Skip to content

Protocol Action Packet Decision v0

Decision

Keep protocol_action_packet_v0 as the hot-path protocol simplification contract for quota should-run.

The opt-in TurnEnvelope may reconstruct this packet from its structured action contracts and omit the repeated summary only when field-level parity succeeds; the full decision continues to persist the compatibility packet.

The hot path remains deterministic and rule-only:

  • llm=no_api
  • primary actor: user or agent
  • user-action requirement
  • agent-action requirement
  • quiet-noop allowance
  • work lane
  • compact action label

Use the Codex CLI wrapper only as an explicit cold-path sidecar experiment. Do not run it during routine quota/status/heartbeat routing.

Defer direct LLM API wiring until a separate backend-comparison experiment proves a measurable gain over deterministic labels on payload size, user/agent-action clarity, and public-boundary safety.

Evidence

The decision is based on four public-safe slices:

  1. protocol_action_packet_v0 made quota should-run expose a compact rule-only summary while preserving the detailed guard payload.
  2. protocol_router_comparison_v0 compared advancement, user-action, and monitor-only scenarios and kept the minimum payload shrinkage above the acceptance floor without model calls.
  3. protocol_action_packet_codex_cli_wrapper_v0 proved the Codex CLI command envelope can be represented as a fake-contract smoke without invoking a model.
  4. The opt-in real Codex CLI probe ran once in an isolated fixture and produced a compact sidecar while leaving the default smoke path fake/no-model.

Operating Rule

quota should-run, dashboard status, and recurring heartbeat routing should consume protocol_action_packet_v0 directly. They should not call Codex CLI, direct LLM APIs, runner adapters, Docker/cloud environments, or paid compute.

Cold-path experiments may call Codex CLI only when the command is explicit, isolated, ephemeral, and sidecar-only. The sidecar may record a final compact summary, prompt length, return code, and stdout/stderr character counts; it must not persist raw stderr, raw session history, private traces, credentials, or local auth material.

Next Work

The protocol simplification spike is complete enough for the current meta lane. Future work should move back to the long-horizon benchmark program unless a concrete protocol regression appears.

The next benchmark-side step is the approved Terminal-Bench/Harbor execution environment readiness lane: check whether local Docker or an approved cloud execution environment is available, then attempt a single-task no-submit Harbor Codex pilot only after the environment and benchmark rules are clear.