ethanreedlabs/sulde/dispatch-task

Dispatch Task

Default Codex dispatch contains only the task instruction and acceptance requirements. Keep the target session's current model and reasoning configuration without showing selectors, model names, capability checks, or a “no switch needed” preamble. Unknown model state does not block dispatch.

Resolve the target

1. Read capability_tier from the task: light, balanced, or deep.
2. Resolve the provider of the target execution session, not the coordinator's machine:

• use the user's explicit Claude Code or Codex statement first;
• when sending to the current session, use the current host;
• when unknown, ask which host will execute the task and emit no dispatch block.

1. Do not inspect the current model/reasoning for ordinary Codex dispatch. Only collect that

information if the user explicitly asks for model advice; omit unknown current-state arguments.

Installing another CLI is not evidence that it owns the target session.

Dispatch an outcome, not a test round

For implementation/repair handoffs, include a bounded continuous-execution agreement in the task: frozen scope and acceptance, baseline/candidate identities, expected impact, evidence path, budget and genuine stop conditions. Read spec/task-authoring.md in the source tree (or runtime/spec/task-authoring.md in a Codex plugin) for the full agreement when authoring or amending that task.

The executor continues authorized fixture repair → defect/fix verification → actual entry checks → affected regression without waiting for a new “continue” after each stage. Updates are not approval gates. For fault-injection work, valid normal controls precede injection; use identical assertions on old and candidate code. Distinguish fixture failures, old defects and candidate regressions. Do not impose injection on documentation-only tasks or run full suites by default.

Two attempts without new evidence trigger a change of diagnostic method, not another identical run or automatic handoff. Pause only at the agreed stop line, budget limit, genuine blocker, or a needed authority/scope decision. Preserve explicit audit-only/tests-only restrictions. A continuation must explicitly amend an earlier “stop after this test” instruction before proceeding beyond it.

Request one consolidated independent review once scoped evidence is ready; collect related blockers together rather than issuing one micro-task per finding. Do not auto-accept, widen authority, or add unrelated discoveries to completion. If an old renderer omits this agreement, append it to the task prose rather than requiring a runtime upgrade merely to dispatch.

Render the instruction

Use the stable launcher:

SULDE_MODEL_DISPATCH="${SULDE_HOME:-$HOME/.sulde}/bin/model-dispatch"
"$SULDE_MODEL_DISPATCH" \
  --provider <claude|codex> \
  --tier <light|balanced|deep> \
  --task <task-path>

Do not use --fresh-session for an amendment or “继续原任务”.

Paste the renderer's stdout verbatim beneath “给 Dev 发送”, then append task-specific prose. An older installed renderer may still emit model controls. For ordinary Codex dispatch, do not forward that legacy preamble: write the task path/content directly and preserve current settings. Do not modify immutable caches or require an installation merely to format a task.

Explicit model advice only

Only when the user explicitly requests model advice, add --model-advice and the available --current-model, --current-tier, and --current-effort arguments. --required-effort is valid only in this advice mode. A task's capability_tier alone is not such a request.

The current default mapping is:

If that exact model is unavailable in the target account, select the same tier or higher from the target Codex /model picker. Preserve a current model/reasoning pair that already meets the floor. xhigh and max are reasoning levels, not model names.

Only advice-mode Codex output may use /model and /reasoning. Codex 派单禁止出现 Claude model families, Claude thinking phrases, /mode, /assign, /clear, or ralph-loop.

Verify

Before sending, confirm:

• the task file still stores only capability_tier, never a provider model name;
• the rendered provider equals the target session provider;
• ordinary Codex output contains no model/reasoning advice or readiness preamble;
• explicit advice output matches the target provider;
• a continuing task stays in the same session unless the user explicitly requested a fresh one.
• a repair task covers continuous execution through its review point, with bounded

retries and explicit stop conditions rather than user-mediated per-test rounds;

• claims separate old defects, candidate repairs and invalid tests, and reference

the tested inputs; unrelated follow-ups do not silently expand acceptance.

If advice-mode rendering is unavailable, report that limitation without inventing selector commands. Ordinary task-only dispatch can proceed without the renderer. This skill controls interactive handoff text, not model selection by managed background Agent runtimes.