WP 6.4 — Resume/cancel semantics + override recording

[AJ-EthereaLogic-ai]

Work package: WBS 6.4 · Phase: 6.0 — Promotion, Consensus & Reporting · Traces: FR-15, US-17, AC-17.1, AC-17.2 · Depends on: WP 6.3

Problem statement

Recovery must never be hidden behind opaque automation (principle P6). Every multi-agent operation the runtime performs — phase runs, gate runs, promotions, consensus rounds — needs cancel / retry / resume / resolve, and every human override must be recorded with actor, reason, and effect in the artifact's audit trail (FR-15, US-17). Because the evidence tree is append-only (NFR-5), an interrupted run is resumable from it without re-running completed attempts, and a cancelled run keeps its partial evidence.

Requested change

Add resume/cancel semantics and override recording as cross-cutting updates to the multi-agent commands under skills/ and their shared helpers under scripts/:

• interrupted-run recovery derived from the append-only evidence tree;
• cancel / retry / resume / resolve on every multi-agent operation;
• an override record (actor / reason / effect) written to the audit trail for every human override.

Acceptance criteria

• Every multi-agent operation exposes cancel, retry, resume, and resolve controls (FR-15, US-17) — enumerated so coverage is checkable rather than asserted: each of the six phase runs (frame, assess, innovate, model, implement, track), the standalone gate run, both promotion boundaries, and the consensus round.
• An interrupted run resumes from its evidence tree without re-running already-completed attempts (AC-17.1).
• A cancelled run retains its partial evidence append-only — nothing in the evidence tree is deleted (AC-17.1, NFR-5).
• Resume derives what is already complete from the append-only attempt directories, not from a mutable status field (AC-15.2, AC-17.1).
• Every human override — dissent resolution, gate exemption, forced rollback — is recorded with actor, reason, and resulting effect and appears in the artifact's audit trail (AC-17.2, NFR-5).

Non-goals

• Adding new agents or validators, or changing any deterministic verdict logic (WP 5.x / 6.1 / 6.2).
• Re-running completed attempts on resume (the append-only tree is the source of completion state).
• Any new blocking behavior outside the promotion boundaries (D-2).

Constraints

• Evidence is append-only with UTC timestamps; a cancelled or resumed run never deletes prior evidence (NFR-5).
• Overrides are recorded with actor / reason / effect (P6, AC-17.2); no --force / --no-verify, explicit-path staging only (NFR-4).

File hints

• Edit/create: every multi-agent surface — the six phase commands skills/frame/, skills/assess/, skills/innovate/, skills/model/, skills/implement/, skills/track/; skills/gate/ and skills/promote/; the consensus agents under agents/ (cadence-critic, cadence-advocate, cadence-grader); and the shared resume/evidence helpers under scripts/.
• Read: docs/design/CADENCE_AUTOMATION_PROJECT_PLAN_WBS.md (FR-15, NFR-5, P6), docs/design/CADENCE_AUTOMATION_USER_STORIES.md (US-17, AC-17.1–AC-17.2, AC-15.2), skills/cadence-method/references/artifact-layout.md.

ADWS run

• source_type: github_issue · output_mode: pr · target_branch: main
• allowed_paths: [ "skills/", "agents/", "scripts/" ]
• test_policy: required · falsifiability: true · task_size: medium

---

Packaging surface updated 2026-08-06 UTC (Q5). The ten practitioner actions ship as per-action skills, skills/<action>/SKILL.md, not commands/*.md. Decided by the operator ahead of the WP 1.4 freeze note, against Q5's pending default, because the live documentation directs new plugins to skills/ and no action was built yet. The invocation string /cadence:<action> is unchanged. See §7 of the authority document and its v1.13 Revision Record row.

---

Reconciled with the authority document and the WP 1.3 spec sheet 2026-08-06 UTC.

• allowed_paths widened from three skills to skills/, plus agents/. The criteria require these controls on every multi-agent operation — phase runs and consensus rounds included — while the path list reached only promote, gate, and implement. As scoped, the package could not touch frame, assess, innovate, model, track, or the Critic/Advocate/Grader agents, so it could not satisfy its own first criterion.
• The operations are enumerated rather than left as a category, so WP 7.4 can walk the coverage instead of taking it on trust (P2).