Matrix Coder Phase 3 — Workflow / loop skills (Ralph, autopilot, ultrawork, ultraqa)

[Interstellar-code]

Epic: #76 · Phase 3 of 6

Goal

Add the workflow/loop skills that give Matrix Coder its "system feel" — orchestration layers that drive the roles in loops and fan-outs. These pair with Phase-2 Kanban so long-running loops are observable live.

Depends on

Phase 1 (roles to orchestrate). Strongly benefits from Phase 2 (Kanban visibility of iterations/branches).

Scope / Deliverables

• Ralph loop — repeat executor → verify until verification passes (bounded, with stop criteria).
• autopilot — full autonomous chain plan → executor → test → review → verify from one ask.
• ultrawork — parallel high-throughput fan-out of roles across files/topics (maps to Kanban child cards).
• ultraqa — test → verify → fix cycle until the goal is met.
• Each workflow emits the shared output contract and (when Phase 2 is present) writes parent/iteration/branch Kanban cards.

Ships when

A user can run an end-to-end autonomous coding loop (e.g. autopilot or Ralph) that self-corrects via verify and is observable on Switch UI.

References

#76 runtime/user-flow comments; chaining diagram.

Attribution

Ralph, ultrawork, ultraqa, autopilot adapt OMC (MIT) — per-skill attribution.

Revisions (post advisory review)

• No built-in loop primitive. delegate_task runs a child to completion and returns a summary. Ralph/ultrawork/ultraqa are MC-implemented orchestration loops over repeated delegate_task calls in harness.py, not wrappers over an existing feature.
• Watch the caps: per-subagent iteration_budget=None (tools/delegate_tool.py:1136) and spawn-depth default 2 (_get_max_spawn_depth) must not cap a loop prematurely.
• Wire cancellation: honor interrupt_subagent (tools/delegate_tool.py:188); add stop/cancel semantics to acceptance criteria.

[Interstellar-code]

Advisory Plan Review — opus critic (code-verified, pre-implementation)

• [MED] There is no built-in loop primitive. delegate_task runs a child to completion and returns a summary; Ralph/ultrawork/ultraqa imply re-dispatch loops MC must implement itself in harness.py. → Scope these as MC-owned orchestration loops over repeated delegate_task calls, not thin wrappers. Verify the iteration budget (iteration_budget=None, tools/delegate_tool.py:1136) and depth/concurrency caps (_get_max_spawn_depth, default 2) don't cap a Ralph loop prematurely.

• [LOW] No stop/cancel semantics specified. Hermes has interrupt_subagent (tools/delegate_tool.py:188). → Wire MC loops to honor it; add cancellation to Phase-3 acceptance criteria.

[Interstellar-code]

✅ Phase 3 v1 complete — workflow skills

Built via sonnet executor → sonnet code-review → review fixes applied. Commit 081988b50, branch feat/matrix-coder (PR #119).

Pragmatic design (day-1, refine during testing)

No Python orchestration engine. A workflow is a persona that instructs the live Hermes agent to run the loop using its own tools — the agent IS the parent_agent, so it can call delegate_task (fan-out/subagents), edit, run tests, etc. MC just composes + injects the workflow procedure via the existing trigger path. This sidesteps the hard "multi-step orchestration from a hook" problem (hooks can't spawn subagents) and lets us harden looping behavior during real testing.

Delivered

• 4 workflow personas (personas/workflows/): ralph (executor→verify loop, cap), autopilot (plan→executor→test→review→verify chain), ultrawork (decompose → fan-out via delegate_task with disjoint file sets / worktree isolation → aggregate), ultraqa (test→verify→fix cycle, cap + no-progress guard).
• Each references _base, emits the four-section output contract, has explicit STOP criteria (so loops terminate), mandates single-writer-per-file for any parallel fan-out, and correctly states the agent uses its own tools (no false "engine" promises).
• intake.py WORKFLOWS set; workflow parses as the role token (no lens). registry.load_persona falls back to personas/workflows/<name>.md. /matrix help lists workflows. plugin.yaml → 0.4.0.

Verification

py_compile OK; all 6 test modules green (stdlib runners; no venv). Review fixes applied: stale grammar + ParsedInvocation docstrings corrected; narrowed a sensitive-pattern false-positive (workflow → ci workflow); added load_persona fallback regression tests.

Refine-during-testing follow-ups (tracked for later, not blocking)

Actual loop robustness (iteration termination in practice, parallel-writer worktree wiring, parent/child Kanban hierarchy for fan-out) will be tuned against real runs — per the "good system day 1, not perfect" approach.

Next: Phase 4 (#116) — domain packs (text composition), or Phase 5 (#117) — implicit routing.