Autonomous missions: charter → plan → phase gates → code-gated done

[BillJr99]

Summary

This PR implements a complete autonomous mission orchestrator that drives a single objective from charter through discovery, product, engineering, validation, and deployment phases to a machine-checkable done gate—with no human follow-up required.

The mission runner wraps existing per-workflow engines (WorkflowRunner, RalphLoop, AutoresearchLoop) in a macro plan→execute→evaluate→refine→iterate loop. Every boundary is trace-committed to git so the execution trail is auditable, and per-cycle telemetry makes the loop observable.

Key Changes

New modules:

• mission.py — MissionRunner orchestrator, MissionPlan/PhaseSpec living artifacts, charter-first ordering
• charter.py — Chief's up-front commitment (mission statement, scope, success criteria, constraints); persisted as .clk/state/charter.json + CHARTER.md
• done_gate.py — Machine-checkable completion gate (tests green, deliverables exist, QA pass, ralph pass, no TODOs, charter criteria); adaptive for test-less projects
• deliberation.py — Self-reflection preamble + blocking Q&A for producing dispatches
• telemetry.py — Per-cycle accumulator (thread-safe) for loop observability; renders compact line + activity event
• noop_guard.py — Detects when producing agents emit prose without file mutations; escalating re-dispatch with repair preamble

Modified modules:

• workflow.py — Integrated done-gate evaluation, telemetry threading, mission-mode flag, supervise-cycles override; added deliberation preamble injection
• evaluator.py — Auto-derives real test commands (pytest/npm/compileall) instead of vacuous pass; marks weak gates for adaptive done-gate
• agent.py — Tracks file mutations applied per run; added mission-level chief dispatch types (charter, mission_plan, phase_gate)
• casting.py — Added CharterProposal and PlanProposal parsing
• cli.py — New mission command (alias auto); run now drives full mission by default with --once for single cycle
• config.py — New config blocks: mission.*, done_gate.*, deliberation.*, validation.auto_derive
• git_ops.py — New commit_trace() for structured execution-trace commits at mission boundaries
• tui.py — Mission dispatch in TUI job handler

Tests:

• test_done_gate.py — Unit tests for gate verdicts (tests red, missing deliverables, QA fail, etc.)
• test_mission_living_plan.py — Plan/charter roundtrip, phase ordering, validation derivation
• test_noop_guard.py — Mutation detection and response-quality flag
• test_debate.py — Adversarial debate-panel refinement
• test_mission_e2e.py — End-to-end mission with shell provider (verifies gates fire, artifacts written, telemetry emitted)

Documentation:

• README updated with autonomous-missions section and new knobs
• .env.example and kickoff.sh updated with mission/done-gate/debate config
• Inline docstrings explain charter-first ordering, gate semantics, and trace-commit strategy

Notable Implementation Details

1. Charter-first ordering: Mission always asks chief to write charter before proposing plan; plan is derived from charter so "done" is measured against committed criteria, not drifting.

2. Living plan artifact: .clk/state/mission.json + MISSION.md persisted after each phase, showing status, iterations used, gate history; enables mid-mission inspection and recovery.

3. Adaptive done-gate: When evaluator marks gate as "weak" (no real test command available for docs/research projects), tests-green requirement is relaxed so mission can still finish.

4. No-op guard: Producing stages that apply zero file mutations are re-dispatched with escalating repair preamble; prevents agents from "describing work instead of doing it."

5. Deliberation: Self

https://claude.ai/code/session_01AHao68VweSg6a9Q1sJUR65