v0.2.0

Released: 2026-09-27

Added

• Official checksum-pinned Herdr 0.9.1 releases replace the vendored patch and Rust/Zig
  build. Senpi's built-in integration owns reports; OLW recognizes pi and omo without
  duplicate session reports. Owned clones use upstream groups and worktree.create.
  olw doctor blocks the operational switch and lists remaining legacy parents. Existing
  servers, installed binaries and caches are untouched; one-release rollback is documented.
  New linked parents (--repo), children and successors of legacy parents are rejected with
  legacy_parent_unsupported before external side effects; close/migrate with approval or
  continue those parents with the retained pre-switch OLW and patched Herdr together.
• Pinned-by-default OMO model routing with accepted baselines, route-level upstream advice,
  selective apply/dismiss, optional follow mode, launcher and manager notices, and catalog
  health review for removed or degraded models. Advice is read-only until explicit acceptance.
• Issue deliverables (pr, report, document) on bindings, briefs, packets and status. Owned-clone
  children publish one PR into their parent's integration branch; pr open reuses open PRs and
  pr merge checks the reported head and merges with a merge commit. PR-less results require a
  deliverable path. Project PRs stop for user review; legacy local merges remain deprecated but unchanged.
• Fetch-before-close publication guards for owned parents and their PR children, with explicit
  --discard, and isolated bare-remote/gh-shim PR lifecycle QA.
• New mapped parents run in OLW-owned, hardlinked clones of repository mirrors, with readable
  per-project Herdr groups and child worktrees owned by the parent clone. Explicit private local-file
  copies, setup deadlines/logs/receipts, and close-time unpushed-commit reports are supported.
  Legacy checkouts remain unchanged; parent create --repo conflicts with a mapping (later rejected; see the official Herdr entry).
• Scope projects can declare a target Git remote and default branch. OLW maintains one locked,
  fetch-only mirror per remote and exposes mirror status and explicit fetching through olw doctor
  and olw repo list|fetch.
• Child modes: child create --mode direct|planned|research. A planned child runs
  a plan stage (opencodex/anthropic/claude-fable-5-1 xhigh, ulw-plan) and then an
  execute stage (opencodex/anthropic/claude-opus-5-5 medium, ulw-execute driving
  mass-ulw) in the same worktree, each in its own Herdr tab and binding. stage complete records the plan hand-off (path, sha256, head) and stage start --stage execute stops the plan session and opens the execute stage. status
  shows mode, stage, stageBindings and openQuestions per lineage generation;
  close on any stage closes that generation and its workspace once. The parent
  sends the issue packet to the execute binding; nothing transfers automatically.
• Questions and answers: question and answer message kinds, the olw_ask tool
  in bound child and parent sessions (native ask_user_question is blocked there),
  and the ask, answer and read-only questions commands. A parent's question
  goes to its ready manager or, without one or with --to-user, to the user inbox;
  answer --as-user answers an inbox question.
• The manager: olw manage opens or reattaches the scope-free management session
  inside the OLW host with the user's default model (fallback Opus 5.5 medium).
  Parents created while it's ready link to it automatically; parent create --no-manager opts out. manage runs an update check at start and puts the
  result in the manager brief.
• Update automation: olw update check reports the pinned OMO and Senpi versions
  against the npm dist-tags (omo-ai beta, @code-yeongyu/senpi latest) and
  never installs; olw update prepare builds an olw/update-omo-<v>-senpi-<v>
  branch from the selected remote's dev (default origin) in a separate worktree, runs install, typecheck, test and
  build there and opens a PR to dev (draft if anything failed). Neither touches
  the live host.
• medium is an accepted thinking level for role models.
• User-managed opencodex model metadata catalog (MODEL_CATALOG in
  src/proxy/model-catalog.ts). It overrides the context window, output limit,
  input modalities and reasoning flag that ocx exports late or wrong, for example
  the 32000 output stand-in on every model (lidge-jun/opencodex#5828). The
  managed omo launcher and role panes load dist/extension/model-catalog.js,
  which re-registers corrected models at session start.
  bun run proxy:routing catalog lists the active changes and any redundant or
  stale entries.
• Issue children start in packet-bound mass-ulw mode, using native phase workflows
  and independent artifact verification before reporting to the existing parent.
• Real child workflow QA for successful execution and recovery of invalid work
  without repeating successful nodes: qa:child-workflow happy|failed-node.
• Model chain warnings on every managed omo start, in olw doctor (under
  chains, still ok) and via bun run proxy:routing chains: a category or
  agent with no working model or a single remaining model, and OLW role models
  missing from the opencodex catalog. Warnings never block a start and print once
  per change.
• Durable execute-stage launch intent: concurrent recovery of an execute stage is fenced to
  one launch, and reconcile settles a recovery whose outcome is uncertain.

Changed

• CI now records exact-commit results for every dev merge and rejects pull requests into
  main. Owner-authorized releases validate that evidence, publish an immutable source tag and
  GitHub release, then fast-forward main with dedicated release credentials. The first release
  after adopting the workflow is bootstrapped by an owner-pushed tag; later releases use dispatch.
• Repository policies follow docs/policy/: PRs target dev and merge with a merge
  commit, main only advances to released commits, and issue and PR text is in
  English. Issue forms are now bug and proposal.
• Pinned the official omo-ai 5.0.0 and @code-yeongyu/senpi 2026.9.26 releases, retaining only runtime-repair hunks that remain necessary.
• Models route through opencodex. The routing synchronizer reads the
  opencodex-owned catalog in models.json, and OLW role pins use opencodex.
• Child fixture standby permits internal workers only after explicit instruction;
  creating additional OLW roles and accessing live Linear remain forbidden.
• Only children may hold an issue-packet goal. Supervisor/parent waiting and model
  assignments remain unchanged; existing bindings are not reinitialized.
• An upstream xAI route is followed by the same model on Cursor when opencodex
  publishes it (xai/grok-4.7, then cursor/grok-4.7). No other same-model host
  becomes a second lane.
• A managed category or agent route whose upstream choices opencodex no longer
  publishes is written without models and listed as unroutable, instead of
  rejecting the whole routing update.
• stage start --stage execute requires an accepted plan report and an unchanged plan file;
  legacy hand-offs without one fail closed. Repeating an identical stage complete attaches
  the accepted report id to the hand-off. Reported plan paths are resolved.
• The manager session keeps the user's native model and thinking level instead of
  pinning them.
• A report addressed to a paused manager goes to the user inbox.
• The child CLI ask command is rejected; children use the olw_ask tool. ask stays
  available to parents.
• update prepare rejects a stale check or one that isn't an upgrade, targets the selected
  GitHub remote with gh -R, reuses an existing remote pull request, and skips typecheck,
  test and build after the install step fails. --remote accepts GitHub remotes only.
• Repository mirror locks use a kernel flock that the spawned git inherits.

Removed

• The CLIProxyAPI model extension (dist/proxy/index.js), its /proxy-refresh
  command, access files and manual-first CLIProxyAPI policy document. Role
  launches and the shared host profile no longer load it. Remove that path from
  OMO's settings.json extensions. A routing receipt without a provider is
  still re-planned at the next start.

Fixed

• Upstream Sol Fast selectors are routed to standard-tier Sol through opencodex;
  Luna keeps its Fast row. Sol's priority tier cost too much for its speed gain.
• The managed omo launcher works with omo-ai installed through bun (5.0.0 and
  later). It finds the omo-ai package from its manifest instead of assuming a
  symlinked bin, so the generated bun launcher script no longer fails the routing
  preflight with ENOENT ... ~/.bun/package.json.
• Once routing has been adopted, a failed routing preflight no longer blocks the
  interactive omo launcher: it prints the error and starts OMO with the retained
  routing. Before adoption, and for OLW role launches, a failed preflight still
  stops the launch.
• pr merge matches the child's report to the requested PR instead of the latest
  report from that binding.

Security

• Explicit local-file copies are anchored to directory handles inside the owned checkout,
  reject C0 and C1 control characters at both path boundaries, and surface cleanup failures.
  Receipts store only a source label, never the source path.
• Git remote arguments get explicit option boundaries (--), so a remote name or URL can't
  be read as a flag.

Migration

Do these steps in order after pulling v0.2.0.

1. Rebuild the control root. Run bun run install:local from the checkout, or follow the
   upgrade steps in docs/policy/releases.md. Running
   Herdr servers and OMO sessions keep the old code until you close and recreate them.
2. Switch to official...

v0.1.0

Released: 2026-09-23

Initial source release of OMO Linear Workflow (OLW).

Added

• A non-overwriting Linux x64 installer for a source checkout and its olw launcher.

• CI for metadata, type checking, lint, tests, the managed Herdr build and installation smoke tests.

• Version/tag validation and source-only GitHub releases triggered by maintainer tags.

• Version, upgrade, rollback and contribution policies, issue forms and a PR template.

• Bun CLI (bun run cli) that imports a revision-pinned Linear initiative scope
  snapshot into local SQLite and creates supervisor, parent and child role
  bindings on top of it.

• Herdr-backed role workspaces: one supervisor per initiative, one parent
  worktree per project on an integration branch, one child worktree per issue
  based on its parent's branch.

• Event-driven delivery between roles through OMO native threads with
  persistent claims, idempotent message IDs and stored receipts. No polling
  loops.

• status, pause, resume, reconcile and close commands that inspect
  and release runtime resources while preserving worktrees, branches and
  session records.

• Managed Herdr runtime pinned by vendor/herdr/manifest.json and
  patches/herdr-0.9.1-omo.patch, built by bun run build with digest-verified
  artifacts and receipts. No fallback to a global PATH binary.

• Proxy model extension (dist/proxy/index.js) that registers CLIProxyAPI
  models in OMO, honours the manual-first model policy, and refreshes the
  catalog at startup, on new agents and on /proxy-refresh.

• Automatic upstream routing tracking (bun run proxy:routing) that maps the
  installed global OMO's category and agent chains onto proxy models while
  protecting manual overrides.

• Ollama Cloud DeepSeek routing through the proxy extension.

• OMO skills olw-define, olw-plan, olw-run and olw-check for defining,
  planning, executing and checking approved scope through the native runtime.

• Isolated QA scripts (qa:events, qa:linear, qa:proxy, qa:routing,
  qa:herdr) that run against owned fixtures and clean up after themselves.

• English and Korean READMEs, plus operator guides for the proxy model policy
  and automatic routing.

Known limits

• Linux x64 is the only supported installer target.
• Live Linear OAuth and real Linear writes have not been exercised.
• The OLW orchestration runtime does not automatically merge work or mutate Linear.
  GitHub source releases require a maintainer to push a version tag.