feat(review): headless JSON verbs + guided assumption-review console

[pofallon]

Implements specs/053-assumption-review-console (all 27 tasks).

Expose the assumption-review lifecycle as headless --json verbs, and add a packaged Claude Code skill that drives them — so clearing the review queue no longer requires a human sitting at a bare terminal.

What's here

US1 — headless JSON verbs. Every review-lifecycle verb is invocable with --json:

Verb	Command
review.list	review --list [--status|--spec|--severity]...
review.answer / review.waive	review <id> --answer/--waive
review.bulk-waive	review --spec <s> --waive <reason>
reconcile.run / reconcile.dry-run	reconcile [--dry-run]
land.status	land --status (new read-only frontier query)
land.run	land

They share one envelope and error-kind registry (cli/json_output.py), and one canonical entry-row projection (assumptions/serialize.py) used verbatim by both review --list and the land report, so the two surfaces can't drift.

US2 — guided console. A packaged maverick-review skill, installed by maverick init and removed by maverick uninstall. It sweeps the open queue one entry at a time via AskUserQuestion, applies each decision immediately through the JSON verbs, and never touches jj/git/bd directly.

US3 — batched flow. One reconcile after the sweep (not per answer), then a frontier report, then landing offered only on explicit human confirmation.

maverick review without --json is unchanged (FR-018) — it remains the fallback for humans without Claude Code.

Review fixes folded in

A code review of the above surfaced 15 findings, all fixed here. The load-bearing ones:

• JSON output was being corrupted. write_json_document went through a Rich Console with emoji substitution left on, so any :name: run inside a string value was rewritten to a unicode emoji. Agent-authored questions and answers hit this routinely, and review --answer "<adopted_answer>" then wrote the mutated text back into the ledger.
• A committed write could be reported as a failure. The post-write projection re-read sat inside json_error_handler, so a transient bd hiccup after the ledger write turned into ok: false, kind: bd-unavailable — telling the caller a recorded decision wasn't recorded. The re-read now fails soft (degraded: true).
• An unreadable ledger looked like a clear frontier. The gate degrades open and materializes zero entries, making frontier_clear trivially true. Added a top-level degraded flag; the skill and contracts now require frontier_clear && !degraded before offering to land.
• bd-unavailable was reported as validation on two review paths, which the skill's preflight branches on.
• WorkflowError.reason_code replaces prose substring-matching for precondition classification, so rewording a message can't silently reclassify it as internal.

Plus: land.run's nothing-to-land return now emits the full result shape; heuristic curation reports absorb_ran/squashed_count (absorb rewrites history without squashing anything); a new cli/commands/land_gate.py removes a private-surface import cycle between land and land --status and takes land.py back under the module-size guardrail; the skill install path is shared and written atomically; bd-readiness is one shared predicate.

Pre-existing bug found along the way: nothing ever set ctx.obj["verbosity"] — the root group sets ctx.obj["verbose"] — so verbose progress rendering was dead for every workflow, not just this feature's paths. Fixed at the source.

Verification

make ci green: format, lint, mypy, and 4379 passed / 1 xfailed. Also smoke-tested end to end against real bd/jj in a scratch repo (seeded ledger entries; exercised list/answer/waive/already-resolved/reconcile/land-status), which is how the reconcile.status documentation fix in land-report-schema.md was confirmed.

Two caveats worth naming:

• tests/integration/workflows/test_reconcile_jj.py and tests/integration/test_assumption_ledger_flow.py intermittently hit the 30s pytest timeout under parallel load with real jj/bd subprocesses. They pass in isolation and are unaffected by this branch — pre-existing flakiness that deserves its own look.
• The skill's guided sweep was validated by tracing its instructions against every acceptance criterion, not by a live /maverick-review session (no interactive Claude Code UI in this environment). Worth one real run against a seeded project before fully trusting the UX.

🤖 Generated with Claude Code

https://claude.ai/code/session_012av72R5SPCc84chsUNjhLC