v0.90.9

Added

• Closed every emitted tool turn with one terminal result across normal, blocked, aborted, timed-out, failed, and resume paths; missing-result repair remains idempotent and duplicate/orphan results fail closed.
• Added a deterministic resource-claim DAG scheduler that preserves source-order artifacts, keeps unknown, bash, and unclaimed extension tools exclusive, and retains waves-v1 as the compatibility rollback path.
• Added execution-bound evidence receipts that bind normalized local command outcomes, artifact/workspace fingerprints, redacted output digests, and replay-ledger state; the built-in CLI and AgentSession bash paths remain outside this optional evidence executor.
• Made compaction transactional behind closed tool turns, revision compare-and-swap, and stale-summary discard.
• Added typed termination, incomplete-run recovery, and omk session doctor, including dry-run repair for unambiguous recoverable state only.
• Added provider-origin-aware omk provider doctor diagnostics with sanitized Level 0–2 probes for native, custom OpenAI-compatible, and local-proxy endpoints.

Notes

• Published to npm as open-multi-agent-kit@0.90.9 (lockstep with omk-ai, omk-agent-core, and omk-tui at 0.90.9); prebuilt binaries are attached to the GitHub release.
• Verification boundary: build/check and the keyless workspace suite passed; live-provider and other-OS coverage remain outside this release. Validate existing integrations against your workload.

v0.90.8

New Features

• GPT-5.6 MoA virtual model — openai-codex/gpt-5.6-moa combines bounded concurrent Sol and Terra analysis into one Sol-synthesized response. See provider setup.
• Path-safe parallel tool-batch waves — independent tool calls run in ordered waves while conflicts and unknown operations remain sequential. See the agent runtime.
• Context-budget v2 controls — enable the global budget with contextBudget.enabled and select an authenticated compaction model with compaction.model. See settings and compaction.
• Evidence-gated Correctness Wall — fixtureless live OA uses a bound MCP handler only and otherwise remains preview-only. See the Correctness Wall example.
• Project-local computer use — the Stagehand extension and omk-computeruse skill add bounded, approval-gated browser workflows. See the extension.

Added

• Added the tool-free openai-codex/gpt-5.6-moa virtual model, which combines bounded, concurrent GPT-5.6 Sol and Terra analyses into one Sol-synthesized response for analysis and review tasks.
• Added partitionToolBatchWaves to run ordered, contiguous safe tool-call waves instead of serializing an entire batch when one call conflicts or is unknown.
• Added compaction.model, allowing auto-compaction and /compact to use an authenticated model such as zai/glm-5.2 without changing the active session model.
• Added the project-local Stagehand computer-use extension and omk-computeruse skill for bounded browser observation, approval-gated actions, and redacted extraction.

Changed

• Added global-only contextBudget.enabled to activate context-budget v2 for every OMK session; OMK_CONTEXT_GOVERNOR=1/0 remain per-process on/off overrides. Each enabled AgentSession reuses an in-memory plan and representation cache without disk or cross-session sharing.

Fixed

• Fixed the CLI help omitting the accepted max and ultra thinking levels, and fixed GPT-5.6 Codex ultra requests failing with an invalid-enum HTTP 400.
• Fixed loop write-scope validation to collapse ./.. path segments, so traversal like src/../package.json can no longer slip past allowedWriteScopes, deniedWriteScopes, or changed-file checks.
• Fixed the context-budget v2 representation chooser to respect the remaining tier budget: it now falls back to a smaller representation (e.g. summary) instead of over-selecting full text and dropping the whole item at the tier ceiling.
• Hardened the guardrails evidence ledger: replay events now carry a prevHash/eventHash tamper-evident chain verified on load (edited, deleted, or reordered ledger lines fail closed), TaskContractBuilder.fromJSON validates the contract shape instead of blindly casting, build()/getEvents()/getLedger() return deep copies, and verification-report table cells escape pipes/newlines from untrusted claim or command text.
• Fixed the footer PKG counter and control-panel ports: label dropping accepted advisory/report-only package ports from every bucket (PKG 12/16 R0 B0 for 16 accepted candidates); the intake summary now counts them via acceptedAdvisory, so all accepted ports read as connected (PKG 16/16).
• Fixed Correctness Wall edit/write hooks to use fixtureless live OA only with non-empty run IDs, explicit MCP transport, and a bound MCP handler; unbound handlers fall back to preview-only evaluation.

v0.90.7

Added

• Added the ultra thinking level (maximum reasoning with automatic task delegation, GPT-5.6 Sol/Terra via openai-codex) to the CLI --thinking flag, thinking selector, settings, theme colors, and router feedback levels; the reasoning router never auto-selects ultra, and models without it clamp down to their highest supported level.

Changed

• Refreshed built-in model metadata via omk-ai regeneration so OpenRouter latest tool-capable models and OpenCode Go (opencode-go) catalog entries stay current for /model and provider defaults.
• Changed the root README project demo link to Mario's Threads profile.

Fixed

• Fixed secret.read_path command-safety false positives for search commands: the pattern argument immediately after -- in grep/egrep/fgrep/rg is no longer treated as a secret file operand, while real secret-file operands (e.g. rg -- needle .env, grep -f .env needle) stay confirm-tier.
• Fixed npm trusted-publishing identity after the GitHub repository rename by aligning package metadata with dmae97/omk.

v0.90.6

Added

• Documented the pinned pi.dev package loadout, slash-command map, and skills.sh install flow in the root README, and added the readmeasset/omkgirl.png hero asset.

• Added the experimental B2C Correctness Wall extension under examples/extensions/correctness-wall/ with fixture/live OA transport wiring, soft/hard/shadow modes, and regressions 018–023; wall evaluation is evidence-gated advisory (not formal correctness proof).

• Added Grok OAuth harness integration: grok-harness domain profile, grok-playbook / grok-harness-dispatch auto-apply for grok-oauth-proxy, compaction model preference for grok-4.5, Imagine chat-model rejection on completion paths, and doctor-provider health probe defaults for the local proxy.

• Added extension host callMcpTool bind path (ExtensionActions.callMcpTool + loader/runtime) so extensions can capture live MCP call capability at load time.

• Added a vendored taste-skill pack under .omk/skills/taste-skill/ (MIT, pinned upstream commit in SOURCE.md) with thirteen Agent Skills (design-taste-frontend, design-taste-frontend-v1, brandkit, imagegen-frontend-web, imagegen-frontend-mobile, and the rest of the Leonxlnx/taste-skill bundle) for project-local design and image-direction workflows without pulling the pack into core runtime code.

• Added an opt-in caveman output-style skill at .omk/skills/caveman/ (MIT, pin 0d95a81, disable-model-invocation: true) so brevity applies only on explicit /caveman or direct user request; it is excluded from every DOMAIN_PROFILES skills allow-list and does not alias the builtin /compact command.

• Added a PATH-gated CodexBar quota CLI (omk quota connect|disconnect|status|usage for target codexbar) with codexbar-connector.json opt-in (enabled, privacyAck, connectedAt only — no credential-shaped keys), tools-manager ManagedTool support for codexbar, and adapter-side JSON redaction before printing usage/cost summaries.

• Extended scripts/check-vendored-skills.mjs (repo root, invoked via npm run check:vendored-skills) to enforce taste-skill thirteen-skill layout, frontmatter names, SOURCE.md pin alignment, and caveman opt-in frontmatter/pin rules alongside existing clone-website and ponytail* vendoring checks; regression external-pack-integration-acceptance.test.ts asserts pin constants match SOURCE.md.

• Added a governed real-world generalization gate for the v4 reasoning-effort router: a synthetic generalization corpus (420 rows, 60 per class, out-of-vocabulary phrasing distinct from the gold set, zero gold-prompt overlap) with a content-blind train/dev/holdout split (frozen 99-row holdout), a pure calibration helper (ECE/MCE/Brier + per-confidence-band error table + aggregate dev gap cards), and the 017-reasoning-router-v4-generalization-governance test that prints aggregate-only full-set and frozen-holdout benchmarks and enforces the escalate-only invariant. Unlike the saturated gold set (100%), this corpus discriminates and gives an honest held-out measure of real-world routing accuracy.

Changed

• Changed the frontend-ui DOMAIN_PROFILES skills allow-list to include vendored taste-skill names (brandkit, imagegen-frontend-web, imagegen-frontend-mobile) for curated activation when domain routing is on; caveman remains absent from all domain gates.
• Changed the v4 reasoning-effort classifier to a table-driven intent lexeme/phrase-cluster model with negative controls and evidence-strength confidence, then evolved its feature model beyond literal keywords with a new pure reasoning-router-v4-normalize module (English light-stemming/lemmatization, Korean josa/eomi normalization, seed-to-cluster intent anchors, and a bounded intent-skeleton / temporal-deferral extractor) so it generalizes over real-world phrasing instead of enumerating vocabulary. On the frozen held-out generalization set this lifts micro from ~0.76 to ~0.92 (debug/refactor/plan reach 100% recall) with zero high-confidence errors on the holdout, while the frozen gold set stays at 100% (210/210 incl. 42-row holdout), the 30-case adversarial probe stays 30/30, and no new dependencies were added.

Fixed

• Fixed threshold auto-compaction compaction_end.willRetry to also signal when agent-level queues remain, so the TUI flush path stays aligned with session resume (compactionEmitWillRetry).

v0.90.5

Added

• Added reasoning-router-resolver.ts as the shared thinking-level resolver core for the v4 reasoning router, centralizing resolveThinkingLevelCore, TASK_CLASS_THINKING_LEVELS, lane-step adjustment, bounded bias/hint fusion, confidence escalation, and available-level clamping.
• Added deterministic capability routing to the subagent example extension, including catalog-backed capability parsing and validation, domain-profile derivation, assigned-capability preambles for bundled agents, optional skill enforcement, unit coverage, the read-only scripts/derive-agent-capabilities.mjs classify/audit CLI, and the scripts/check-agent-capabilities.mjs live-catalog guard.

Changed

• Collapsed automatic reasoning-effort routing to a single /think auto backed by the v4 confidence-bearing router. The /think auto-v1, /think auto-v2, /think auto-v3, and /think auto-v4 variants (plus their auto vN / auto:vN forms) are removed, and /think auto now routes via v4 instead of the former v1 classifier; those variants were opt-in experimental surface introduced in 0.90.4. Manual /think <level> still always takes precedence.
• Changed context-budget v2 observability to inline the legacy token-optimizer compatibility status in the planner/types instead of importing the removed optimizer module; telemetry still reports optimizer_id="legacy-token-optimizer", status="quarantined_compatibility", and active_context_budget_optimizer="context-budget-v2".

Fixed

• Fixed v4 reasoning-router misclassification of real-world, out-of-gold-set phrasing by extending the keyword families with vocabulary the classifier previously had no signal for: review (sanity-check, once-over, eyeball), refactor (tidy up), plan (map out, phases, and Korean 계획 … 세워 with an intervening particle), debug (track down, leak), plus a narrow negation-aware review cue. An out-of-vocabulary adversarial probe improves from 22/30 to 30/30 while the frozen gold-set holdout and full set stay at 100%, guarded by the new 016-reasoning-router-v4-generalization regression test.

Removed

• Removed the v1, v2, and v3 reasoning-router modules (reasoning-router.ts, reasoning-router-v2.ts, reasoning-router-v3.ts, reasoning-router-weights.ts), their regression suites (003-007), and the v3 calibration script; only the v4 router remains, reachable via /think auto.
• Removed the unused compactor.ts module.
• Removed the legacy token-optimizer.ts runtime module and its dedicated test while retaining fixed compatibility telemetry in context-budget v2.

v0.90.4

Added

• Added per-task reasoning-effort routing: /think auto keeps the stable v1 local classifier, while /think auto-v2 / /think auto v2 select the calibrated v2 weighted router for the current session; manual /think <level> always takes precedence and levels are clamped to model capabilities.
• Added the benchmark-backed v3 contextual reasoning-effort router: /think auto-v3 / /think auto v3 select v3 for the current session while /think auto remains v1 and /think auto-v2 remains v2.
• Added the confidence-bearing v4 reasoning-effort router: /think auto-v4 / /think auto v4 select v4 for the current session, adding per-turn confidence-band, margin, and fallback-reason metadata so a low-confidence or fallback-decided verdict can only raise, never lower, the resolved thinking level; /think auto remains v1, /think auto-v2 remains v2, and /think auto-v3 remains v3.
• Added a gold-set evaluation governance harness for the reasoning-effort routers: fixed train/dev/holdout splits plus accuracy, macro-F1, severe-under-allocation, and class-flip/McNemar reporting scripts, which confirmed v4 classifies identically to v3 across the full gold set, including the held-out rows.
• Added global opt-in v4 learning-bias wiring: reasoningRouterLearning.enabled lets /think auto-v4 load one validated bias snapshot per session, apply a bounded local bias, and append privacy-safe feedback records without storing raw prompts, file paths, diffs, session identifiers, model/provider payloads, tool output, or hook output. Project-local settings cannot enable or redirect it. The Adaptorch advisory bridge remains default-off groundwork with no runtime transport or session call site.
• Added a safety caveat to the docs-writing domain loadout: slides-grab is external, heavy, and runtime-dependent, so it must never be vendored, must never default to reusing a private local session as a credential, and must never have its mutable remote install instructions followed blindly. Part of the Goal 011/012 external-skill-integration review, which also vendors a pinned, provenance-stamped clone-website skill and six pinned Ponytail markdown skills (ponytail, ponytail-review, ponytail-audit, ponytail-debt, ponytail-gain, ponytail-help) while keeping strix's executable pentesting runtime out of any integration.
• Added !omk bang-launcher routing for OMK role hubs: exact aliases such as frontend, backend, loop, and plan select the matching OMK hub skill, free-form prompts are scored against the hub vocabulary, and !omk never falls through to bash.

Changed

• Changed the OMK//CONTROL runtime status panel so the MCP badge counts only stable server entries, the skills total excludes OMK hub/routing-only skills, and the Headroom line shows the detected installed version while tolerating legacy output formats.

v0.90.3

Removed

• Removed the entire jailbreak subsystem that shipped in 0.90.2 and earlier (the omk jailbreak command, --jailbreak-mode/--jailbreak-target CLI flags, jailbreak-extension, agents/jailbreak/, utils/jailbreak/, types/jailbreak.ts, and the fuzzing/, routing/, encoding/, multiturn/, and modules/ attack modules). These generated cross-provider godmode/parseltongue/LRL-bypass payloads and are not a legitimate OMK feature. Defensive GOD Mode resistance in the system prompt and the context-file sanitization marker are retained.

Fixed

• Fixed the 96-column control-panel visual-QA test to run its live-render coherence assertion unconditionally and only compare the recorded .omo/ visual-QA artifact when it is present, so CI checkouts (where the gitignored artifact is absent) no longer fail.

v0.90.2

Added

• Added the max thinking level to the CLI (--thinking-level max), thinking selector, and settings, exposed for models that support the top effort tier (e.g. Claude Opus 4.7/4.8).
• Added Zyloo provider setup documentation, CLI env-var help, and /login display name.
• Added the reverse-skill routing module and project extension for generating OMK reverse-engineering workflow skills.
• Added regression coverage for the context-budget v2 cache hit path, including stale and negative cache rejection (the telemetry itself shipped in 0.90.1).
• Added computer-use MCP preset catalogs (Playwright MCP as the default, browser-use as an optional advanced runner) and public MCP preset metadata with risk/auth policies, as groundwork modules that are not yet wired to a user-facing command.
• Added release consistency checks for package-backed OMK//CONTROL version display and release metadata drift, including git tag-lineage checks (flags a tagged/released version whose commits never merged to main, degrading gracefully when tags are absent) and README/.github/RELEASE_NOTES_v*.md release-surface drift checks; both report as warnings outside --release mode and as errors during --release.
• Added ! skill launcher support with !skill:name and !name turn-scoped active skill invocation while preserving ! command and !! command bash shortcuts.

Changed

• Changed release-surface consistency checks to accept the pre-release alignment state (README release surface and RELEASE_NOTES link ahead of the latest changelog release) outside --release mode, so docs(release): align ... commits can land before the release version bump.
• Changed the changelog consistency check to require the [Unreleased] top section only outside --release mode, since the release script renames it to the release version before the release commit and re-adds it afterwards.

Fixed

• Fixed the legacy hooks/ deprecation warning blocking startup on every interactive launch. Project-level .omk/hooks/ is now auto-archived to hooks.migrated/ (project shell hooks are never executed), and the remaining global hooks//tools/ deprecation notice no longer waits for a keypress.

v0.90.1

Added

• Added GitHub-focused v0.90.1 release notes and README release documentation for the current OMK package line.
• Added an autopilot automation profile with browser-use integration for repeatable operator workflows.
• Added materialized context-budget v2 representation and plan cache reuse with validation and telemetry.

Changed

• Changed release docs, badges, install references, and OMK//CONTROL TUI copy to the package-backed omk v0.90.1 · OMK//CONTROL release surface.
• Changed standalone package release metadata to keep open-multi-agent-kit, omk-ai, omk-agent-core, and omk-tui lockstep across npm and GitHub release surfaces.

v0.80.8

Added

• Added GitHub-focused v0.80.8 release notes and README release documentation for the OMK hard-fork release.
• Added source-backed public hooks fail-closed primitives for the ./hooks export.
• Added containerization documentation and a Gondolin extension example for routing built-in tools into a local micro-VM.
• Added Ant Ling provider selection and setup documentation.
• Added NVIDIA NIM provider selection, setup documentation, and direct NIM request attribution headers.
• Added ctx.mode to extension contexts so extensions can distinguish TUI, RPC, JSON, and print mode.
• Added ctx.getSystemPromptOptions() for extension commands to inspect the current base system prompt inputs.
• The interactive control footer now shows sampled process CPU% and RSS memory (width-adaptive segment).
• Two new built-in themes omk-control and omk-control-light.
• The working loader now exposes the model thinking level while streaming.
• The OMK//CONTROL startup dashboard now displays the package-backed omk v0.80.8 · OMK//CONTROL release header.

Changed

• Changed internal runtime package dependencies to OMK-scoped packages and publish them before open-multi-agent-kit so installed omk can resolve its runtime imports.
• Changed published coding-agent package metadata and source config defaults to use OMK/Open Multi-Agent Kit names, omkConfig, .omk, and the OMK repository.
• Changed release metadata to lockstep OMK publishable packages at v0.80.8.
• 256-color theme quantization generalizes the cube/grayscale selection for near-neutral colors (colors with channel spread < 10 keep byte-identical output).

Fixed

• Fixed auto-compaction to trigger before provider requests when the projected prompt would exceed the configured context headroom.
• Fixed temporary extension package installs to use a private ~/.omk/agent/tmp/extensions directory with 0700 permissions instead of a shared OS temp extension directory.
• Fixed git package source handling to reject unsafe host/path components and keep managed clone paths inside install roots.
• Fixed stored XSS in HTML session exports by sanitizing Markdown link and image URLs with a scheme allow-list after stripping control characters.
• Fixed SDK embedding in bundled Node apps failing with ENOENT when package.json is not present next to the bundle entrypoint. The package metadata reader now gracefully handles missing package.json by using defaults, enabling createAgentSession() without requiring package-adjacent files at runtime (#5226).
• Fixed HTTP timeout setting not being respected for non-Codex providers (e.g., llama.cpp via OpenAI-compatible API). The httpIdleTimeoutMs setting (set via /settings HTTP timeout) now applies as the default SDK request timeout for all providers that support it, not just OpenAI Codex Responses. Disabling the timeout (HTTP timeout = false) now correctly disables SDK timeouts for all supported providers by sending a maximum int32 value (effectively infinite) instead of 0, since SDKs treat timeout=0 as an immediate timeout (#5294).
• Fixed opening and listing very large JSONL session files by reading session entries line-by-line instead of materializing the full file as one string (#5231).
• Fixed renderShell: "self" tool renderers that emit no component lines leaving a blank chat row (#5299).