Orchestration digest — Tier B (single-agent clients)

Who this is for: OpenAI Codex CLI, Google Gemini CLI, Cursor, and VS Code Copilot — any client that drives a single agent and cannot fan out to parallel subagents or run a hook system. If you are Claude Code, read the Tier A digest instead.

This is the compact, session-start digest. The full reference is get_guidance("orchestration") and the agents standard.

How you work

You follow the same workstream logic as Tier A, but sequentially, driven by AGENTS.md plus MCP calls. You do not have a .claude/ native layer, and you must not invent equivalents — there are no Codex agent files, no hooks.json, no fabricated subagents. Your "orchestration" is the ordered sequence of steps below, performed by you, one at a time.

For a workstream (build / review / refactor / test / document / investigate / operate), run the steps in order yourself:

1. Plan — restate the task, list the files you will touch, and the checks that will gate it.
2. Do the work — make the change.
3. Review it yourself — re-read the diff against the applicable standards (you are your own reviewer; there is no reviewer subagent to spawn).
4. Validate deterministically — see the MCP tools below.
5. Human approve — stop and get sign-off before anything irreversible (operate/refactor with API changes always gate).

Lean on the MCP for the checks you can't fan out

Because you cannot spawn a validator or reviewer subagent, use the MCP's deterministic tools directly:

• validate(path, profile) — run the repo's profile checks and read the results yourself.
• check_drift(repo=… | path=…) — canonical-layout conformance; fix each DRIFT finding using its remediation hint.
• which_standards_apply(repo) + get_standard(domain, scope) — the exact rules to self-review against.
• get_guidance("orchestration") — the full flow when this digest is not enough.

Rules that still bind you

• No metered API misuse / no fabricated tooling. Use your client's own harness and model picker. Do not invent Tier-A-only machinery.
• Secrets: never commit tokens, keys, passwords, or connection strings — reference by Key Vault name only.
• Commits: type(scope): short description with an AB#<id> work-item reference.
• Security-sensitive or infra changes: require explicit human approval before executing.

Session protocol

Read .ai/state/ then .ai/memory/ at session start; summarise believed state before changing anything; update .ai/state/HANDOFF.md before ending. This is how your work hands off to (and from) other tools. Full contract: ai-workspace standard.