Repository navigation
Releases
deep-loop v1.28.0
Fixed
- The closed
review configurerecovery profile for a Codex plus Agy review pinned models two generations old (#79). Agy 1.3.5 no longer recognizesgemini-3.6-flash-high, so with--no-fallbackthe Agy reviewer failed withERROR_UNSUPPORTED_MODELon every run. The profile now pinsgpt-6.1-solat high effort for both Codex roles andgemini-3.8-flash-highfor Agy. Both ids were checked against their CLIs on 2026-10-14 (codex-cli 0.162.1, agy 1.3.5).
Changed
- The profile's canonical name is now
codex-agy-static. Likecodex-only-static, it names the providers rather than a model generation, so the next pin change does not make the name wrong. gpt56-agy-staticis still accepted. It is an alias that expands to the same flags ascodex-agy-static. The result and thereview-configuredevent record the canonical name asprofileand the requested name asrequested_profile. A request that uses a canonical name produces the same shapes as before.- The Agy effort stays in the model id suffix (
-high), because deep-review's Agy adapter has no effort transport./deep-loop-statusand both READMEs now say this. They also say that a rejected pin fails closed, and that the Agy role runs the agy CLI inside the reviewed tree. scripts/lib/review.mjsexports its frozen profile and alias tables. A docs test reads the pins from them. It fails when either README or the status skill stops naming a profile, an alias or a pinned model. It also fails when runtime code or skill text names a retired pin again.
Compatibility
- Runs already configured with
gpt56-agy-statickeep the flags recorded in theirloop.json. The kernel does not rewrite them. Every run on this machine that used the profile was alreadycompleted. A live run with the old flags can be moved to the new pins through the existing recovery route: abandon the failed checker as an operational failure, then runreview configure --profile codex-agy-static. - No durable format changes. 1.27 reads everything 1.28 writes;
requested_profileis an extra field in event data that nothing reads.
deep-loop v1.27.0
Fixed
- Old run history no longer switches off cwd run selection for the whole project (#77). More than 64 run directories, or one finished run with a claim an older version recorded as an absolute path, used to make
run resolve,run status, the status band and the PreCompact, PostCompact and SessionStart hooks fail for every session, so no compact checkpoint was written. - Terminal worktree claims that overlap — two finished runs on one path, or a finished run and an active run — no longer make the whole project ambiguous.
- PreCompact and SessionStart passed their fixed business time as the selection deadline clock, so their selection time limit never applied. They now use a real clock.
- PostCompact counted every entry under
.deep-loop/runs/and gave up on a regular file named like a run id; it now counts run directories like the kernel does. - PreCompact answers a selected finished
workstream-sessionrun withno-run-terminalbefore any affinity or checkpoint step.
Changed
- Run selection reads
completedandstoppedruns without the run lock when theirloop.jsonmatches.loop.hash, passes the schema and binds to the project root. Other runs get the full locked verification, and a selected run is always re-read with it. The event log, pending transactions and artifacts of terminal history are not checked on this path;--run-idreads andvalidatestill check them. - Bounds: 256 run directories (was 64), 64 full verifications, 4096 worktree claims, 32 MiB of state read while classifying. Time: 500 ms + 2 ms per run directory to classify, then 500 ms + 100 ms per full verification, capped at 3000 ms (
cli-read) or 6900 ms (hooks,run list). - A terminal claim affects only a cwd inside it (
none,terminal-residue,source: worktree). Inside a worktree path that a new run reused, the safety net stays off; PreCompact now writes one diagnostic line there and the band says so. - A legacy absolute claim of a finished run that points into
.claude/worktrees/or.worktrees/is read as its relative form; other unresolvable claims of finished runs are isolated and counted inrun resolve'shistory. - An active run's non-conforming claim now fails with reason
invalid-worktree-claim(wasrun-set-integrity/integrity-invalid). A run directory withoutloop.jsonis kindstate-missing; a lock that stays held islock-busy. run listrows carryverification(state-hashorfull) and stay filled when another run errors.run resolve, hook and headless diagnostics carryphaseandboundfor an exceeded bound (enumeration,classification,full-capture-count,lock-retry,claims,recapture).- The status band names four run-selection failures (
run selection unavailable (<reason>)) and a cwd inside a finished run's worktree./deep-loop-statusexplains each reason and the human remedy, and says where no CLI remedy exists yet.
Compatibility
- A terminal run with a leftover transaction no longer appears as
reconciliation-requiredinrun listor blocks selection. - No durable format changes; 1.26 reads everything 1.27 leaves behind.
deep-loop v1.26.0
Added
- Read-only
run status [--json] [--cwd DIR] [--run-id ID] [--project-root R] [--now T]. It reusesrun resolve's selection and prints one closed-shapestatus_version: 1envelope: status, pause reason, budget, comprehension debt, pending human reviews, breaker, workstream counts and the next action. Exit 0 or 1 always carries the envelope; exit 2 is a usage error with empty stdout. It takes no lease, writes no state and evaluates no goal proof. - Claude Code status band (
hooks/status-band/, Claude Code 2.1.287 and later). One line above the prompt for the selectedrunningorpausedrun, or a notice when several active runs make the choice ambiguous.Status,AckandHidebuttons only fill an empty prompt; nothing is submitted. One merged toast on breaker, debt-gate, budget-stop and completion transitions. It runs onlyrun status --jsonand is active only while a terminal or desktop surface is attached.
Changed
- The static assertion grader recognizes Claude Code Mods host calls (
$.process.run,$.process.spawn,$.http.fetch), including aliases, optional chaining, wrapper members and helpers, so the band is covered byno-external-action-routes. /deep-loop-statusstates correctly thatbudget checkreturns{ok, reason, tier_after}; the figures come fromstate getorrun status --json.
Compatibility
.claude-plugin/plugin.jsonnow points athooks/hooks.claude.json, which repeatshooks/hooks.json's command hooks verbatim and adds the module. Grok 1.0.46 reads only the manifest hooks file, so the repetition keeps its hooks unchanged. Claude Code runs identical command hooks once (measured on 2.1.291, including a real/compact, and on 2.1.200). Codex reads onlyhooks/hooks.json, which is byte-identical.- Claude Code without Mods support, Codex and Grok show no band and behave as before. No schema or migration change.
- Test and CI results demonstrate the covered contracts, not statistical real-agent efficacy. The real Claude desktop-app composer was not exercised.
Implementation and validation history: #76
deep-loop v1.25.0
Added
- Read-only
router probe --json. It reports the realpath of the deep-model-router install that the kernel locator selects, that install's manifest version, the run's frozen policy digest, and thepolicy_pinto send. - Route requests after a run's first routed episode carry the frozen digest as
policy_pinwhen the router is 1.16.0 or newer. A local overlay publication no longer stops an in-flight run.
Changed
- The continue skill's shared router procedure (§1.75) runs only the path returned by
router probe, adds the pin to RouteRequestV1, and spells out the request and--routingshapes. It is used by the legacy maker, the legacy checker and v0.5 fresh attempts. - A pin that the router cannot reproduce is reported as
router-policy-pin:<reason>with recovery guidance. Band rules are unchanged. - The router locator rejects the relative
../deep-model-routercheckout by path segment, including the.././,..//and Windows drive-relative spellings. It compares personal-skill markers case-insensitively.
Compatibility
- No schema or migration change. Records carrying
request.policy_pinremain valid for the previous kernel. - Older or unidentifiable routers receive no pin, and the recording-time digest check still applies.
pin_suppressed_by_offrecovers in the same run. A router plugin update, a later revocation or a missing generation still needs a new run.- Test and CI results demonstrate the covered contracts, not statistical real-agent efficacy.
Implementation and validation history: #74
deep-loop v1.24.1
Fixes
- Resume an explicitly confirmed needs-human pause with the same active owner and open workstream affinity. Budget, breaker, lease and handoff fences remain enforced inside the transaction.
- Treat explicitly abandoned inline maker attempts as cancelled work when checking whole-goal prerequisites. Unknown external attempts and invalid, unconfirmed or unfinished work still block completion; cancellation does not fabricate review credit.
Compatibility
No durable schema migration or relaxed review/finish requirements. Test and CI results demonstrate the covered contracts, not statistical real-agent efficacy.
Implementation and validation history: #72
deep-loop v1.24.0
Added
- Read-only
goal drive --checkwith typed support/remediation results and
host-local execution plans bound to the active controller, fence, profile,
review configuration and trusted checker doctrine. - Supervised same-requested-model
subagent-checkerexecution with fresh CLI
session/process evidence and exact existing maker-claim/import provenance. - Per-call time limits, fresh run/host budget admission, exactly-once host call
charges, bounded per-workstream/episode progress monitoring and explicit
recovery diagnostics. Unknown usage or termination cannot become success. - Versioned repeated real-agent evaluation with fixed scheduled denominators,
three additional behavior tasks, source/argv/raw-trace evidence and a separate
controlled host-loss safety lane. V1 result/profile formats remain available.
Compatibility and evidence boundary
- Experimental
goal drivesupports explicitsubagent-checker,same-model
and empty review flags. Unsupported default/cross-model/sibling-loop policies
refuse before maker execution. Installed deep-review provides criteria only;
served-model identity and automatic lost-host reattachment remain unavailable. - V0.4/v0.5 durable schemas, lease fences, proof-derived completion and the exact
owner-only terminal-cost exception remain unchanged. - Token caps govern admission after measured calls and may be overshot in flight.
Fixture success, safe refusal and model availability are not efficacy evidence. - Real-agent completion remains experimental; no reliability or efficacy promotion is claimed.
PR: #71
Validation: all six required Linux/macOS/Windows × Node 20/24 CI axes passed on the identical pre-merge tree. Full Node 24 Linux lane: 3,034 passed, 8 skipped, 0 failed. Unit lane: 285/285.
Measured limitations: fixed pilot planned 45 rows, started 6, left 39 unstarted. Two native trials passed, three minimal trials exceeded budget, one current trial timed out with incomplete usage. No harness trial reached kernel completion. The user explicitly accepted incomplete dual independent review after two Fable timeouts; one protected Opus review was valid. Missing review evidence is not a completed review.
deep-loop v1.23.0
Added
- Opt-in v0.5 goal contracts with immutable requirement identities, typed obligations, current source/artifact snapshots, and independent whole-goal completion proof.
- Recoverable execution attempts, explicit continuation stages, and experimental measured Codex
goal drive. - Separate real-agent evaluation profiles and bounded POSIX process-group evidence.
Changed
- Goal-driven defaults use delegated supervision, same-owner continuation, and an implementation review point. Existing v0.4 runs without a goal contract keep released behavior.
Verification
- GitHub matrix: Ubuntu, macOS, and Windows on Node 20/24 — 6/6 successful on merge SHA 16f12ee.
- Live current-profile smoke for requested
gpt-6-astrais unavailable (GPT models treated as unavailable). The live gate is not claimed.
Integration
- Plugin PR: #70
deep-loop v1.22.0
Added
- Optional C-lite
decision_fingerprintandrequest_sha256linkage passthrough. - Deterministic, post-commit, fail-open
RouteObservationV1emission with no-clobber publication. - Opt-in deep-model-router observation validation for synchronized router installs.
Fixed
- Windows canonical validator path handling, UTF-8 parity, and LF-stable golden fixtures.
- False
integrity-invaliddiagnoses caused by transient.lock.release-<token>artifacts during concurrent verified reads.
Verification
- Local preflight: 2,755 tests; 2,738 passed, 17 skipped, 0 failed.
- GitHub matrix: Ubuntu, macOS, and Windows on Node 20/24 — 6/6 successful.
- deep-model-router compatibility: 393 router tests plus the actual validator gate passed.
Integration
- Plugin PR: #62
- deep-suite pin PR: Sungmin-Cho/deep-suite#101
- deep-suite pins both marketplaces to
b388f400c1b5dc9098cfca9bc7e8abb099719cfb.
The offline fixture bank's FULL_BANK_ACCOUNTING result remains a documented v1.21.0 baseline issue and is not counted as release-green evidence.
deep-loop v1.20.0
deep-loop v1.20.0
Suitability-gate release. Merge: 4a39168 (PR #60). Marketplace pin: deep-suite#91.
[1.20.0] — 2026-08-19
Added
- Entry suitability gate (
/deep-loop§2-1.5). Before a run is created, the
entry skill judges once whether the confirmed goal is loop-shaped. A fitting
goal proceeds with zero extra friction; a clearly misfit goal (bounded
one-shot question/analysis, review-only, single-session implementation,
trivial fix) gets one reasoned alternative prompt — AskUserQuestion on Claude
Code, the same options as a conversational question elsewhere. Ambiguous
goals proceed as a loop, no option is preselected or auto-applied, and no
answer or a tool failure stops without a default. Choosing an alternative is a
terminal branch: the goal is handed to the sibling skill or done inline in the
same conversation and never falls through torecipe-match/init-run.
Prose, docs, and marker tests only — no kernel, recipe, schema, or skill
frontmatter change.
Notes
- Evals profile
deep-loop-current-v1.19is replaced byv1.20, in lockstep
with the plugin minor asloadFixtureProfilederives the expected id from it.
The durable schema stays0.4.0— this release does not touch the kernel.
Deep-suite re-pin is proposal-only.
deep-loop v1.19.0
deep-loop v1.19.0
Kernel-contract-trust release. Merge: 3586e1d (PR #58). Marketplace pin: claude-deep-suite#87.
[1.19.0] — 2026-08-17
Added
-
helpread-only surface.help,--help,-h, and empty argv list everyROUTE_FLAGSroute and exit 0. Empty argv previously exited 2 (unknown subcommand: <none>).help <handler>prints that handler's verbs and(LOCATOR ∪ allow) − rejectedflags. -
npm run test:unit. Thetests/unit/lane runs throughscripts/run-unit-tests.mjsso Node 20/26 both discover it.
Changed
-
Unknown flags and unexpected positionals are usage exit 2 on every dispatcher-owned route. The table is the allowlist. Review denylist spellings stay in
allowso they still reachREVIEW_METADATA_FORBIDDEN(exit 1). Handler-owned grammar routes keep fence-first polarity.
Compatibility: callers that passed unread or mistyped flags, or stray positionals, which used to be silently dropped now fail closed. -
Grok guidance.
compact_supportedis false, so grok loops never receiveadvice:'compact'. Handoff continuity useshandoff_continuity_note(grok attended-Darwin sentence; claude/codex strings unchanged). SessionStart no longer defaults the host token to'claude'.
Fixed
-
Nine mutating routes now return exit 3 for
PROJECT_ROOT_FENCED.pause,breaker reset,recover,recover --supersede-affinity,recovery acquire,attended-launch revoke,spawn-style reset-desktop,budget extend, androot recovery acquireused to fold that fence to exit 1. Read-onlynext-action/state getstay exit 1;path resolvestays 3. -
Windows lock-release rename. Transient
EACCES/EPERM/EBUSYon lock directory rename is retried throughrenameAtomicWithRetry, so a still-alive parent does not leave.lockbehind for concurrent diagnose.
Notes
- Schema stays
0.4.0. Evals profiledeep-loop-current-v1.18is replaced byv1.19. - Claude/Codex attended compact advice is unchanged. Grok remains attended Darwin only,
compact_supported=false. - Deep-suite marketplace pins (
.claude-plugin/marketplace.jsonand.agents/plugins/marketplace.json) are3586e1d6361ec7e7df0d4767e98fb4b1b64c510c.
Part of Deep Suite — a harness layer for Claude Code and Codex.
/plugin marketplace add Sungmin-Cho/claude-deep-suite
/plugin install deep-loop@claude-deep-suite