5.1: Slim harness to single-stage git-lifecycle runner

[savitharaghunathan]

5.1: Slim harness to single-stage git-lifecycle runner

Part of Stream 5: Harness → AgentPlaybookRun alignment

The harness is currently a full CLI with init, run, status, resume, step commands and a hardcoded 5-step pipeline (detect → plan → execute → verify → fix-loop) baked into Go code. Per ADR-0001, the harness should be a thin git-lifecycle wrapper — clone, strip creds, checkout branch, launch goose serve, commit, push. Pipeline orchestration belongs to the AgentPlaybookRun controller, which creates sequential AgentRuns (one per stage).

Tasks

• Reduce to a single run command driven entirely by env vars (KONVEYOR_MODEL_*, GIT_REPO_URL, KONVEYOR_INSTRUCTIONS, etc.)
• Remove interactive init command and file-based config — use config.LoadFromEnv() only
• Discover exactly one skill from /opt/skills/*/SKILL.md at startup
• Build prompt from 4 context layers: Agent prompt (KONVEYOR_PROMPT), playbook context (KONVEYOR_PLAYBOOK_INSTRUCTIONS), skill content, stage task (KONVEYOR_INSTRUCTIONS)
• Create a single ACP session, send one prompt, block until completion
• Final commit + push on exit
• Remove dead packages: internal/detect, internal/plan, internal/execute, internal/verify, internal/fixloop, internal/metrics, internal/rundir, internal/handoff
• Remove internal/goose/recipe.go, internal/goose/acp_runner.go (pipeline-specific ACP orchestration)

Acceptance Criteria

• Harness has a single run subcommand, no interactive prompts
• All configuration comes from environment variables
• Exactly one skill is discovered from /opt/skills/*/SKILL.md
• Agent receives a composed prompt (agent prompt + playbook context + skill + instructions)
• On exit: harness commits all changes and pushes to the target branch

References

• ADR-0001 §Git-as-persistence
• ADR-0001 §Agent container image strategy
• PR ✨ Harness init #33