dispatch.codex.nu — Hybrid Spec Dispatcher

Routes specs to Codex CLI (gpt-5.3-codex) or Auto-Claude (run.py) per spec. Pre-plans every spec via Ollama cloud (qwen3-coder:480b-cloud) before dispatch. Tracks state in .codex/dispatch-state.json. Spawns named Zellij panes.

Projects

Quick reference

nu dispatch.codex.nu plan                        # show all waves + agent + plan status
nu dispatch.codex.nu status                      # done / failed / pending counts
nu dispatch.codex.nu pre-plan                    # queue all specs through Ollama now
nu dispatch.codex.nu run --wave 1 --yes          # run wave 1 (skips confirmation)
nu dispatch.codex.nu run --yes                   # run all pending waves
nu dispatch.codex.nu watch-handoff --from-wave 1 --to-wave 2 --yes
nu dispatch.codex.nu mark-done 011              # manually mark complete
nu dispatch.codex.nu mark-failed 011            # manually mark failed
nu dispatch.codex.nu reset --spec 011           # reset single spec to pending
nu dispatch.codex.nu reset                      # reset ALL (nuclear)
nu dispatch.codex.nu retry                      # re-queue all failed specs
nu dispatch.codex.nu review 011                 # print spec.md
nu dispatch.codex.nu list                       # table of all specs
nu dispatch.codex.nu lock                       # check active run lock

Launching panes (two-pane pattern)

# Pane A — run current wave
zellij --session wise-xylophone run --name "proj-wave1" --close-on-exit -- \
  nu -c "cd $HOME/PROJ && nu dispatch.codex.nu run --wave 1 --yes"

# Pane B — auto-chain to next wave when A finishes
zellij --session wise-xylophone run --name "proj-handoff" --close-on-exit -- \
  nu -c "cd $HOME/PROJ && nu dispatch.codex.nu watch-handoff --from-wave 1 --to-wave 2 --yes"

Spec table fields

Cooldown delays (rate limit pacing)

Batch pause of 2 min every 4 specs on wave 2+.

Ollama pre-planning

Runs qwen3-coder:480b-cloud on each spec.md before Codex touches it. Writes .auto-claude/specs/{spec}/plan.md with:

• Risk register (top 5 risks + mitigations)
• File order (what to create, in order)
• Crate audit (APIs to verify before writing)
• Gotchas (known pitfalls)

Codex receives both spec.md and plan.md as context → fewer wrong turns.

# Pre-plan all specs now (run while waiting for wave 1):
nu dispatch.codex.nu pre-plan

State file

.codex/dispatch-state.json
{"completed": ["001","002",...], "failed": [], "skipped": []}

Pane shell calls mark-done NNN or mark-failed NNN on exit. Signal files also written to /tmp/triz-done-NNN for cross-script coordination.

Rescue playbook

# Stale lock (process died without clearing):
rm .codex/dispatch-lock.json
nu dispatch.codex.nu retry

# Pane exited but mark-done never fired:
nu dispatch.codex.nu mark-done 011

# Spec produced bad output, redo it:
nu dispatch.codex.nu reset --spec 011
nu dispatch.codex.nu run --wave 2 --yes

# Rate limited mid-wave:
nu dispatch.codex.nu retry
nu dispatch.codex.nu run --wave 2 --yes

Bootstrap a new project

Copy the template and edit the specs [] table and constants:

cp ~/.claude/skills/dispatch-nu/dispatch.codex.nu.template /path/to/project/dispatch.codex.nu
# Edit: PROJECT_DIR, specs table (id/name/wave/cost/agent/mode)
# Seed completed: ["any-pre-done-specs"] in state
nu dispatch.codex.nu pre-plan   # plan everything via Ollama
nu dispatch.codex.nu plan       # review
nu dispatch.codex.nu run --wave 1 --yes

dub-specific notes (dispatch.codex.nu)

• Uses gpt-5.3-codex -c model_reasoning_effort=xhigh
• Wave 1 is strictly sequential (each spec is the next's input)
• Isolated specs need manual merge before next wave
• watch-handoff waits for wave N then auto-starts wave N+1

triz-specific notes

• 001+002 pre-seeded as completed in state
• Codex specs: 011, 017, 018, 019, 020 (pure Rust / TypeScript / config)
• AC specs: 003, 004, 005, 006, 007, 008, 009, 012, 013, 014, 016 (complex LLM/Nu)
• Wave 1: 011-core-commands via Codex (no LLM deps, deterministic Rust)

Verification contract (auto-injected by proveryay)

Every agent dispatched through this skill receives the following appended to its prompt automatically:

<see ~/.claude/skills/proveryay/CLAIMS-TEMPLATE.md — injected verbatim>

This means every agent MUST emit a <claims> block when it makes any deployment/state-change assertion. The proveryay SubagentStop hook will verify each claim in parallel and inject structured verdicts into the dispatcher's next-turn context.

If the agent's work is purely planning / research / narrative (no state changes), the agent may omit the <claims> block. Proveryay treats absence as silent (no noise).

Invocation example with template injection:

# Prompt injection: template auto-appended (idempotent)
let tpl = (open --raw ~/.claude/skills/proveryay/CLAIMS-TEMPLATE.md)
let prompt = if ($original_prompt | str contains "<claims>") { $original_prompt } else { $"($original_prompt)\n\n($tpl)" }

See:

• ~/.claude/skills/proveryay/SKILL.md — hook behavior
• ~/.claude/skills/proveryay/CLAIMS-TEMPLATE.md — canonical format
• ~/.claude/skills/fleet-eval/SKILL.md — probe taxonomy