Instruction file imported from Marsouino/roblox_test (.cursor/rules/agentic-workflow.mdc). Copyright stays with the author.

Agentic Workflow

The project uses a strict agentic TDD workflow.

Entry points (skills)

• /onboard-session -- fresh session briefing
• /plan -- shape a scoped idea into TDD-ready steps (architect-agent; product brainstorming happens upstream with the human)
• /tdd-step <plan> <ID> / /tdd-step <plan> <START>..<END> -- run one Red-Green-Refactor-Commit cycle on a slice, or chain the slices of a range (majordome orchestrates red/green/refactor/commit). scripts/governance/autopilot_scope.py is the single reading of the line; nothing about plan or range is inferred. Every role runs on the engine its charter pins (harness: cursor | claude in .cursor/agents/<role>.md), through scripts/governance/role_runner.py (brief, record, measure) on both engines, with no per-run override (same charter, same status vocabulary; see norme_orchestration "Execution harnesses")
• /commit -- finalize a validated step into a high-signal commit (commit-agent)

Specialized subagents

• architect-agent -- architecture and TDD slicing; writes the plan file docs/backlog/<PREFIX>_*.md (its only write)
• majordome -- lightweight orchestrator of the red-green-refactor-commit sequence
• red-agent -- writes or updates tests, confirms GOOD_RED for the right reason
• green-agent -- smallest honest implementation change to pass targeted tests
• refactor-agent -- improves design without changing behavior
• commit-agent -- synthesizes a high-signal commit message and executive summary

Discipline

• One step per cycle. No silent multi-step enchainment in manual mode.
• Tests are only written or changed by red-agent.
• Production code is only changed by green-agent and refactor-agent.
• The human validates plans before execution. A plan is a committed file in docs/backlog/ (Statut : plan validé humain), written by architect-agent; slice definitions never live only in a chat or an agent's memory, so any session can resume the arc from git.
• The human validates commit content before it is recorded in manual mode. In autopilot mode, the slice spec from the plan IS the pre-validated contract; commit content is recorded without per-cycle human approval.
• Pushing to the remote is always a human action, in both modes.
• Never bypass pre-commit hooks. Never use --no-verify. (Applies in both modes; autopilot does NOT relax this.)
• Stable project facts live in the typed registry (src/registry/ + docs/specs/project_registry.yaml). No agent — main agent included — plans, briefs, or implements a parallel config store for them.
• Autopilot is the only opt-in exception to per-cycle human pause points. Activated explicitly by a range, /tdd-step <plan> <START>..<END>; the range on that line is the scope of the run. See docs/agent/norme_autopilot.md for the full contract.

Doctrine to apply during the cycle

• docs/agent/norme_coding.md -- code constraints (fail-fast, no silent fallback, explicit imports, Pydantic BaseSettings, timeouts, no legacy naming)
• docs/agent/norme_testing.md -- test discipline (behavior over implementation, specific assertions, error path coverage, no hollow tests)
• docs/agent/norme_commit.md -- commit format (type(scope) summary, Context / Decided / Remains)
• docs/agent/norme_autopilot.md -- autopilot contract (hard stops, soft stops, override, push behavior); applied only when /tdd-step carries a range <START>..<END>

Each execution agent loads the relevant document(s) before acting.

Inter-agent communication

• No persistent handoff files are required between phases. The plan file is not a handoff file: it is the arc's contract, written once by architect-agent, validated by the human, committed, and read by every phase of every slice.
• Subagents return structured textual responses with an explicit status field.
• The orchestrating agent reads that status and routes the next prompt.
• Durable state lives in git: diff, tests, and commit messages.

Expected status vocabulary

• architect-agent: DISCUSS_MORE, HANDOFF_TO_MAJORDOME
• red-agent: GOOD_RED, BAD_RED, BLOCKED (TESTS_REPAIRED in repair mode: a test a gate demanded, the pre-gate flagged or GREEN named as wrong exists and passes)
• green-agent: GREEN_CONFIRMED, STILL_FAILING, BLOCKED
• refactor-agent: REFACTOR_DONE, REFACTOR_PARTIAL, BLOCKED
• commit-agent: READY_TO_COMMIT, NEEDS_MORE_EVIDENCE, BLOCKED If a subagent returns a status outside this vocabulary, treat it as a protocol violation and require correction before continuing.