Parallel Implementation Orchestrator
Use this skill only after resolving the task with ../setup/references/task-context.md. The
orchestrator owns shared task/review records and the final combined verification.
Decide
Parallelize only when two or more work items are independent: they have disjoint owned paths, no
ordering dependency, and no shared mutable fixtures or state. If there is one item, an overlap, or
uncertain independence, continue directly through /si and state the reason once. Keep a wave to a
small set, normally no more than four workers.
Before dispatch, read git status --short, the task entrypoint, and the expected paths for each item.
Record the dispatch base revision for every isolated worker and the pre-dispatch inventory for a
shared checkout. If relevant inputs are already dirty, do not create workers from a clean HEAD; use
a coordinated shared checkout or continue sequentially. Leave unrelated changes untouched.
Dispatch
Use whatever worker and isolation capabilities the environment supports. Prefer isolated execution
when available. A shared checkout is acceptable only when each worker has disjoint owned paths and
the orchestrator coordinates all shared state. Give every worker one item, its task path, optional
criterion/context paths, owned paths, and the exact git authority it inherits. Default to no commit.
If no supported worker capability exists, continue directly through /si.
Dispatch independent workers up to the host's supported concurrency before waiting. With one
worker slot, execute items sequentially or directly while preserving their coverage. Adapt to the
available API and isolation mechanism. A worker returns complete, failed, or blocked using
the developer-agent JSON contract.
Consolidate
For each result:
- Inspect the reported files and the worker's status; do not trust a summary without reading the changed tree.
- For an isolated worker, compare the full owned tree with its recorded dispatch base, including committed, staged, and unstaged tracked changes. Use a binary diff from that base plus an explicit untracked-file listing for added, deleted, binary, mode, and symlink changes. For a shared checkout, inspect the changes in place against the pre-dispatch inventory; never replay them as a patch.
- For isolated workers, transfer every owned path from that manifest, then resolve conflicts before reading the destination back. Preserve ownership, file content, modes, symlinks, and binary data.
- Read back the destination inventory and compare it with the complete manifest. After all transfers, validate the combined tree from the commands and working directories resolved from project configuration. Worker checks are evidence for the item, not combined verification.
If a commit is authorized, stage only the assigned changes; whole-path staging is suitable only when that file's entire diff is in scope. Inspect the staged diff and name-status/stat so unrelated dirty changes remain outside the commit. Publication requires its own authority when not already granted.
Verification and handoff
Run the required lint, type, test, render, schema, or plugin checks for the combined tree. Apply TDD
to behavior changes; validate docs, formatting, and metadata with fit-for-purpose checks without
inventing tests. Treat configuration as behavior when it changes runtime behavior. Mark the combined
work complete only when every required check passes; record not_applicable with a reason where a
check does not apply, and never turn a skipped or unresolved check into passed.
Update the task record once after consolidation with the parallelization decision, worker evidence, combined verification, and next action. Workers do not edit shared task documents. Preserve the complete final file inventory and any real blocker for the next stage.
Safety rules
- One worker owns one scoped item; no scope creep or shared task writes.
- Reuse authorization already granted for the current task/request. Workers inherit explicit scoped git authority from the orchestrator and do not re-ask for blanket approval.
- Default: no worker branch and no worker commit. Publication is a separate gate when not granted.
- If a worker fails or is blocked, record the cause and resolve it sequentially; do not claim the combined task is complete from partial results.
Output
Report the items dispatched, each worker's status/files/evidence, transfer conflicts, combined verification, and the next sequential action. Keep the report concise and evidence-based.