parley-deck-cli 1.46.0

1.46.0 — 2026-08-21

A protocol-and-skill audit by six agents, and the fix-up cycles it took to close. 37 findings were
confirmed; 33 are fixed here, 4 remain deferred with recorded reasons. Nothing in this release is
a feature — it is enforcement catching up with what the protocol already said.

The defect class this release is mostly about

A printed rule binds only where enforcement lives. Nine of the fixes below are the same shape:
a rule stated in COOPERATION.md or in a prompt, enforced on one of two entry points, silently
absent from the other. The audit found it, then committed it twice more while fixing it.

Fixed — gates that accepted emptiness

• consensus draft accepted a blank round artifact, or one with no headings, as a filed round.
• consensus finalize closed an idea around an unwritten FINAL.md scaffold, and — through the
  manual path only — around a substantive FINAL belonging to a different idea, or one whose
  status: was not final. protocol.ValidateFinal is now the single gate for both the manual
  command and the driver; internal/driver's second section list and second section checker are
  deleted.
• consensus draft --review accepted review artifacts with no reviewed-commit and no
  ## Refutation attempts. The driver had validated these since the rule was introduced; the
  manual command never did. ValidateReviewArtifact moved to internal/protocol and now binds
  both.
• The implementation gate accepted any non-empty status — status: banana validated — and
  matched ## Summary of work as a substring, so a bare heading and a prose mention both passed.
  Now a closed vocabulary and a non-empty-section check. Measured before enforcing: zero live
  artifacts in this deck are newly rejected.
• The pipeline treated the first-step FINAL scaffold as a completed block, because the scaffold
  carries status: final from step one.

Fixed — quorum and roster

• The deliberation review quorum silently dropped the implementer's required signoff. §6 forbids
  the implementer authoring a review; §4.0 requires all participants to sign on the
  deliberation track. One list answered both questions. reviewConsensusVoters now splits them.
• The implementer's own signoff made every review consensus it appeared in malformed —
  measured across this deck: 24 consensuses flipped, two of them in flight.
• Standalone parley preflight probed every installed adapter family instead of the roster. On
  this deck it reported on seven adapter families including agy, which is not in the roster, and
  printed "Ready: no pending gates". parley run was already correct. §9.0 requires probing every
  rostered participant.

Fixed — freshness and metadata

• Two missing hashes compared equal, so a fresh deck was reported "in sync" whatever its
  protocol said — including an altered one. An equality test is evidence only when both sides exist.
• The gate that fix introduced could not be cleared: its own displayed --yes remedy had no
  branch for it, so a freshly initialized deck could never be reported ready. --yes now hashes the
  live protocol (and the packaged body when the installed skill exposes one), persists both, and
  states exactly what it compared. It never claims "in sync" without a packaged hash.
• parley init records the hash of the COOPERATION.md it just wrote, and deliberately does not
  invent a packaged hash it never computed.
• parley init used to leave **Workspace:** <workspace-name> and **Created:** <date> literally
  in the header — false provenance on the protocol's own first lines.

Fixed — things the protocol told you to do that did not work

• parley roster render, which COOPERATION.md instructs the bootstrap to run, broke the repo's
  own drift guard. The generated §2 table does not reproduce the anchor's hand-typed column
  padding and adds a State column. The anchors now match on column signature.
• masked-by-env was in the closed STATUS vocabulary with nothing able to emit it. roster set
  printed it once to stderr after a write; roster show had no path to it at all. A status nobody
  can observe is documentation of a behaviour that does not exist.
• The driver instructed its own FINAL drafter to produce an artifact its own gate rejects. The
  prompt is now generated from protocol.RequiredFinalSections, and a contract test builds a FINAL
  from nothing but the prompt's own text and feeds it to the gate.

Fixed — protocol text

COOPERATION.md §2 now documents all three roster-authority states; §12.12's dangling slug, the
Quickstart's missing §15/§10, §3's layout omissions and §11.B's self-contradicting branch-protection
advice are corrected. learn and preset list are visible in --help. 20 closed ideas carried a
stale non-terminal status:.

Deferred, with reasons

codex-1/F6 (needs a designed semantic signal for an adversarial alternative — substring inference
would not safely enforce §15.6), codex-1/F8 (a missing collapsed fast-track close path),
codex-1/F14 (needs per-knob precedence; applying standard defaults blindly would overwrite
explicit idea configuration), kimi-1/F1 (needs a manifest describing the installed core shape,
which includes payload absent from the source add-on manifest).

Known environment limitation

go test ./... exits 0 on an unsandboxed macOS host and 1 inside a Codex Seatbelt sandbox, failing
only TestDurableKillEndToEndRealProcess: seatbelt denies sysctl kern.boottime, so procctl