Skip to content

EXECUTION_MODES.md

Latest commit

 

History

History
80 lines (50 loc) · 4.38 KB

EXECUTION_MODES.md

File metadata and controls

80 lines (50 loc) · 4.38 KB

Execution modes — JAE v2.1.x

This document is normative for how the skill may combine governance (eight phases, scoring gates, consent tiers) with host execution (Claude / CoWork / Manus) without drifting.

Machine-readable: rules.json → execution_modes.
Skill behavior: SKILL.md — Execution mode gate and Mode 2 re-anchor.

1) Definitions

Key Name Default
agent_supported Agent-supported (Mode 1) Yes — used unless the user opts into Mode 2 on a supported host
cowork_autonomous CoWork autonomous (Mode 2) Opt-in — only on Claude CoWork when the user explicitly requests it

Session entry modes (unchanged): Mode A (discovery) vs Mode B (direct apply) are orthogonal to execution mode. You still run Mode A/B detection as today; execution mode only affects how aggressively the host may drive multi-step work inside the same phase law.

2) Mode 1 — agent_supported (default)

  • Full A0 → Phase 0–7 path with scoring gate every phase and Tier 2/3 consent exactly as in automation-registry.json.
  • CoWork: parallel Tier-1 work only where rules.json and SKILL.md allow; irreversible work on the main thread.
  • Claude.ai / Manus: sequential governance as already documented in platform-capabilities.md.

3) Mode 2 — cowork_autonomous (opt-in, CoWork only)

Intent: Use CoWork’s ability to run larger multi-step host execution where the product allows, without giving up JAE’s workflow law.

Non-negotiable (anti-drift):

  • Governance is always the agent-supported phase model — the eight phases, invariants, and registry scope remain the source of truth.
  • The skill does not “skip to submission” or collapse phases. Any fast host execution must be chunked and followed by a re-anchor step (A15/A16) that verifies alignment with the current phase and checklist.
  • Irreversible actions (submit, send, fill, etc.): default remains unchanged — same Tier 3/2 requirements unless a future, explicit, versioned policy says otherwise.
  • If the user’s request, tool output, or host behavior conflicts with phase scope, stop, return to a normal agent-supported checkpoint, and reconcile.

Opt-in phrases (one is enough, exact match):

  • JAE MODE: COWORK_AUTONOMOUS
  • I opt in to JAE CoWork autonomous mode with phase governance

To leave Mode 2 and return to Mode 1 in-session:

  • JAE MODE: AGENT_SUPPORTED

If the user never sends an opt-in phrase, remain in Mode 1.

4) Per-chunk loop (Mode 2)

For each autonomous chunk (see A14 in automation-registry.json):

  1. Pre-chunk: State current phase, objective, and allowed tools/scope for this chunk.
  2. Run chunk (host execution as permitted, with A14 consent if required by tier).
  3. Post-chunk re-anchor (A15): Confirm phase status, checklist row, invariants, and that no phase boundary was crossed without a completed gate.
  4. Drift check (A16): If ambiguity, tool conflict, or scope creep — pause and return to full agent-supported presentation for that step.

5) What Mode 2 does not mean

  • Not permission to add undeclared automations or “ghost” Axx.
  • Not permission to run Tier 2/3 on subagents.
  • Not a guarantee that every host feature (e.g. computer use, scheduling UI) is available on the user’s plan or region — see platform-capabilities.md and the vendor’s pages.

6) References (external, read-only for wording discipline)

JAE does not claim these pages describe this repo; they inform responsible wording.