跳转至

Codex Subagent Independent Context and Delegation Protocol

Core Capabilities

  • A subagent handles specialized work in an independent context, keeping intermediate exploration and large reads out of the main session.
  • Independent context does not inherit every parent fact automatically; the main agent must provide a minimum complete, unambiguous task packet.
  • Every subagent consumes model and tool resources. Multiple agents do not inherently save tokens.
  • Independent context is not permission isolation. System read-only status requires a read-only parent session or runtime evidence.

Minimum Delegation Packet

  1. Functional boundary, goals, and non-goals.
  2. Current phase, execution profile, risk, and stopping conditions.
  3. Baseline, HEAD, packet hash, and freshness.
  4. Unique responsibility, file/symbol/call-chain scope, and exclusions.
  5. Existing validation, failures, unverified items, and evidence index.
  6. Permitted/prohibited operations and isolation level.
  7. Approved dispatch profile, permit reference, reserved units, and result schema.

Return to the Main Session

  • Checked scope.
  • Consolidated root-cause findings.
  • Evidence levels, unverified items, and necessary validation.
  • Approved dispatch profile, reserved units, and isolation state.
  • Brief conclusion.

Do not return the full parent session, complete diff, long raw logs, every file, or internal reasoning.

Dispatch Profiles

luna-low -> luna-medium -> terra-medium -> terra-high
  • Luna: read-intensive, bounded, repetitive, structured subtasks.
  • Terra Medium: business semantics, multi-file logic, and ordinary specialist judgment.
  • Terra High: transactions, concurrency, security, irreversible migration, core state machines, and blocking conflicts.
  • Automatic flows never exceed Terra High and never use Sol, xhigh, max, or ultra.
  • Exact model requests are validated only transiently before dispatch; Agents do not read, infer, or report host runtime model identity.

Harness Principle

V7.4 follows “a map, not a thousand-page manual”: global rules define boundaries and routing; Skills provide on-demand capability; scripts maintain state, evidence, idempotency, and gates; repositories and external memory retain verifiable facts.