codex-dispatch

name

codex-dispatch

description

Mandatory dispatch preflight before any subagent delegation. Use before the first child-work call in every task, and when starting Codex work that needs an explicit Direct vs Single/Parallel/Sequential choice: code edits, debugging, reviews, repo exploration, multi-file changes, ambiguous scope, high-risk judgment, or requests mentioning agents, delegation, parallel work, model choice, cost, token use, or over-delegation.

Codex Dispatch

Mandatory preflight

Before the first call that creates or assigns substantive work to any child agent in a user task, the root MUST load this skill and emit the visible dispatch decision below. This is a hard gate: without the decision, do not call spawn_agent, followup_task, or send a scope-expanding task to an existing child.

One decision covers later status checks, waits, messages within the approved scope, interrupts, and result collection for that task. A materially new user request or a new child scope requires a new preflight before more delegation. If this skill is unavailable, stay Direct or tell the user delegation is blocked.

If a child was started before the preflight, stop further delegation, disclose the miss, load this skill, emit the decision, and audit every existing child against the chosen mode, scope, and three-child maximum. Keep conforming children; interrupt or rescope non-conforming children. Do not restart conforming work merely to simulate compliance.

Core rule

Make one visible dispatch decision before substantial work and always before child delegation. Loading this skill never means “spawn”; it means decide.

Root owns requirements, shared context, file-conflict control, final synthesis, and the user-facing answer. Delegate only when a child result is independently useful enough to justify coordination overhead.

When the root model is strong (for example Sol-class) or other skills already push toward subagents, be stricter, not looser. Children are for context isolation, true parallelism, bounded implementation, hard debugging, or independent review—not confidence theater.

Required decision

Before running repo inspection, editing files, or debugging deeply, emit exactly one decision:

Dispatch: Direct | Single | Parallel | Sequential
Reason: <why this mode is sufficient>
Children: <none, or role -> bounded result>

No visible decision means the skill has not been used and child delegation is forbidden.

Bias

Start at Direct. Stay Direct unless at least one hard delegation gate below is met.

Choose Direct when:

• the only reason to spawn is task size, confidence, model strength, or habit;
• files are tightly coupled or likely to overlap;
• the user has not authorized subagents and active runtime policy requires explicit authorization;
• the root can finish faster than it can coordinate.

Hard delegation gates

Delegate only when one of these is true:

Gate	Delegate when	Prefer
Context isolation	A large read-only search, inventory, extraction, or comparison would bloat the root context before implementation.	luna-explorer
True parallelism	Independent areas can be explored or verified without overlapping files or shared state.	luna-explorer / terra-worker
Bounded implementation	One clear change has acceptance criteria, expected files, and a matching test or verification command.	terra-worker
Hard debugging	Reproduction or root cause is uncertain enough that a focused investigator can return a concrete failing case or fix plan.	terra-debugger
Independent review	The result is high-risk, architectural, security-sensitive, or likely to suffer from root blind spots.	sol-reviewer

Modes

Outcome	Selected children
Direct	No hard gate is met, or coordination would cost more than it saves.
Single	Exactly one bounded artifact.
Parallel	Two or three independent, non-overlapping artifacts.
Sequential	Two or three artifacts where a later child depends on an earlier child.

Hard maximum: three child tasks per routing decision. Root synthesis does not count. If more branches exist, combine them into three bounded scopes or keep excess work with root.

Never parallelize agents writing overlapping files. Keep delegation root-to-child; children must not spawn agents. Respect active runtime policy; this skill requires a decision, not a spawn.

Choose agents

Agent	Route here
luna-explorer	Clear read-only exploration, extraction, inventory, classification, or summaries.
terra-worker	Everyday implementation, writing, straightforward debugging, and tests with clear acceptance criteria.
terra-debugger	Difficult reproduction, root-cause analysis, cross-module reasoning, or high-effort implementation.
sol-reviewer	Independent review of ambiguous, architectural, security-sensitive, or high-value results.

Cost preference for compatible work: Luna Low -> Terra Medium -> Terra High -> Sol High. Cost never overrides role semantics. Max and Ultra require explicit user choice.

Child contract

For every child, provide: goal, scope, completion criteria, allowed mutations, and expected return format. State the outcome, not tactics. Wait for required results, compare cited evidence, and synthesize one answer; never concatenate child responses.

Reuse an existing agent with follow-up instructions instead of spawning a duplicate for the same role and scope. Ask for concise findings, verification, and remaining risk rather than raw logs.

Direct is not a shortcut around verification. Use the task's domain skills, tests, and evidence requirements normally.

Common mistakes

• Spawning agents merely because a task is large.
• Spawning agents because a strong model is available.
• Treating “I used the skill” as true without a visible dispatch note.
• Starting or assigning child work before the mandatory preflight.
• Re-running conforming child work after a late preflight instead of auditing it in place.
• Using Sol for clear bounded work that Luna or Terra can complete.
• Running dependent stages in parallel.
• Giving two writers ownership of the same files.
• Letting children expand scope or spawn grandchildren.
• Reporting success without waiting for required verification.