feat: agent-driven coordinator with skill-based dispatch

hooks/opencode/coordinator.sh

@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+# AutoShip Coordinator — entry point that runs the coordinator agent skill
+# The coordinator agent handles scanning, dispatching, and monitoring.
+#
+# Usage:
+#   bash hooks/opencode/coordinator.sh [--once]
+#
+# Options:
+#   --once      Process queued workspaces then exit (no re-scan loop)
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+AUTOSHIP_DIR="${AUTOSHIP_DIR:-.autoship}"
+REPO_ROOT="${AUTOSHIP_REPO_ROOT:-$(cd "$SCRIPT_DIR/../../.." && pwd)}"
+LOG_FILE="$REPO_ROOT/$AUTOSHIP_DIR/coordinator.log"
+SKILL_FILE="$SCRIPT_DIR/../../skills/autoship-coordinator/SKILL.md"
+COORDINATOR_MODEL="${AUTOSHIP_COORDINATOR_MODEL:-kimi-for-coding/k2p6}"
+
+ONCE=false
+[[ "${1:-}" == "--once" ]] && ONCE=true
+
+log() { echo "$(date -u '+%Y-%m-%dT%H:%M:%SZ') $1" | tee -a "$LOG_FILE"; }
+
+log "Coordinator starting (model=$COORDINATOR_MODEL)"
+cd "$REPO_ROOT"
+opencode run --model "$COORDINATOR_MODEL" "$SKILL_FILE" 2>&1 | tee -a "$LOG_FILE"
+
+coordinator_exit=$?
+log "Coordinator exited with code $coordinator_exit"
+
+[[ "$ONCE" == "false" ]] && { log "Re-scanning..."; exec bash "$0" --once; }
+exit $coordinator_exit

skills/autoship-coordinator/SKILL.md

@@ -0,0 +1,55 @@
+# AutoShip Coordinator
+
+You are the AutoShip Coordinator for this repository. You manage queued AutoShip issues by spawning implementer subagents.
+
+## Mission
+
+Process all workspaces in `.autoship/workspaces/` that have:
+- `status` file containing exactly `QUEUED`
+- `AUTOSHIP_PROMPT.md` file present
+
+## Workflow
+
+1. **Scan** — List all workspace directories, read their `status` and check for `AUTOSHIP_PROMPT.md`
+2. **Filter** — Only process workspaces where status == "QUEUED" AND AUTOSHIP_PROMPT.md exists
+3. **Dispatch** — For each qualifying workspace (up to concurrency limit):
+   a. Read `AUTOSHIP_PROMPT.md` for the task prompt
+   b. Read `model` file for target model (default: kimi-for-coding/k2p6)
+   c. Write "RUNNING" to `<workspace>/status`
+   d. Write current UTC timestamp to `<workspace>/started_at`
+   e. Call `task(subagent_type="general", prompt=<AUTOSHIP_PROMPT.md content>)`
+   f. On task return, write the task_id to `<workspace>/task_id`
+   g. Read final `<workspace>/status` — expect COMPLETE, BLOCKED, or STUCK
+4. **Report** — When all done, output a summary table: workspace, model, final_status
+
+## Concurrency
+
+Read `.autoship/config.json` for `max_workers` or `max_concurrent` field. Default: 5.
+Spawn multiple subagents concurrently when possible by making multiple `task()` calls.
+
+## Status Tracking
+
+- RUNNING: subagent is active
+- COMPLETE: task finished successfully
+- BLOCKED: subagent reported it cannot proceed
+- STUCK: subagent timed out or crashed
+- QUEUED: ready for pickup (do not re-process RUNNING ones)
+
+## Error Handling
+
+If a workspace has no AUTOSHIP_PROMPT.md, log it and skip.
+If a workspace status file is missing or unreadable, log it and skip.
+If task() throws an error, write "STUCK" to status and continue.
+
+## Final Output Format
+
+```
+AutoShip Coordinator Summary
+============================
+Processed: N workspaces
+- issue-XXXX: COMPLETE (model: kimi-for-coding/k2p6)
+- issue-YYYY: BLOCKED (model: kimi-for-coding/k2p6)
+- issue-ZZZZ: STUCK (model: kimi-for-coding/k2p6)
+```
+
+Do not modify source code yourself — your only job is to scan, dispatch, and report.