LoopSkill v3.3.0

LoopSkill v3.3.0

LoopSkill v3.3.0 replaces the session-style State-Writer in new schema-v3
loops with a deterministic MCP State Gateway. It atomically prepares canonical
routes and evidence, binds one host-cooperative App send observation to the
original outbox, accepts only target-staged reports, recovers lost stdout on
that same outbox, captures binary Git diffs without model-transcribed patch
bytes, and initializes fresh successors without modifying terminal
predecessors.

Highlights

• state_gateway is the sole canonical writer for new schema-v3 loops.
• Atomic PREPARE_ROUTE, RECORD_ROUTE_SENT, ACK_ROUTE_RESULT, and
  original-outbox REPORT_RECOVERY.
• Current artifact + current dispatch + formal PASS report evidence binding;
  stale, wrong-thread, wrong-digest, BLOCKED, or unattested reports fail closed.
• Runtime-owned binary-safe CAPTURE_COMPLETE_DIFF with reverse-apply proof.
• Deterministic transport degradation/pause handling, successor handoff, and
  derived loop-efficiency metrics.
• Finalization remains nonterminal after PREPARE_FINALIZATION; only a real
  PAUSED heartbeat readback followed by ACK_FINALIZATION yields
  FINALIZATION_ACKED.
• Schema v1/v2 compatibility remains; v2-to-v3 migration is explicit at a
  paused safe point.
• Native Goal generation recovery remains intentionally unavailable.

Release identity and local authority

• Protected-main commit:
  f31456d78f4cb9c13805b99a8ee0c43fd155d682.
• Local complete gate: 650 tests PASS, 80.24% branch coverage, generator fuzz
  5000 PASS, state fuzz 5000 PASS, Skill/SPEC/schema/installer/risk/secret
  scans PASS.
• Independent Reviewer: P0=0, P1=0, P2=0.
• Same-SHA install receipt: version 3.3.0, configuration readback true,
  source/install drift empty, source and installed manifest digests equal.

Real Codex App evidence boundary

The required product-capability aggregate is PASS using the immutable A/B/C
receipts plus independent positive fields in D1:

• A: task/heartbeat registration and long automation-ID boundary.
• B: cross-task send/verify/stage/ACK, strict report binding, and original
  outbox report recovery without a second product dispatch.
• C: binary diff capture and bounded transport pause with PAUSED readback.
• D1: FINALIZATION_ACKED, LOOP_COMPLETE, PAUSED heartbeat readback,
  SUCCESSOR_INITIALIZED, fresh successor controller, empty successor
  leases/outboxes, and byte-identical predecessor state/control-plane digests.

D1-D7 remain individually and immutably recorded as FAIL. Their harness or
scenario errors are not rewritten as PASS: invalid status/codec arguments,
hand-transcribed fields, temporal/order/diff-reference mistakes, incompatible
scenario prerequisites, an empty heartbeat parameter object, and a task-root
mismatch. These rejected calls had zero canonical side effects. No D1-D7
evidence shows an invalid request being accepted, canonical pollution, a wrong
report being ACKed, send observation creating PASS, premature terminal state,
or successor mutation of its predecessor.

The aggregate release decision is a product-capability decision under the
user-authorized host-cooperative evidence model; it is not a claim that D1-D7
all passed or that the harness quality was clean.

GitHub CI is not used or claimed as the release authority for this release.
The authority is the exact-main local gate, same-SHA installation identity,
independent review, and the real App product-capability aggregate described
above.