Dispatch Tickets

Use the dispatcher for scheduling, process lifecycle, validation, local commits, serial integration, and durable state. Agents own design, bounded coding, failure diagnosis, independent review, and acceptance. Scheduling requires no model calls. Each implementation ticket gets a fresh CLI process and worktree.

Prepare The Approved Graph

Read references/graph.md for the executable schema, commands, worker contract, and recovery behavior. Derive the graph from the approved plan; do not invent scope or publish tracker issues to make it runnable. Tasks need settled contracts, explicit owned paths, prerequisites, and meaningful checks. Split oversized tasks before execution. Record approval provenance in the parent decision record; approved: true is a machine gate, not user consent.

Use the current isolated integration worktree and a clean, committed baseline. Keep graph, briefs, state, and logs under ignored .agent/work/<slug>/. Worker worktrees live in a sibling directory outside the checkout; each has its own ignored task packet. Follow repository Git rules; use explicit branch/base refs. Do not overwrite unrelated changes or silently change an existing branch.

Configure worker_command as an argv array, with an optional per-ticket override. The runner must accept its task on stdin, work in its current directory, and exit when the assignment is finished. Codex, Claude, or another CLI/wrapper can use this protocol. Resolve the executable, model, permissions, and authentication before running; inherit the user's configured access, and never silently switch providers or relax sandboxing. Do not put credentials in command arguments.

Main automatically selects a smaller, lower-cost coding model for implementation unless the user specifies otherwise or a task has a documented reasoning need for a larger model. Resolve current supported models; do not inherit the main agent's model or an unchecked CLI default. Encode the selection in the command or a verified profile, including every per-ticket override. If unavailable, disclose actual fallback and honor user cost/model constraints. The runner executes these commands; it does not choose or verify model capability.

Preview And Execute

Resolve <skill-dir> to this installed skill's absolute directory. Run its bundled launcher; it compiles bundled Rust source on first use and lets Cargo rebuild changed source. Artifacts stay in the user cache, outside the skill and target repository. Requires Rust/Cargo, Git, and first-build access to crates.io; no prebuilt binaries or repository-root tooling are needed.

On macOS/Linux:

"<skill-dir>/scripts/run" --repo <integration-worktree> \
  --graph <work-item>/tickets.json status
"<skill-dir>/scripts/run" --repo <integration-worktree> \
  --graph <work-item>/tickets.json run --jobs 3

On Windows, invoke & "<skill-dir>/scripts/run.ps1" with the same arguments. Launchers retain the caller's working directory and forward arguments and exit status. Set XDG_CACHE_HOME on macOS/Linux to select another build cache.

status validates and previews without mutations or model calls. run starts local coding-agent processes and performs Git writes. A request to execute the approved implementation covers those local actions; preview-only requests do not. The dispatcher does not push, publish a PR, write to issue trackers, deploy, or conduct the final independent review.

Monitor the foreground process through the runtime's resumable execution tools; keep user updates concise. In Claude Code, start run with Bash run_in_background (foreground calls cap at 10 minutes, below the worker timeout) and wait for its exit notification; read dispatch/state.json and attempt logs meanwhile. The dispatcher launches only tickets whose prerequisites have passed integrated checks, fills available capacity, and serializes tickets with overlapping owned paths. It integrates completed results one at a time and runs the graph's integration checks before releasing dependent tasks.

Workers receive only their own brief plus workspace/ownership constraints. They edit and may validate their assignment, but the dispatcher owns commits and merges. Retain the resulting logs, exact starting commits, worker commands, result commits, merge commits, and check outcomes as execution evidence. A zero worker exit code alone does not establish acceptance.

Recover And Accept

On failure the loop stops and terminates active worker groups, retaining work, logs, and attempt history. Inspect status, logs, branches, and live processes. No automatic retry, destructive reset, or worktree cleanup occurs. Retry only selected tickets after diagnosing the failure; never retry an integrated ticket as if it were unfinished. Use a new graph for new repair scope or changed intent. Graph and brief changes invalidate an existing run rather than silently moving its goalposts.

Reconcile an externally committed, inspected repair with --accept-head <commit> when needed. After a merge whose integration checks failed, --retry <id> reruns those checks instead of reapplying the worker result. Resolve an interrupted Git operation before restarting; the script only recovers a completed merge when its recorded parents match exactly. Do not edit state to bypass a failure.

The main agent verifies the combined result and planned acceptance, performs the selected independent review, and directs bounded repairs. Script completion means graph checks passed, not that review or publication is complete. Use $explain-implementation and $pr for the authorized visual PR handoff.