Dev Cycle

Coordinator (user session or burndown per issue): dispatch and sequence roles; never write code, post plans, transition board status, or merge.

Project facts from AGENTS.md: § Project Board, § Branch Map, § Build & Validation, § Repositories. Missing section: name it and stop.

Use issue-plan § Board Status Transitions: explicit board: none skips board operations, not ledger ownership; missing board facts stop. Deliver fully resolving PRs only, never partial Refs.

Inputs

GitHub issue number/URL anchors ledger, claim, and Closes #n. No issue: run issue-raise first. Batch coordinator may supply owned file/module scope and validation requirements.

Subagent Dispatch

All subagent orchestration defaults to foreground execution: wait for dispatched work to finish and collect its result before continuing dependent work. Use background mode only when the user explicitly requests it; pass that choice through nested coordinators within the requested scope. Set the tool's foreground option explicitly where available (for example, background: false or run_in_background: false).

Parallel work does not imply background permission. Use concurrent foreground calls when supported; otherwise dispatch sequentially. If the harness only offers asynchronous dispatch, wait for completion before continuing and state that limitation rather than claiming native foreground execution.

Phases

• Fresh context per phase: issue number/URL and needed facts only, no inherited conversation.
• Never skip planning as "trivial"; the architect establishes scope before implementation/review. The soft gate avoids unnecessary round trips.
• Validate each result against its request before dispatching the next; inconsistencies stop.

Phase Prompts

Diagnose: Run issue-diagnose on issue <number> (<url>). [Re-diagnosis authorized by this cycle; RCA REST id: <...>; trigger: <blocker or review finding>.] Return RCA link/REST id, conclusion, criteria, refinement or evidence needed.
Refine: Run issue-refine on issue <number> (<url>). [Apply the RCA's requested change: <...>.] [Approved: <user go-ahead>.] Return proposal and questions, or the applied refinement and verified status.
Plan: Run issue-plan on issue <number> (<url>). Return ledger link, approach, task count, blockers/decisions.
Implement: Run issue-implement initial mode on issue <number> (<url>). Claim/resume evidence; ledger REST id/link: <...>. [Owned scope: <files/modules>.] Validation: <§ Build & Validation>. Return PR URL, paths, ledger state, validation output, risks.
Review: Run code-review PR mode on PR <number>, round <n>. Merge iff criteria hold; perform after-merge steps. Return assessment, comment URL, merge/blocker, board transition, cleanup.
Revise: Run issue-implement revision mode on PR <number> (issue <n>, <url>). Current-cycle ownership; ledger REST id/link: <...>. Findings from <review comment URL>: <enumerated>. Validation: <...>. Return commits, fixed/contested findings with reasoning, validation output.

Flow

1. Load issue (gh issue view <n> --repo <repo-slug> --json number,title,body,state,url), complete board lookup, and all REST ledger comments per issue-plan. Closed/done: finished, stop. Existing ledger or in-progress/in-review: claimed, stop unless explicitly user-authorized resume, even boardless; batch dispatch is not authorization. Backlog/unrefined: issue-refine; it reports needs diagnosis: Diagnose, then refine again. Authorized resume with delivery PR goes to review, not initial implementation: first verify unique ledger, open PR identity, exact intended closing reference, and in-review or board: none. In-review without matching PR blocks.
2. Plan, unless explicit resume. claimed: report/stop. Decisions instead of ledger: return to user/batch coordinator and stop, never answer for the architect. needs diagnosis: Diagnose, then plan again. RCA needs refinement: Refine with the requested change, holding the user's approval, then plan again; once per RCA, a second request stops and reports both. Confirm unique ledger and in-progress or board: none; retain claim/resume evidence and ledger REST id/link for handoffs.
3. Implement. Require fully resolving PR URL, exact closing-reference verification per issue-implement, validation output, and verified in-review or board: none. Blocked: return ledger once to architect in issue-plan same-cycle blocked-plan repair mode, with ownership evidence/blockers. Architect returns needs diagnosis: Diagnose with this cycle's authorization, then finish the same repair. Continue only after ledger blockers clear; decisions or repeated blocker stop, never generic auto-resume.
4. Review, round 1. Merged: finish. Needs diagnosis (takes precedence): Diagnose with this cycle's authorization, run the blocked-plan repair with the review finding, then Revise with the ledger's changed tasks as findings; this spends the cycle's single repair and one revision round. Repair already used, a second needs diagnosis, or unchanged: park with both positions. Critical/High: revise. Otherwise any unmet merge criterion, including important Medium findings: report the blocker and leave the PR open for the user.
5. Revise, at most twice (review rounds 2/3). Developer fixes enumerated findings only; fresh architect re-reviews against prior comment. Contest ends loop immediately for human decision. Cap/contest: park PR with round, exit reason, both positions.
6. Report: issue/PR URLs, ledger link/approach, merge commit or parked round/reason or unmet criterion, verified board status, validation, residual risks, incomplete work.

Without Subagents

One fresh session per phase, loading only that phase's skill and prompt. Never review in the session that implemented.