feat: autonomous delivery harness — event-driven worker/watcher split

[IcyWednesdays]

Summary

Replaces Craft's in-worker PR polling loop with a separate, event-driven harness. Workers now exit after creating a draft PR; a deterministic bash watcher takes over post-PR, polls gh, and dispatches narrow remediation agents (e.g. /fix-ci-failure, /address-review-comment) only when events fire. Adds a pluggable state backend so the harness can later swap to Linear (or any tracker) without changes to the orchestrator or skills.

What's in the change

5 commits, each a coherent unit:

#	Commit	What it adds
1	docs: autonomous delivery harness design	Architectural spec covering all subsystems
2	feat(state): pluggable state backend abstraction	state.sh dispatcher + state-local.sh file backend with high-level operations (state_claim_task, state_mark_waiting, state_mark_done, state_mark_blocked, state_append_note, state_create_subtask, etc.); orchestrator + bin/craft migrated; STATE_BACKEND=local config
3	feat(watcher): worker/watcher split + event-driven dispatch	bin/lib/watcher-events.sh (pure event detection), bin/watcher.sh (polling loop), spawn_watcher_pane on tmux + cmux, orchestrator watcher tracking, /work-task rewritten to exit after PR, first remediation skill /fix-ci-failure
4	feat(ui): split-pane watcher UI + dashboard freshness	bin/render-watcher-status.sh snapshot + tmux/cmux split (status above, log below, refreshes every 5s); bin/render-dashboard.sh with per-watcher freshness signals, CI badges, recent activity feed; milestone-completion log dedupe
5	feat(watcher): comment dispatch + detached remediation panes	/address-review-comment skill; new_comment_received and new_review_received (filtered to non-approvals) → dispatch the skill; spawn_task_pane_detached so remediation runs in its own unfocused pane; smart completion (pane-exit OR PR head SHA changed OR 5min timeout) with force-kill at the end

Architecture

Architect (Claude Code)            User-facing planning agent
       ↓
Orchestrator (bash daemon)         Polls queue, spawns workers/watchers
       ↓
Worker (short-lived agent)  →  exits after creating draft PR
       │ handoff via state_mark_waiting
       ↓
Watcher (long-lived bash)          Polls gh, detects events
       ↓ on event
Remediation agent (transient)      Narrow skill: /fix-ci-failure, etc.

Watcher event taxonomy:

Event	Action
ci_failed	dispatch /fix-ci-failure
merged	state_mark_done, terminal
closed	state_mark_blocked, terminal
draft_to_ready	fire on_ready plugin hook
new_comment_received	dispatch /address-review-comment
new_review_received (non-approval only)	dispatch /address-review-comment
new_review_approved, mergeable_conflicting, first_poll, ci_passed	log only (deferred)

Validation

Tested live on real Linktree PRs (developer-platform-libraries):

• Worker → PR creation → watcher spawn → polling → event detection → remediation dispatch → terminal exit on merge: end-to-end working.
• Comment-driven dispatch: posted comments triggered /address-review-comment, which made fixes and pushed; watcher detected the SHA change and tore down the remediation pane cleanly.
• Approval suppression: pure approvals don't fire the LLM dispatch (verified by upstream filter in watcher_diff_events).

Test plan

• make test green: 149 assertions across 4 suites (test-queue.sh, test-state.sh, test-watcher-events.sh, test-render-watcher-status.sh)
• bash -n clean on all modified shell scripts
• End-to-end smoke on a real repo: worker creates draft PR, watcher polls, comment triggers /address-review-comment, merge tears down watcher cleanly
• Approval-only review correctly suppressed at the watcher level (no LLM dispatch)
• For reviewer: kick the tyres on cmux — the cmux mux provider has a known issue with workspace creation that's not fully resolved (tested fully on tmux)

Known limitations / deferred follow-ups

• Comment ID idempotency — currently the watcher tracks comment_count only, so the agent's own replies could re-dispatch; folded into a future plan.
• Skill template propagation — new skills under templates/.claude/commands/ aren't auto-copied into existing projects; requires either a craft sync command or a fallback in dispatch_remediation.
• Additional remediation skills (/resolve-merge-conflict, /investigate-pr-close) and operator-override skills (/pause, /kill, /restart, /watch) are designed but not built.
• WORKER_AUTO_PERMISSIONS config knob to gate --dangerously-skip-permissions — not yet added (workers run autonomously by default, which is the existing upstream behaviour).
• Worker pane /exit doesn't terminate Claude Code's process (only fixed for remediation panes via the force-kill in commit 5).

🤖 Generated with Claude Code