Skip to content

Kbediako/evergreen-codex-skills / skills/native-subagents-first

native-subagents-first

Verified

Coordinate native Codex subagents for requested parallel work or independent execution and review scopes. Use for ownership, worker selection, and lifecycle management.

FileKindSize
SKILL.mdskill md5.5 KB
agents/openai.yamlconfig243 B
references/astra-routing.mdmarkdown2.7 KB
references/briefs-and-ledger.mdmarkdown5.3 KB
references/read-only-boundaries.mdmarkdown2.1 KB

references/briefs-and-ledger.md

View on GitHub
# Briefs and coordination ledger

## Child brief

Use a compact, self-contained brief. The fields below are a menu, not a form to fill for every assignment; include what the child needs to act and what the parent needs to verify:

```text
Stream: <short label>. <read-only/write-enabled>.

Objective: <one concrete outcome>
Scope: <exact paths, modules, artifacts, or questions>
Scope key: <milestone:objective:paths:mode>
Coupling: <independent scope or one coupled outcome owned end to end>
Allowed changes: <none or exact ownership>
Worker condition: <role and effort; fresh or bounded inherited context>
Expected budget: <small/normal/deep; checkpoint; hard decision point>
Authority and safety: <task-local permissions and prohibitions>
Contradiction rule: if evidence falsifies the requested mechanism, report evidence and a replacement hypothesis without exceeding authority
Dependencies: <none or named peer and required handoff>
Subagent boundary: Complete this directly. Do not spawn unless this brief explicitly grants it.
Nested run authority: <none, or exact reserved scope granted by the parent>
Acceptance:
- <observable criterion>
- <observable criterion>
- <observable criterion>

Return:
Outcome: done/partial/blocked
Changes: exact files or none
Validation: commands/checks and results
Findings: prioritized bullets
Evidence: paths, line refs, artifacts, or rollout ids
Peer handoffs: messages sent/received or none
Open questions: blockers only
```

For a write-enabled child, add:

```text
You are not alone in the workspace. Do not revert unrelated edits; adapt to changes made by others.
```

Do not ask every child for a plan when the assignment is already bounded. Do not send the intended answer, hidden diagnosis, or another child's conclusions unless the dependency requires them.

## Child-run accounting

Before the initial batch, and after resume or compaction, record:

```text
stable milestone ID = M
independent scopes = N
initial cumulative child-run cap = C
cumulative child runs used = U
reserved nested runs = R
remaining unreserved child-run budget = B = C - U - R
configured/observed free child slots = S
fanout now = min(N, B, S)
queued scopes = N - fanout now
```

Debit every `spawn_agent` and `followup_task` at any depth, including retries, replacements, failed calls, and interrupted runs. Reserve nested allowance before granting it; each nested spawn converts one reservation into one used run. Release unused reservations only after reconciliation.

`send_message` is free only for an in-brief correction, handoff, or already-requested evidence/checkpoint. A new objective, ownership boundary, deliverable, or acceptance criterion requires a budgeted child run.

## Ledger

Keep the ledger small enough to survive compaction:

| Field | Record |
| --- | --- |
| Milestone and terminal synthesis point | Stable ID, name, and stopping condition |
| Acceptance | Observable criteria for the assigned outcome |
| Child-run budget | Initial cap, used by tool/depth, nested reserved, remaining |
| Launch batches | Runs and parent delta unlocking each later batch |
| Capacity | Configured limit, active children, free slots |
| Scope keys | Child, mode, ownership, status |
| Evidence received | Final answer/artifact identity and parent verification |
| Rejected paths | Candidate and reason |
| Progress delta | Verified change, proof, evidence, decision, or metric; unused or batch unlocked |
| Next action | Parent action, wait, queued scope, or handoff |

Update it after any spawn/follow-up at any depth, reservation change, meaningful message, final answer, interruption, progress delta, or scope change. Do not narrate unchanged wait windows or snapshots.

## Peer handoffs

Use direct peer messaging when a child discovers a dependency another active child needs. Send only:

```text
Dependency: <what changed or was discovered>
Evidence: <path, line, artifact, or check>
Impact on your scope: <specific consequence>
Requested action: <one bounded response>
```

Require the sender to note the handoff in its final answer. The parent remains responsible for resolving disagreement and verifying the integrated result.

## Progress before another batch

Runs launched from the same parent state form a batch. The first batch needs available budget. Before a later batch, including a follow-up, retry, replacement, or nested run, record a new parent-owned result or decision from the previous batch: an integrated artifact, passed check, closed decision, verified evidence, or advanced goal metric. Reading a child answer counts only after verification or integration.

That result unlocks one declared batch and does not replenish the cumulative cap. Check both progress and unreserved budget. Without them, consolidate or complete authorized work in the parent only when the acceptance criteria permit it. Independent-review, controlled-trial, and delegate-only requirements still apply. Preserve the milestone, cap, used runs, reservations, and last verified result across steering and compaction.

## Baselines

For writes, record the relevant starting revision and existing changes within the assigned scope. For advisory read-only work, use the bounded manifest and index procedure in [read-only boundaries](read-only-boundaries.md). That reference owns link handling, ignored files, and before/after comparison; a Git status check alone does not prove the read-only invariant.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall