Imported from nebius/nebius-ps-services (
skills/task-implementer/SKILL.md). Install upstream withnpx skills add nebius/nebius-ps-services --skill task-implementer. Copyright stays with the author.
Task Implementer
Help
For $task-implementer --help or $task-implementer -h (including native Claude forms), return concise help
and stop before any workflow step. State the purpose and invocation policy.
Show exact usage for every public action. Describe each public action,
positional argument, and flag in one concise line, including -h, --help; say
"No additional public flags" when there are no others. Use only the documented
public interface. For internal or coordinator-only skills, state that boundary
and that no standalone public workflow action exists. After the selected
SKILL.md is loaded, help is report-only: do not call any additional tools,
inspect project state, or modify files, private state, Git, or external
systems. Never expose private helper actions or flags or treat help as workflow
authorization.
Agent Compatibility
Before workflow reads, load ../global-context-management/references/agent-hosts.md
for selected-home paths, native identity and required runtime setup.
Use $task-implementer in Codex, /task-implementer in Claude Code, or
/skills:task-implementer in the Claude plugin. Dollar-prefixed skill examples
refer to the same named skill on either host; use the native invocation syntax.
Preserve the declared invocation policy, approvals and workflow ownership.
Use available native tools; an unavailable required capability is a blocker,
never permission to bypass a guard or claim unobserved behavior.
Purpose
Coordinate a complex brownfield request from one durable private Markdown prompt. A run compiles accepted intent, plans dependency waves, dispatches isolated workers, integrates verified commits, promotes each wave to a persistent lane, and resumes from durable state until complete or blocked.
Invocation Policy
Public entry requires explicit invocation. Keep
policy.allow_implicit_invocation: false in agents/openai.yaml.
prompt-session-intake may capture a safe project-intent projection for an
already-bound session, but it never selects or invokes Task Implementer.
That durable project-intent projection excludes shell/tool control text.
Capture and execution are separate operations; capture never authorizes
starting or resuming the workflow.
Public Interface
Expose exactly these actions:
$task-implementer workspace init [project-folder]
$task-implementer workspace reuse [project-folder]
$task-implementer run <prompt-ref-or-file>
$task-implementer integrate [project-folder]
$task-implementer workspace remove [project-folder]
workspace initinitializes or verifies the exact project workspace.workspace reusereopens an existing verified workspace without repair or lane mutation.runaccepts one exact managed prompt reference, ID, filename, or path.integrateconsumes completed lane generations into the recorded source branch after exact validation.workspace removeremoves only an idle, clean, fully integrated lane and preserves private prompt/run history.
When rendering report-only help, add this note after the workspace remove
description:
If the primary checkout is on a different branch from the lane's recorded
source branch, omitting [project-folder] from that checkout may resolve a
different workspace. Target the project directory inside the existing lane:
$task-implementer workspace remove "/Users/<username>/repos/<repository>-worktrees/<lane>/services/<project>"
State that each angle-bracketed path component is an illustrative placeholder
that the user must replace. Never insert an observed username, repository, lane
identity, or local path into help. This example uses only the public
[project-folder] selector; it is not an internal worktree identifier.
- Never require a user to provide run, wave, task, branch, worktree, lifecycle, or private-state identifiers.
- Do not add compatibility aliases or public internal transitions.
When To Use
- The user explicitly invokes one of the five actions.
- Durable steering, dependency ordering, isolated worker commits, combined validation, and resumable integration materially help the request.
- Parallel tasks have complete and disjoint write ownership.
When Not To Use
- Do not invoke it implicitly for a complex task or generic parallel request.
- Do not use it for one-shot implementation, Agentic SDLC, design-only work, standalone review/commit/PR/merge, release, or publication.
- Do not dispatch parallel workers when dependencies or ownership are unclear.
Inputs
- One exact explicit public action and selected project folder.
- The accepted immutable prompt reference plus current steering disposition.
- Verified workspace, lane, coordinator, task, assignment, result, journal, Git, Worktree, and applicable repository-instruction evidence for the active transition.
- Focused code, tests, docs, and validation commands needed by the locked plan.
Required Reads
- Read
references/prompt-workspace.mdbefore workspace, prompt, steering, queue, state-recovery, or sandbox decisions. - Read
references/prompt-requirements-refinement.mdbefore compiling or revising accepted intent. - Read
references/implementation-loop.mdbefore wave planning, dispatch, integration, promotion, cleanup, or finalization. - Read the bound prompt snapshot, handoff, coordinator, active wave, assignments, results, journals, applicable repository instructions, relevant code/tests/docs, and current Git state.
- Every worker uses
code-reviewand exactly one$commit. The coordinator uses combined validation, integrationcode-review, and final$align.
Writes
- Owner-only prompt, plan, queue, coordinator, task-plane, assignment, result, evidence, lease, and recovery records under the verified private workspace.
- Exact registered lane, worker, and integration worktrees plus journaled Git refs and commits owned by the active run.
- Coordinator-owned requirements, design, documentation, and changelog deltas only through the exact reviewed integration path.
- Never lifecycle state, project-instruction state, unrelated source dirt, or unregistered external systems.
Canonical Project-Spec Contract
- The root coordinator classifies every direct root-user statement and uses
maintain-project-specsas the single parser, paired publisher, and receipt owner. Durable intent is reconciled into requirements and a ready design before the implementation plan locks. - Missing or invalid canonical specs block only the spec-dependent plan lock or publication that cannot be performed safely. They never deny tools, Stop, cleanup, session completion, or independent recovery work.
- Spec status and receipts never authorize or prevent a Task Implementer transition.
- Task Implementer owns prompt-impact claims, receipts, plan settlement, and resource gates under its private run state. It binds the exact root-intent digest and project-spec receipt as plan evidence, not workflow authority.
- Workers inherit the root-intent digest and exact project-spec receipt. They
never reclassify direct user intent or edit canonical specs; discoveries are
returned as typed
spec_gapsfor coordinator reconciliation and replanning. Enumerate dirty and committed worker paths with rename folding disabled, and reject sensitive text in every gap summary/evidence item before persistence. - Do not expose or call a lifecycle authorization bridge.
- Do not require project-instruction inspection, rendering, mutation,
verification, reload, or sealing as part of a run. Read every already-
effective
AGENTS.mdin the applicable instruction chain. - A separately and explicitly invoked project-instruction workflow owns its mutations and safety checks. Its result may inform a later run but cannot become a prerequisite or terminal mutation for the current run.
- The project-instruction workflow remains separately responsible for every requested repository mutation and reload decision.
- Historical lifecycle evidence may remain readable for diagnosis. Ignore it for active decisions; do not reconstruct, migrate, or refresh it automatically.
Private State
Keep prompts, immutable revisions, queues, coordinator state, assignments, results, journals, validation evidence, and summaries under the verified private Task Implementer root outside Git. Files are owner-only, bounded, regular, non-symlinked, and atomically written. Never persist prompt bodies, diff bodies, secrets, private endpoints, raw logs, or credentials in public status or repository artifacts.
Process
Workspace initialization and reuse
- Resolve the exact Git common directory, primary checkout, named non-default source branch, selected project scope, and applicable instructions.
- Initialize or verify the Worktree-owned persistent full-repository lane and generated prompt workspace. Preserve unrelated source-checkout dirt; never copy it into the lane.
- Treat generated workspace mismatch as tampering unless the existing narrow
executable-relocation repair contract applies to
workspace initorrun.workspace reuseis strictly non-mutating. - Return only public-safe paths and observed status.
Run intake and planning
- Bind one accepted prompt revision and one active lane generation under the scope and execution locks. Queue a different explicit prompt FIFO when another run is active.
- Compile the prompt into a complete Task Implementer prompt-impact claim and
stable clarification IDs. Classify statements, then publish the canonical
requirements/design pair through
maintain-project-specs. Material ambiguity blocks planning; historical lifecycle state does not. - Bind the immutable plan to the latest accepted revision, root-intent digest, current canonical v2 spec receipt and bytes, task IDs, dependencies, exact/prefix write claims, conflict domains, validation commands, and done criteria.
- Review a resource-free dirty lane completely before the built-in checkpoint transaction. Preserve unrelated, sensitive, unsupported, or ambiguous changes as a blocker. Never reset, stash, clean, amend, unstage, or manually commit the managed lane.
- Reconcile steering only at a proven resource-safe boundary. Material change requires a distinct plan identity; never rebind unchanged plan bytes.
Dependency waves
Wave states are:
planned -> preparing -> running -> integrating -> promotion_pending -> promoted -> cleanup -> done|blocked
Task states are:
planned -> assigned -> running -> committed -> merged|failed
For each wave:
- Require the persistent lane to be clean at its recorded branch and
HEAD. - Journal resource intent; create one locked full-repository integration worktree from that exact commit; re-observe Git identity.
- Create immutable assignments only for the active capacity batch. Each assignment binds absolute scope cwd, base commit, helper digest, task digest, root-intent digest, canonical project-spec receipt, claims, domains, validation, criteria, guardrails, and predecessor evidence.
- Dispatch one fresh worker per assignment. Prefer native subagents; the
sequential fallback uses ephemeral Codex exec or Claude print mode with the
selected private home and a fresh identity. Preserve native permissions and
hooks; unavailable capabilities never authorize bypassing them.
A worker may mutate only its own
registered worktree and claims, validates its task, performs
code-review, invokes$commitexactly once, and publishes immutable result evidence. Its result always containsspec_gaps; a non-empty list requiresREPLAN_REQUIRED. - The coordinator independently verifies assignment, worker session, commit ancestry/tree/paths, result digest, and validation evidence. It never implements a worker task.
- Merge verified worker commits into the integration branch in stable task-ID
order with
git merge --no-ff --no-edit. Never cherry-pick, rebase, squash, or push internal branches. - Apply coordinator-owned requirements/design/docs/changelog reconciliation through the exact journaled coordinator commit path. Record implementation and verification evidence in the canonical pair after combined proof. Product fixes become correction tasks.
- Run combined validation and integration
code-review, bind evidence to the unchanged integration tip, then remove only verified clean worker resources. - Fast-forward the persistent lane under the shared promotion lock after exact preconditions and postconditions. Tasks become done only after promotion.
- Reconcile workflow-owned prompt impact, remove the exact integration worktree/ref, record retained resources if any, and continue. No project lifecycle receipt or project-instruction mutation is part of cleanup.
Recovery and resume
- Resume from authoritative coordinator-v7, task-plane, journal, Git, Worktree, interop, and lease evidence. Handoff Markdown is a projection, not authority after coordinator creation.
- Every mutation is intent-journaled and re-observed. Replays reuse only exact validated effects and immutable arguments.
resume-control-v1binds controlled transitions and returnsRESUME_STALEbefore effect when its token or observed basis changed.- A fresh heartbeat waits. Expired workers or ambiguous prior ownership require confirmation. Dirty, divergent, malformed, or unknown-writer state remains blocked with resources retained.
- Coordinator v1 through v6 remain unsupported. Do not add a compatibility reader, state shim, or inferred migration.
- If promotion reports failure, classify the actual lane
HEADas unchanged, promoted, or unexpectedly moved before retry.
Finalization and source integration
- After the last wave cleanup, run changed-surface
$alignand pass its bounded evidence to the private finalizer. - Finalization requires every wave done, no retained internal resource, a
clean lane at the final promoted
HEAD, current workflow-owned prompt impact, and a stable summary projection. It never requires lifecycle seal evidence. - Release the immutable generation and activate the unchanged FIFO queue head. Repeating an unchanged completed prompt returns the same summary.
- Public
integraterequires no active generation, contiguous released receipts, and clean exact source/lane identities. Build and validate one exact two-parent candidate when histories differ, then advance the source ref only by expected-old compare-and-set. - Never push, open a PR, publish, or merge remotely without a separate authorized workflow.
Idempotency
- Reopening an unchanged workspace verifies and reuses the same lane without repairing or duplicating resources.
- Replayed transitions reuse only exact journaled effects, immutable arguments, and current compare-and-set preconditions.
- Repeating a completed prompt returns its existing released summary; it does not create another generation, plan, commit, or integration candidate.
- Stale tokens, changed prompt revisions, head drift, partial effects, or ambiguous ownership stop before effect and route through owned recovery.
Failure Handling
- Worker, merge, validation, review, steering, promotion, cleanup, or source integration failures leave the last proven branch head unchanged and retain exact recovery resources.
- Use precise codes such as
WORKSPACE_BUSY,REPLAN_REQUIRED,WORKTREE_CONFLICT,WORKER_SCOPE_VIOLATION,VALIDATION_FAILED,PROMOTION_BLOCKED, andCLEANUP_BLOCKED. - Lifecycle advisory failure is never converted into one of these workflow blockers.
- Do not force-remove worktrees, force-delete refs, initialize submodules, widen permissions, or perform external writes to make recovery pass.
Must Not
- Expose internal IDs or transitions in user commands or status.
- Let workers touch the primary checkout, other worktrees/refs, shared coordinator files, undeclared paths, Git maintenance, or external systems.
- Run overlapping tasks in parallel.
- Trust PID, timestamp, Markdown, or lifecycle state as proof that a writer is quiescent.
- Persist or print secrets, private identities, raw diffs, or private paths.
Completion Criteria
- The public interface remains exactly five explicit-only actions.
- Every task has deterministic dependencies and locked ownership.
- Every promoted task has one verified worker commit plus combined validation and review evidence.
- Cleanup leaves no unreported temporary resource.
- Final
$alignpasses; the generation is released from a clean exact lane head without lifecycle gating. - Public integration, when requested, advances only the recorded source ref through an exact validated candidate.
Output Contract
Return the versioned public run summary or one precise blocker and next public action. Status is strictly read-only, double-observed, and identity-safe. Do not expose prompt content, lifecycle/private state, raw logs, branches, commits, or internal paths.
Learning Loop
When using this skill, capture durable, reusable, public-safe learnings in the narrowest appropriate surface only when the task contract allows source edits. For read-only/report-only work, or when a learning is not public-safe, evidence-backed, in scope, or free of unverified/vendor-specific claims, do not edit skill sources; report that it was skipped. Do not capture secrets, private URLs, customer data, raw logs, or one-off local state.
References
references/prompt-workspace.mdreferences/prompt-requirements-refinement.mdreferences/implementation-loop.mdassets/handoff-template.md