The team¶
Trazo gives Claude Code and Codex the same roles, with a generic AGENTS.md adapter for other compatible agents. The team works through GitHub, not chat memory.
Roles and agents¶
| Role / Agent | Command | Claude Code | OpenAI Codex | Cline | Use for | Edits files? | Updates GitHub? |
|---|---|---|---|---|---|---|---|
| Engineer | /eng, /work |
Haiku default; /work runs in forked Haiku context |
Luna | User chooses provider and model | Building: code, tests, git, pull requests | Yes | Yes |
| Senior engineer | /senior-eng |
Sonnet for the command turn | Sol | User chooses provider and model | Escalated issues labeled tier:senior (ADR 0016, 0018); same building role, not a reviewer of ordinary PRs |
Yes | Yes |
| PM / advisor | /pm |
Sonnet for command turn; Sonnet pm subagent |
Sol | User chooses provider and model | Status, planning, priorities, "is this result real?", charter/budget/stop rule, writing issues | No | Issues only — labels, parent/sub-issue, blocked-by, milestone, assignee, type, plus comments. Never rewrites an issue --body, never closes one |
| Reviewer | subagent only | Sonnet | Sol | User chooses provider and model | Reviewing pull requests and diffs before merge | No | Verdict, as a comment on the pull request |
| Security | subagent only | Sonnet | Sol | User chooses provider and model | Secrets, permissions, workflows, dependencies, infra, repo security settings | No | Findings as a PR comment, or a new issue |
| Skeptic | subagent only | Sonnet | Sol | User chooses provider and model | Breaking a research/analysis result before it is acted on | No | Verdict, as a comment on the pull request or issue |
| Architect | /architect (on demand) |
Opus for the command turn; advisory in the session | Astra | User chooses provider and model | System structure and design fit, on trigger only (ADR 0021) | No | Finding on the issue or pull request |
Which model runs? A fresh Claude Code host install defaults to Haiku. /work runs in the forked Haiku context; /eng runs its command turn on Haiku. Claude PM, reviewer, security, and skeptic subagents are pinned to Sonnet, and /pm runs its command turn on Sonnet. A Claude role command's model applies for that turn only, then the session returns to its previous model. Codex role agents are pinned to Luna for engineering and Sol for the other roles. Cline has no Trazo model setting: choose the provider and model in Cline. Existing Claude or Codex settings can override defaults.
No role other than the engineer edits a file — that is the invariant, and it is why a review cannot quietly grade the same conversation's work. But "no file edits" is not "read-only": every one of these four writes to GitHub. The PM reshapes the issue graph, and the reviewer, security and skeptic agents all record their verdict where the next session can see it, because a verdict that lives only in the conversation gates nothing.
Role-authored issue descriptions and GitHub comments end with a role attribution footer under the shared rules. Claude Code and Codex use their native adapters; other tools, including Cline, use the AGENTS.md adapter. The footer includes the exact active model and agent ID (or session ID when no agent ID is exposed) only when the role context provides them. Unknown fields are omitted. GitHub still shows the shared account as author, so the footer is self-reported attribution, not an independently verified identity.
Review verdicts are comments, not approvals
Reviewer, security and skeptic post gh pr review --comment, never --approve or --request-changes. GitHub rejects both while the agent and the PR author are the same account — which is every pull request here, because agents share the owner's identity (#44). The moment agents get their own identity, these become real blocking reviews. Until then, a green CI run is your merge gate, not a reviewer approval.
needs-decision is the owner's gate. The PM removes the label only to record your decision, in the same step as a comment that opens Owner decision (<date>):, quotes your words verbatim, and says where you gave them: in the session, typed by you, or the URL of a comment you wrote on the issue or PR. A quote relayed by another agent never counts, even if it claims to be verbatim. The PM never clears the label on its own judgment, and never applies the decision itself; engineers build from the issue.
They can be used in two ways:
1. Command turns: /pm runs its command turn on Sonnet, and /eng runs its command turn on Haiku. The host session then returns to its previously selected model. Reviewer, security, and skeptic are subagent-only — never role-switch commands — so a review can't grade the same conversation's own work (see .trazo/project/adr/0003-review-security-github-tracked.md).
2. Subagent delegation: /work runs in a forked Haiku context; PM, reviewer, security, and skeptic agents are pinned to Sonnet. Claude Code delegates to reviewer and security automatically as part of /work and /check-pr, or ad hoc, without switching the whole conversation.
Commands¶
You don't need to memorize these — CLAUDE.md maps plain-English requests to the right one, and /trazo shows a menu. Typing / in Claude Code lists all of them.
| Command | Does |
|---|---|
/trazo |
Menu of what you can do right now |
/start / /wrapup |
Begin / end a work session |
/work 12 |
Implement issue #12 in the forked Haiku context → pull request (reviewer checks it first) |
/check-pr 15 |
Review pull request #15 (reviewer, plus security if needed) |
/pm |
Run the PM command turn on Sonnet; use the pinned pm subagent for subsequent PM work |
/eng |
Run the engineer command turn on Haiku; later turns use the session's selected model |
/senior-eng |
Run the senior engineer command turn on Sonnet, for issues labeled tier:senior only |
| ask the security subagent | Secrets, permissions, infra (no role switch) |
| ask the reviewer subagent | PRs, diffs, safety (no role switch) |
| ask the skeptic subagent | Break a research/analysis result before it counts (no role switch) |
/brief |
Quick status, changes nothing |
/architect |
Run the Architect command turn on Opus; advisory in the session |
/decide … |
Draft a decision record |
/kickoff |
New-project setup |
| note a lesson for Trazo | record it in the project; promote by hand when working in Trazo |
Plain English → routine¶
The owner shouldn't need to remember commands. CLAUDE.md maps requests like these to a routine or role:
| If you say something like… | It does |
|---|---|
| "catch me up", "where are we", "what's next" | /start routine (or /pm for strategy questions) |
| "what can I do", "help", "menu" | /trazo |
| "work on issue 12", "fix X" | /work routine |
| "is this PR ok", "review #15", "can I merge" | /check-pr routine (or ask the reviewer subagent directly) |
| "is this secure", "check permissions" | ask the security subagent |
| "is this number real?", "poke holes in this analysis" | ask the skeptic subagent |
| "should we…", "is this worth it", "plan the next milestone" | /pm |
| "back to building", "ready to code" | /eng |
| "we decided…" | /decide routine |
| "wrap up", "done for today" | /wrapup routine |
| "GitHub/CI says …" (settings, failures) | handled directly; ask the security subagent for protection/permission settings |
See the playbook for how this plays out over a normal day, and the guide for setting up the optional GitHub Actions versions of the PM and reviewer.