v5.6.0 — Fable→Opus + reviewer context wiring (incl. v5.5.0 hybrid Workflow acceleration)

[egerev]

Summary

Hybrid (not a rewrite): the markdown DAG and Agent-based dispatch stay the default everywhere and the only path on Codex runtime. Two opt-in saved workflows accelerate the hottest Phase 2 fan-outs, and Phase 2 launch now suggests a /goal watchdog.

Saved workflows (deployed to ~/.claude/workflows/ via the v5.4.0 checksum sync)

• /superflow-review — unified review fan-out: product + technical reviewers in parallel; fenced-JSON verdicts parsed fail-closed in the script; pass computed in code, not model memory. workdir arg makes codex review the sprint worktree checkout (not the repo root); product: false handles light-governance single-reviewer cells. Provider value: workflow-review.
• /superflow-wave — implementation-only parallel wave per the v5.4.0 dispatch rule; orchestrator keeps review → docs → PAR → ship per sprint.

references/workflow-orchestration.md — single Workflow authority

Documented API facts with official-doc citations, opt-in policy, permission gates by mode, limits, runtime conventions verified by live smoke runs (top-level script body, parallel(thunks), meta pure literal, args may arrive as a JSON string → normalization required), an UNDOCUMENTED-API warning box, the /goal watchdog section, and the Codex fallback chain.

Integration

• context.use_workflows opt-in captured at Phase 1 plan approval (state + Autonomy Charter), with an honest permission-mode heads-up.
• Phase 2 launch prints a ready-to-paste /goal suggestion (user-only command — documented as such; subagent goal-direction remains charter injection).
• Codex runtime untouched: no Workflow tool, no /goal — existing spawn_agent paths.

Test plan

• Live smoke run, fail-closed path: /superflow-review with empty args → instant {pass: false, error: "missing required args: …"}, 0 agents, 12 ms
• Live smoke run, full pipeline: /superflow-wave with a fake sprint → 1 agent dispatched, fenced JSON parsed, position-bound result returned exactly per contract
• Live probe: confirmed args arrive as a JSON string → defensive normalization shipped in both scripts
• tools/verify-phase2-dag.sh 33/33 PASS; jq schema checks; llms.txt line counts swept
• Dual review (technical + coherence) → fixes → scoped re-review APPROVE
• CI green on this PR

🤖 Generated with Claude Code

[egerev]

v5.6.0 added on top of this branch

Fable access blocked → Opus everywhere

• Deep judgment roles (deep-spec/code/product-reviewer, deep-analyst): model: fable → model: opus; depth now differentiated by effort (deep=max, standard=high). Implementers stay on Sonnet.
• model_profile selection removed (Phase 1 Step 2c, state schema, charter, plan recap, all dispatch docs).
• Every Claude secondary pinned to claude-opus-4-8; stale "inherits Fable" warnings corrected.

Reviewer context wiring

• code-reviewer gains <spec_or_plan> + <autonomy_charter> slots; product-reviewer gains <product_brief> + <autonomy_charter> — the plan-completeness/charter checks finally have structured inputs.
• /superflow-review takes optional spec_path/plan_path/brief_path, wired into both lenses; missing inputs flagged unverifiable.
• review-unified.md dispatch contract made explicit.

⚠️ One follow-up commit (CI forbidden-token gate: add claude-fable-5 guard + fix pin message) is held locally — pushing a .github/workflows/ change needs the workflow OAuth scope. Will push once granted.

[egerev]

✅ Follow-up c255f79 landed — CI forbidden-token gate now also blocks the Fable model pin and the message reflects the claude-opus-4-8 pin. v5.6.0 fully on the branch.