Feature request: mission-entropy detection and bounded sub-agent delegation #21858

What variant of Codex are you using?

Codex desktop app on macOS, across multiple local repositories and customer-facing release workflows. This also applies to CLI and IDE workflows where long-running missions span implementation, QA, release, and documentation.

What feature would you like to see?

Codex should detect mission entropy in long-running work and proactively propose, or optionally auto-spawn, bounded sub-agents while the parent thread remains mission control.

I use Codex for real operational work across multiple local repos, including:

• building and packaging macOS and Windows desktop apps
• testing installed customer artifacts such as /Applications and Windows install locations
• creating GitHub releases and preparing storefront fulfillment installers
• updating a company website
• verifying product copy against actual release artifacts
• preserving local handoff and mission notes for continuity across sessions

Long-running Codex threads can become muddy when build failures, QA defects, release work, storefront work, user corrections, and handoff notes accumulate in one context. Codex already supports multiple agents, but today the user must decide when to spawn them and how to scope them. In practice, the moment delegation is most needed is often when the mission is already confusing.

Suggested trigger behavior

• Default: propose a delegation plan first.
• Optional workspace setting: auto-spawn bounded sub-agents when entropy is high.
• Per-workspace override: ask first / auto-spawn / never auto-spawn.

Suggested mission-entropy signals

• repeated failed attempts at the same task
• long-running work without evidence of progress
• conflicting source-of-truth claims, such as docs vs code vs release artifacts vs storefront
• too many unresolved branches or todos
• scope drift before acceptance criteria are closed
• repeated user corrections
• mixing implementation, QA, release, storefront, marketing, and memory or handoff work in one thread

Suggested defaults

• Rolling window: last 20 turns or last 30 minutes, whichever captures more signal.
• Soft trigger: entropy score at or above 70/100.
• Hard triggers:
  • 3 repeated failures with the same root cause
  • 2 explicit user stop/reset/muddy corrections
  • contradiction in source truth for a release, storefront, or customer artifact

Delegation model

The parent agent should remain mission control. It owns intent, constraints, acceptance criteria, synthesis, contradiction resolution, and durable handoff.

Sub-agents should be bounded:

• narrow task
• explicit success criteria
• time or step budget
• isolated worktree when applicable
• read-only by default unless implementation is explicitly needed
• no spending money or changing customer-facing systems unless already authorized
• if budget is hit, report attempted actions, evidence found, blocker, and recommended next step instead of looping

Suggested spawn defaults:

• Max concurrent sub-agents: 3
• Max total sub-agent spawns per mission phase: 8
• Default sub-agent budget: 10 to 20 minutes or a fixed tool-action cap

Required sub-agent report

Each sub-agent should report:

• assigned task
• scope boundaries
• evidence inspected, with concrete references
• findings, separating observations from inferences
• confidence level and why
• files changed, if any, with diff or commit references
• blockers
• recommended next action

Evidence examples: command outputs, log paths, file paths, git commits/tags/releases, PR or diff references, screenshots when relevant, URLs checked, and checksums for release artifacts.

Durable continuity

When appropriate, Codex should update canonical handoff files:

• FACTS.md: verified facts only
• DECISIONS.md: decisions plus rationale and evidence pointer
• NEXT_ACTIONS.md: prioritized next steps
• CONFLICTS.md: unresolved contradictions and resolution plan
• OUTCOME.md: what was actually completed

Only evidence-backed facts should enter durable handoff or memory. Hypotheses, failed attempts, emotional context, and stale assumptions should not become durable operating truth.

Additional information

Example workflow:

Parent mission: certify and release three desktop apps.

Entropy trigger: Codex detects repeated installer/test failures, conflicting local vs GitHub artifact versions, and user corrections about source truth.

Codex proposes:

1. Sub-agent A: verify GitHub release assets, versions, and checksums.
2. Sub-agent B: inspect installed macOS apps from /Applications and collect logs/screenshots.
3. Sub-agent C: compare storefront fulfillment files and copy against release artifacts.

Each sub-agent reports evidence and findings. The parent thread synthesizes, resolves contradictions, updates FACTS.md and NEXT_ACTIONS.md, and continues from verified state.

The failure mode in long-running Codex operational work is often not lack of intelligence. It is mission entropy. Codex should automatically manage complexity, delegation, verification, and durable handoffs so the user sets intent and boundaries while Codex keeps execution reliable.