Repository navigation
kane-cli 0.7.1
Highlights
Multi-source extraction lands with full pipeline support: sources are cited, deduplicated, queued, and drained automatically. A new trust and review layer catches unreviewed or conflicting designs before they commit. Paused and crashed runs are now durably resumable with honest, auditable provenance.
[0.7.1] - 2026-08-08
Multi-source extraction and queued drain
- Extract from multiple sources in one run — each tracked with its own
source_idand completeness record. Op-union (keep/set/remove) lets the same criterion be built from contributions across sources. - Queued drain:
extract --queued— pre-drawn work items are claimed from a durable queue and processed in budget-aware batches (--budget n; default 10 per composite root). Mirrored items complete without spending budget; a budget stop names the remaining queue and prints the exact command to continue. - Agent failures inside a queued batch no longer abort the whole batch — a failed claim is recorded as
AGENT_ERRORand the batch continues. Early exits (including pause) release claims cleanly. --queuedis incompatible with--source,--resume, and--plan— kane-cli exits 2 immediately if these are combined, rather than producing a silent bad run.
Trust gate and review surface
- Unreviewed design targets now block extraction — attempting to extract onto an unreviewed-derived design causes an explicit disclosure prompt (interactive: yes/no confirm; agent: exit 2 with a runnable review command in
next; CI: exit 2). Pass--allow-unreviewedto bypass. - Duplicate hits go to a review queue instead of silently promoting — an explicit match onto an unreviewed-derived target mints a new derived entry and records a durable
PAIRcard for review, rather than laundering a promotion. - One review walk covers REVIEW, PAIR, MODIFY_HELD, and QFAIL cards — all verdict types surface through the same interactive walk grammar. A feeder fault never blocks the reconcile walk.
- Held archives require explicit consent —
--verdictsdefault holds rejected entries as non-destructivepending_archivefacts (exit 0, loud summary). Destruction requires--allow-archive --because <reason>; CI refuses archives under any flag (exit 2, atomic). Structured flags--approve/--skip/--deferare the agent-mode surface.
Durable sessions: pause, crash, and resume
- Paused runs are fully resumable — the session is written to disk from turn 0, before the agent spawns. A pause persists the checkpoint and releases locks cleanly; the printed resume command is always runnable.
- Crashed runs exit 3 and name the resume command — a crash no longer exits silently or leaves the session in an ambiguous state. Persist failures exit 1 and name the
sidand sequence number. - Resume now reconciles store truth against checkpoint truth — on
--resume, committed work is read from the durable record (not recomputed), externally-verdicted rows are pruned and never re-presented, and foreign watermark conflicts are surfaced loudly with a typed error. - Wrong-verb resume is refused honestly —
extractno longer silently picks up a design session, and vice versa. Unknown-verb sessions (written by a newer kane-cli) get an honest no-command line rather than a fabricated hint. sessions listandsessions showare verb-honest — sessions written by a newer kane-cli version are shown with an honest unknown label rather than a guaranteed-refusal extract resume hint.
Design phases and cite verification
- Design now tracks phases durably — each finalized design record carries
{phase, generation, completeness}.--phase <name>on design tests lets a run enter at any phase; missing predecessors prompt interactively or exit 2 with a parseablenextin agent mode. - Already-designed and stale-design gates convert to actionable asks — interactive mode offers view/redesign/pick-another with impact pairs shown; agent mode exits 2 with structured
nextincluding--forceandevolverows; CI stays fail-closed.--becauseis required for redesign. - Cite verification runs before every design commit —
validateCitechecks the pinned source version's citable text before appending aDERIVESedge. Failure routes through the existing repair channel asCITE_UNVERIFIED; sessions on older templates are byte-identical.
New coverage / gaps surface
--mode agent|cion coverage tests — both modes ride the same NDJSON envelope as a decorator over one implementation.--jsonoutput is captured as a singlecoverage/gapsevent; refusals emit an error event and exit 2.nextcontains the gap engine's ready commands, grammar-filtered, deduped, and capped.
Headless and --source / --resume parity
--resume --with-sourcenow works in headless mode — headless runs have full parity with interactive runs for source-attached resume.- The
@affordance in free-text options — an option carries its associated text through the agent, so prefill and coverage work correctly end to end. --trust auto|holdregistered —autois today's derived auto-commit behavior (pinned);holdforces all new BUs into the review queue. Under CI,--trust holdexits 2 with nothing executed.
Fixes
- Dedup probe now uses the run's live auth — a stale bearer token was causing commit-moment 401s off-CI, holding every new BU across sessions. The probe now resolves credentials the same way the driver does.
prefillholds by question identity and never auto-submits — a question that was already answered is not re-asked; a prefilled answer is not submitted without user confirmation.- No-proposal terminal under
--queuedno longer re-queues forever — a clean done with no proposal now recordsNO_PROPOSALand continues the batch rather than re-queuing the item with attempts unchanged. - Answered-question filter covers both the annex and session ledgers on resume — a crash at an ask no longer re-asks the same question after resume.
MODIFY_HELDapprove no longer fabricates provenance — cites are verified against the pinned source before any held-update verdict is applied.- Verb/template mismatch is now detected at bind time — a design conversation that echoes a valid
extract@5binding no longer renders the extract prompt and tools under clean digests. Kane-cli exits 2 withCONTRACT_DIGESTmismatch. --resumenever re-fires the design-entry trust gate — a session that passed the gate at entry is not re-prompted mid-journey.sessions listreview walk no longer crashes on review cards —WalkCard.impactis now a legal union; all three readers guard by shape and render the review preview correctly.
Install
# npm
npm install -g @testmuai/kane-cli@0.7.1
# Homebrew
brew install lambdatest/kane/kane-cli
# Shell installer (latest release)
curl -fsSL https://raw.githubusercontent.com/LambdaTest/kane-cli/main/install.sh | bashCompare: 0.7.0...0.7.1