifact instead of taking the external action. Preserve user config and private state unless the user specifically asks for a scoped change.
71 
72## Multi-Agent Support
73 
74Codex now supports multi-agent workflows behind the experimental `features.multi_agent` flag.
75 
76- Enable it in `.codex/config.toml` with `[features] multi_agent = true`
77- Define project-local roles under `[agents.<name>]`
78- Point each role at a TOML layer under `.codex/agents/`
79- Use `/agent` inside Codex CLI to inspect and steer child agents
80 
81Sample role configs in this repo:
82- `.codex/agents/explorer.toml` — read-only evidence gathering
83- `.codex/agents/reviewer.toml` — correctness/security review
84- `.codex/agents/docs-researcher.toml` — API and release-note verification
85 
86## Key Differences from Claude Code
87 
88| Feature | Claude Code | Codex CLI |
89|---------|------------|-----------|
90| Hooks | 8+ event types | Not yet supported |
91| Context file | CLAUDE.md + AGENTS.md | AGENTS.md only |
92| Skills | Skills loaded via plugin | `.agents/skills/` directory |
93| Commands | `/slash` commands | Instruction-based |
94| Agents | Subagent Task tool | Multi-agent via `/agent` and `[agents.<name>]` roles |
95| Security | Hook-based enforcement | Instruction + sandbox |
96| MCP | Full support | Supported via `config.toml` and `codex mcp add` |
97 
98## Security Without Hooks
99 
100Since Codex lacks hooks, security enforcement is instruction-based:
1011. Always validate inputs at system boundaries
1022. Never hardcode secrets — use environment variables
1033. Run `npm audit` / `pip audit` before committing
1044. Review `git diff` before every push
1055. Use `sandbox_mode = "workspace-write"` in config
106