Canonical flow aligned with global .claude execution policy.
/prompt → F-xx extraction → /research (optional) → /planner → DB approval → /execute {id} (TDD) → Thor per-task validation → Thor per-wave validation → closure (all F-xx verified) → learning loop.
Any bypass of this chain is a policy violation for multi-step work.
| Step | Claude Code | Copilot CLI |
|---|---|---|
| Capture goal | /prompt "<goal>" |
@prompt "<goal>" |
| Research (optional) | /research |
@research |
| Create plan | /planner |
@planner or cplanner "<goal>" |
| Execute tasks | /execute {id} |
@execute {id} |
| Validate | Thor validator | @validate {id} |
| Close | PR + CI + merge, or validated non-code deliverable | Same |
- Extract F-xx requirements from user intent.
- Persist prompt artifact in
.copilot-tracking/.
- Generate waves and tasks mapped to F-xx.
- Require explicit review/approval before execution.
Per task: RED → GREEN → REFACTOR.
Required checks before submission:
- Relevant tests pass
- Type/lint checks pass where applicable
- Artifacts are present and scoped
Thor validates task submissions independently and blocks progression on failed gates.
- Validate wave completion
- Ensure CI green before merge
- Complete plan only when all F-xx are verifiably satisfied
Capture reusable lessons after closure:
- Generic rules in
.claude/rules/(bounded, justified) - Project-specific conventions in repo docs
- night-agent (
night-maintenance): off-hours hygiene, triage, bounded remediation. - sync-agent (
claude-syncflow): environment alignment and sync continuity.
These operational flows support the main chain; they do not replace planner/execute/validate for feature work.
- Max 250 lines per file
- No self-declared completion without evidence
- Batch CI fix policy
- Zero deferred technical debt at task completion