Subagent + Worktree Parallel Development

Parallelize the doing; serialize the merging. Worktree isolation removes write conflicts during implementation but only defers integration conflicts to shared "hotspot" files. Whether the model pays off depends almost entirely on how much your slices touch shared code — so fan out independent work, but treat merging as an ordered, serial, verified step.

See REFERENCE.md for the full recipes (dispatch prompt template, merge configurations, the silent-merge-hazard catalog, the pre-launch checklist, and the cost/benefit table).

When to use

• Fan out across genuinely independent modules/subsystems whose code is disjoint.
• Serialize instead when slices are tightly coupled — two slices that both extend the same component, or both append to the same shared file, will collide at merge. Coupled work is sequenced, not parallelized.

Choose the operating mode and permissions

• Planning-only: produce the dependency map, slice briefs, wave plan, merge order, validation plan, and permission requests. Do not create worktrees or branches, edit files, commit, push, open change requests, merge, or write to a remote tracker.
• Execution: perform only the local mutations that the user authorized. Worktree creation, file edits, and commits do not imply permission to push, open or update a change request, merge, reply to reviews, or change remote tracker state. Record remote-write permissions separately and stop at a local handoff when they are absent.

Core principles (non-negotiable)

• Group-external parallel, group-internal serial. Only independent modules run in parallel. Within one module, land the foundation slice first, then fan out its dependent follow-up slices. A foundation slice establishes the smallest shared contract or scaffold that its follow-ups require; parallel follow-ups before it lands duplicate that foundation and collide.
• Small batches (≤ ~5), merge before the next wave. Each rebase then lands on a stable base; large waves create a merge treadmill (every merge re-conflicts the rest) and raise the odds a subagent is truncated at a run limit.
• The real integration boundary is part of Definition of Done. A fast tier that stubs that boundary can pass on a broken merge. Run every locally reproducible gate that covers the change, including the real integration/e2e/compile/parse path and relevant non-test checks. Classify gates that require protected infrastructure, secrets, policy evaluation, or unavailable hardware as CI-only. If a gate cannot run, record why, the substitute evidence, and the remaining risk; never report it as passed. (REFERENCE §3, §6)
• The orchestrator (the "lead") independently re-verifies before merging. Subagent implements and produces the authorized local artifact in its worktree; the lead re-runs locally reproducible gates and spot-checks the diff. "Done but no artifact" = needs takeover.
• Remote writes need explicit authority. Pushes, change-request creation or edits, merges, review replies, and tracker updates are separate permissions. When authorized, the orchestrator owns the integration record and its completion semantics. Mark a work item complete only when the slice satisfies all of it; a foundation slice, dependent follow-up slice, or other partial slice remains non-closing.
• Stacked change requests are live dependencies until merged. When a base change receives review fixes or is squash-merged, every dependent change must be re-evaluated: rebase or retarget it, update stale descriptions and completion semantics, then rerun the applicable gates on the new head. Old validation on the stacked base is not merge evidence.
• Public docs and agent-facing docs are integration surfaces. If a slice changes a user/agent-visible behavior or public shape, verify the whole surface chain for that slice and its dependents — CLI/help, schemas, user docs, generated or translated docs, and bundled skills/agent guidance. Code tests alone can miss a documentation contract gap or a stale generated marker.
• Disjointness is a merge-cost heuristic, not an architecture goal. Never let "keep slices disjoint / avoid the append hotspot" suppress a sound design decision — a legitimate shared-module edit, deep-module reuse, or a single source for a public shape. When they conflict, serialize that slice's merge rather than degrade the architecture; and make "did this slice reuse the existing deep module (not re-implement it in isolation)?" part of each subagent's DoD — wired into the dispatch-prompt DoD and the pre-launch checklist, not just prose. (REFERENCE §1, §3, §9)

Workflow

choose mode/permissions → decompose + dependency analysis → plan waves
  planning-only → return the plan
  execution → fan out → verify locally → publish/merge only when separately authorized

1. Choose mode and permissions. Record planning-only or execution, the allowed local mutations, and each allowed remote write. Do not infer execution from a planning request or remote authority from local implementation authority. (REFERENCE §3)
2. Decompose + analyze dependencies. Split the work into end-to-end slices. Mark which are independent (parallel-safe) vs. coupled (must serialize). Up front, identify the append hotspots — central registries, enums, dispatch tables, render/plugin maps, shared test files that every slice edits — that is where merge cost concentrates — and give each hotspot exactly one owner slice for the wave; the others flag needed changes instead of editing. (REFERENCE §1)
3. Plan waves. Group independent slices into waves of ≤ ~5; sequence each foundation slice before its dependent follow-up slices. Decide the integration order now. In planning-only mode, return the plan here. (REFERENCE §2)
4. Fan out to implement. In execution mode, launch one subagent per slice in its own git worktree. Its dispatch prompt pins the worktree, states local and remote permissions, requires early durable local artifacts when authorized, and includes the applicable validation gates. (REFERENCE §3)
5. Verify, hand off, or publish. As each implementer finishes, run locally reproducible gates, serialize shared-global-resource tests, audit affected public surfaces, and record every CI-only or unavailable gate with substitute evidence and remaining risk. If remote writes are not authorized, return the local artifact and stop. If they are authorized, the orchestrator performs only the listed remote actions. (REFERENCE §6, §7)
6. Merge serially in dependency order when authorized. Foundation slice first → rebase followers onto the new base → independent groups can merge in any order; a clean rebase still gets the applicable integration gate. Re-poll mergeability after each merge. For stacked changes, retarget followers after the base lands and update stale descriptions, completion semantics, or verification text before merging. Watch for the two marker-free conflict traps. (REFERENCE §4, §5)

This path is not one-shot: independent review sends merged-ready slices back, and remediation reshapes the plan. A review/fix round is a re-dispatch — resume the original implementer with its context where possible, restate the full dispatch discipline (worktree pinning, permissions, and the applicable local and CI-only gate inventory). Follow the repository's commit-history policy. Require one commit per finding only when the repository or user explicitly requires it; otherwise map each finding to its resolution in the handoff. The lead then re-verifies and closes the loop on the review channel only when that remote write is authorized — e.g. a reply mapping each finding → resolution — keeping the change description current where the host supports it. Re-derive the overlap map and merge order as fixes land, verify each slice against its source requirement (not just its own green tests), and fix a finding at the altitude of its true cause, not where it surfaced. (REFERENCE §1, §6, §7)

When append hotspots keep dominating merge cost, the durable fix is architectural — split them into per-module fragments that auto-aggregate (REFERENCE §8). Run the pre-launch checklist (REFERENCE §9) before every wave.