agent-pipeline

OMP Native Agent Pipeline

Use for /auto go SPEC-ID [--execution-owner omp|orca]. This entrypoint selects work and context; open only the reference needed for the current step.

Need	Open
Phase entry, work, and exit conditions	references/phases.md
Worker ownership and context isolation	references/delegation.md
Change classification and gate evidence	references/gates.md
Tests, smoke checks, and UX verification	references/verification.md
Review authority and termination	references/review.md
Completion and sync readiness	references/completion.md
Configured quality and permissions	references/quality-modes.md

Execution Decision

Work in the current session by default. Dispatch only genuinely independent slices, a necessary specialist review, or work that needs isolated context. File count and line count do not justify delegation. Keep dependent or same-file work sequential; one migration-numbering lane has one writer.

A verified low-risk compact change contract omits separate planning and test scaffold dispatch: implementation still starts with the relevant failing test, then validation and review run. Missing, ambiguous, or escalated authorization keeps the full route. Report the actual route_phase_set from auto pipeline run, not an imagined phase checklist. Explicit --solo, --team, and --multi requests must retain their requested semantics.

Execution Owner

Choose exactly one DAG owner before any dispatch: --execution-owner omp|orca. Omission selects omp. Reject duplicate, aliased, or conflicting owner flags; never retry a failed explicitly selected owner as another owner.

• Owner omp: this session owns progress through native task and todo, coordinating jobs with wait, read proc://, write proc://<id>/kill, and write agent://<id> (OMP 18.3+) or hub (older OMP). --solo keeps work in this session; --team requests a native batch with explicit responsibilities. The two flags conflict.
• Owner orca: only an explicit --execution-owner orca selects it. First run and read orca skills get orchestration --full; the Orca Run owns the DAG. Do not create a competing OMP task or progress DAG. --team and --solo conflict with this owner.
• --multi requests risk-tiered provider-diverse evidence, not a second DAG.

auto pipeline run SPEC-ID --platform omp --execution-owner orca performs the integrity check and records pipeline_execution_owner_receipt.v1 before any worker session. The body-free .autopus/pipeline-state/<SPEC-ID>.execution-owner.json records owner, source, identity, and verification status.

Native Dispatch

Before dispatch, inspect the tools actually exposed by this runtime. Use their current schemas rather than copying static payloads. If the selected topology requires an unavailable tool, report that blocker; never simulate dispatch.

OMP serves this project's 17 Autopus agents from .omp/agents/: annotator, architect, debugger, deep-worker, devops, executor, explorer, frontend-specialist, perf-engineer, planner, reverse-engineer, reviewer, security-auditor, spec-writer, tester, ux-validator, and validator. Select the agent of the role a worker plays:

• planner, architect, and spec-writer plan and design; executor, debugger, tester, and the other implementation roles write code.
• explorer is read-only exploration only. Never give it work that writes.
• reverse-engineer only analyses an existing app or binary into a behaviour specification; never give it implementation work.
• reviewer, validator, and security-auditor review and verify; keep them off implementation.
• Omitting the agent runs the bundled task agent, the default for work no role fits. Every other bundled agent (scout, security-reviewer, sonic, designer, librarian) is disabled through task.disabledAgents; never select them.

The agent field names the role; the assignment text still states its scope. State the responsibility, scope, and exit condition in the task itself.

Batch independent owned slices together. Native OMP owns isolated worktree creation, integration, and cleanup. Request isolation only if the current task schema exposes it; do not manually merge or remove its worktrees.

Each worker receives scope, forbidden paths, acceptance criteria, and only its needed context. Request the five-field result contract: owned_paths, changed_files, verification, blockers, next_required_step. Use native schema validation where supported; worker assertions are not execution proof. Record actual dispatch count and the native agents used. Follow up with the same revivable worker through write agent://<id> (OMP 18.3+) or hub (older OMP); an isolated worker whose workspace has been released needs a new assignment. Only the parent owns todo.

A worker that stops at its turn limit returns a partial result; the turn-budget rule every assignment carries (references/delegation.md, Worker hygiene) keeps that rare. Never treat a partial result as a completed phase or feed it to the next phase. Follow up once with a narrowed instruction (finish and write the artifact, minimal further exploration), through that follow-up path or as a new assignment as above; if that run is partial again, stop and hand the phase to the user with what was produced.

Context and Models

Keep required core/SPEC bodies complete and hash-verified. Ordinary auto workflow context includes optional architecture only when explicitly selected. The managed OMP evidence path independently rebuilds its canonical full context: do not substitute a smaller CLI manifest, alter its signed protocol, or promote an unmeasured compaction result.

Give workers task deltas and retrievable references instead of replaying raw provider/tool output. Validate paths and hashes; reject missing required context, traversal, symlinks, stale or wrong-SPEC input before dispatch. Use native context management rather than layering an extra summarization loop.

Inherit the current OMP session model and thinking configuration unless an explicit validated role-model profile has written task.agentModelOverrides for the native agent being dispatched. That configuration is the only model authority: never pass a guessed model ID, and never pick a cheaper agent to fit a larger task.

Gates and Completion

Use auto spec gates for applicability and exact-input evidence reuse. Security, validation, data-loss, and deterministic-oracle gates remain active; UI changes retain accessibility and UX verification. Annotation is opt-in via --annotation or a direct request, not an automatic extra worker phase. Honor explicit project/route numerical gates; do not invent universal floors.

Integrate changes before the merged verification run. Keep a verdict and real evidence per acceptance criterion. Failed tests or security findings cannot be overruled by provider consensus. Re-review only changed/open findings; unchanged input is not a reason to repeat discovery. Respect the existing retry limits.

Finish only after the changed path has run, required acceptance is satisfied, and blocking findings are closed. Use references/completion.md for the sync receipt. Status combines project/SPEC state with the selected owner's native job state; external commands must not inspect user-owned OMP session roots or claim they replace native job observations (read proc:// on OMP 18.3+, hub on older OMP).