Skip to content
OrbitOrbitOrbitv0.37.0

Changelog

What changed in each Orbit release, newest first. Breaking changes come before highlights. Follow a pull request link to find the delivery on GitHub. The newest three releases are expanded; select an older version to read its notes.

Check for a newer release:orbit update --checkNew here? Install Orbit.

Changelog

0.37.0 Latest

Breaking Changes

  • Ship sweep removed: the orbit run ship-sweep command, the seeded ship-sweep routine and its job are gone, and workflow.auto_ship is deprecated and ignored. Unattended dispatch → use orbit run auto drains or orbit run ship, and delete the key. (Pull request)
  • Regression severity: regression_from links and review findings carry one five-level severity (nit to critical). Older binaries refuse the new field, so upgrade the owner and followers together. (Pull request)
  • Regression sites always recorded: workflow.regression_risk.record_sites is deprecated and ignored, because regression.sites are now always kept. Config that sets it → delete the key. (Pull request)
  • Store migrations: opening a workspace applies SQLite v69 (run_resource_minutes_freshness_index) and v70 (reopen_legacy_experiment_admissions). Both are additive but change persisted storage, so upgrade hosts together. (Pull request)

Highlights

  • Size-based riskiness: task regression risk is now fitted on footprint size (declared files before a diff exists, lines changed once one does), with the size-band evidence shown in place of driving files and a crew adjustment. (Pull request)
  • Experiments from the dashboard: Analytics › Experiments can create, edit and start experiments, and task.crew now records the crew that actually ran, including experiment arms and provider-hold redraws. (Pull request)
  • Severity-weighted scorecards: regressions and findings share one severity scale, and scorecards weight them by it. (Pull request)
  • Site-attributed hotspots: Analytics › Codebase counts a regression against the files its recorded sites name instead of the whole footprint, and a file view lists its regressed symbols. (Pull request)
  • Steadier landing: a failed check is rerun once before the candidate is blamed, and a handoff waits for the base fix after an inherited failure and reconciles PRs merged outside Orbit. (Pull request)

0.36.0

Breaking Changes

  • Resource-minute migration: opening a workspace applies SQLite v68 (resource_minute_regime). The migration is additive (two nullable columns older binaries never read) but changes persisted storage; upgrade hosts together. (Pull request)
  • Scheduled task field: v0.35.0 and older refuse unknown task fields, so they can’t read a task whose not_before is set. Mixed-version hosts → upgrade the owner and followers together. (Pull request)

Highlights

  • Scheduled tasks: a task’s optional not_before stops every dispatcher from starting it earlier, and the hold lifts by itself. orbit task list --scheduled and not_before range filters list upcoming work; the dashboard shows the schedule in task detail and the forecast. (Pull request)
  • Workspace presets: orbit workspace init --preset <file> applies a distribution’s TOML preset that picks which shipped routines and auto-tasks are seeded, enables plugin-seeded definitions and sets workspace config. The whole file is validated before anything is written. (Pull request)
  • Count-based windows: forecasts and per-task analytics can count the newest deliveries instead of a time period (last50, last200, last500, or orbit run profiles --accuracy --last N), so a quiet month no longer blanks them. (Pull request)
  • Dashboard: task detail shows the run’s branch and worktree path. Locks held lists locked files with the tasks holding them, marks paths several tasks share and filters by a task’s footprint. Fleet drains shows one tab per host. (Pull request)
  • Symbol-level footprints: task-pilot and the review auto-tasks declare symbol: selectors where they can, and a regression_from relation can name the defect’s sites, recorded once workflow.regression_risk.record_sites is on.

0.35.0

Breaking Changes

  • Lease queue migration: opening a workspace applies SQLite v67 (task_reservation_lease_queue). The migration is additive (three nullable columns older binaries never read) but changes persisted storage; upgrade hosts together. (Pull request)
  • Hard leases queue by default: an overlapping hard orbit.lease.request now waits in the lease queue (granted=false, queued=true, holding nothing) instead of being refused, and orbit.lease.list includes queued leases. Immediate refusal → pass wait=false (CLI --no-wait). (Pull request)

Highlights

  • Path lease queue and listener: hard leases queue in priority-then-age order with aging, orbit.lease.expand adds paths with deadlock refusal, and the lease listener streams notices over SSE, terminates TLS natively and rate-limits every request. The dashboard gains a Leases view. (Pull request)
  • Backlog forecast v2: the forecast simulates the drain’s real admission, including dependencies, file locks, dispatch order, CPU and memory fit, blocked, failed and retried runs, review modes and expected regression rework. The drain card shows it on one line. (Pull request)
  • Task riskiness: each task shows a regression riskiness score from its files’ regression history, with the top driving files, on Tasks rows and in task detail. (Pull request)
  • Claim recovery and bulk settling: orbit task recover-claim and orbit.task.recover_claim return a live follower claim to blocked or backlog, and the dashboard’s Settle merged PRs button settles every task whose delivery PR merged. (Pull request)
  • Auto-task last fire: orbit.auto_task.list reports each auto-task’s last fire time, minted task and outcome, and routine retrieval exposes the last-fire run id and step outputs. (Pull request)
  • Held, not failed: agent-declared task blocks, fail-closed pr_complete stops and required commands that fail identically on the base are recorded as held, so they no longer count as failed runs. (Pull request)

0.34.0

Breaking Changes

  • MCP task list: orbit_task_list returns compact rows by default instead of full records, and view=bounded always needs an explicit workspace. Full records → pass fields for description, plan or acceptance criteria. (Pull request)

Highlights

  • Experiments: Run randomized experiments over agent treatments, with sticky assignment at admission, instruction, tool, crew, config and skill overlays, and a per-arm readout with intervals and sample-ratio checks, from the CLI, MCP or Analytics › Experiments. (Pull request)
  • Path leases: Interactive sessions can take advisory or hard leases on paths through orbit lease, MCP and loopback REST, with a durable notice stream covering overlap, landing and expiry. (Pull request)
  • Granted tools: Operators can lift the implementer’s deny list for named read-safe tools with granted_tools, with classification at grant time, an owner-host placement hold, and mutation proposals applied after approval. (Pull request)
  • Explore brief: Task-pilot stores a revision-keyed code map for each task, and runs receive it, with a turns-to-first-edit series and an accuracy report. It runs in shadow mode for now. (Pull request)
  • Backlog forecast: Orbit forecasts drain time, cost and tokens from delivery history, records forecasts on a schedule, and shows them on the Tasks drain card. (Pull request)
  • Host labels and type crew pools: Tasks can require host capability labels through requires_labels and machine.labels, and workflow.type_complexity_crews routes crew draws by task type. (Pull request)

0.33.0

Breaking Changes

  • Job-run task recency migration: opening a workspace applies SQLite v59 (job_run_task_recency). The migration is additive (a new table, index and triggers older binaries never touch) but changes persisted storage; upgrade hosts together. (Pull request)
  • Cursor, Pi and OpenCode lanes retired: configuration and dispatch now reject the cursor, pi and opencode providers; recorded history still displays them. Crews on those providers → move them to a supported provider. (Pull request)
  • New task history decision: tag edits record a tags_changed decision, which v0.32.0 and older cannot read, and the distributed admission contract now carries review-skip tags. Mixed-version hosts → upgrade the owner and followers together. (Pull request)
  • orbit tool run --pretty removed: the hidden error-formatting alias is gone, so scripts passing it now fail. --pretty removed → drop the flag. (Pull request)

Highlights

  • Decision channel: Orbit records decision requests on held work and raises them through a new state trigger. Operators list and answer them from orbit decision, MCP or the dashboard, and registered plugins can submit authenticated answers. (Pull request)
  • Why stuck: a dashboard card links each blocked or held task to the run or finding that stopped it, with one-click requeue, settle-merged and pilot-admission actions. (Pull request)
  • Workspace delivery conventions: configure pull request titles, labels, commit messages and branch names per workspace so automated deliveries follow the repository’s conventions. (Pull request)
  • Review-skip tags and default review: workspaces can map a task tag to the paths whose changes skip agent review, and workspace init seeds before-landing review for PR-route workspaces. (Pull request)
  • Portable maintenance sweeps: new built-in auto-task definitions cover code health, architecture, structure, dependency freshness, flaky tests and CI budgets. (Pull request)

0.32.0

Breaking Changes

  • Crew performance snapshot migration: opening a workspace applies SQLite v58 (crew_performance_snapshot). The migration is additive (two new tables older binaries never touch) but changes persisted storage; upgrade hosts together. (Pull request)
  • Pull protocol fingerprint: follower pull requests now carry per-crew usage pressure, which changes the distributed-drain protocol fingerprint. Mixed-version pull drains → upgrade the owner and followers together. (Pull request)
  • Blocked-task triage auto-task retired: orbit workspace sync removes seeded copies of the blocked-task-triage auto-task, even enabled ones, and keeps edited copies. blocked-task-triage auto-task removed → set enabled: true on the blocked_task_triage routine. (Pull request)
  • workspace init --refresh-defaults removed: the hidden no-op flag is gone, so scripts passing it now fail. --refresh-defaults removed → drop the flag. (Pull request)
  • Rust API: AgentFamily::Muse and StateTriggerKind::{PilotWarningsHeld, BlockEscalated} are new variants; OrbitError::RemoteTool.code is RemoteToolCode, AutomationError::Deferred carries DeferredReason, and PolicyDecision is removed. Unused orbit-store and orbit-common items are deleted. Downstream exhaustive matches and constructors → update them. (Pull request)

Highlights

  • Usage-pressure crew selection: with workflow.crew_selection = "usage_pressure", draws favour crews whose weekly allowance would otherwise go unused, and admission re-draws system-chosen crews. Pull followers weigh their own readings; the dashboard and orbit run readiness show pressure per account and crew. (Pull request)
  • Performance-weighted crew draws: workflow.crew_performance weights draws by an hourly per-task-type implementer snapshot. orbit agents scorecard --by-type shows the cells, and --shadow compares recorded shares with what the factor would give. Off by default. (Pull request)
  • Escalation triage routines: new pilot_warning_triage and blocked_task_triage routines, disabled by default, send pilot-held proposed tasks and escalated blocks to a hard-tier triager that records one requeue, proposal or human-action report per episode. (Pull request)
  • Wider provider usage coverage: Grok, Antigravity and Muse runs now record usage-window readings, so provider limits, holds and usage pressure cover them, and Muse crews run on Linux hosts. An auto-task whose crew hits a usage limit is reassigned within its complexity pool before it runs. (Pull request)
  • Built-in argon crew: Gemini 4 Argon through agy is a built-in crew, and orbit init adds it to the hard complexity pool when agy is installed. (Pull request)
  • Dashboard refresh: Settings is split into per-topic tabs, task History reads as a timeline, Context files group by crate or folder, and provider usage chips open a detail dropdown. Edits to config.toml now show without restarting the server. (Pull request)

0.31.0

Breaking Changes

  • Executor census migration: opening a workspace applies SQLite v55 (executor_census). The migration is additive at runtime but changes persisted storage; upgrade hosts together. Rust callers must replace automatic_dispatch_cmp_for_host and AdmissionOrdering.owner_os with the affinity/census inputs. (Pull request)
  • Crew accounts migration: opening a workspace applies SQLite v56 (provider_crew_accounts). Environment-backed crews now have separate account identities → explicitly allow those crews when an allow-list previously relied on their provider’s default account. (Pull request)
  • Handoff storage compatibility: SQLite v57 moves handoff settlements into a sidecar and guards stale inline writers. Older binaries reopened afterward are read-only; already-running intermediate writers reject handoffs → upgrade and restart Orbit clients, including long-lived MCP servers, to resume writes. Rust settlement constructors and exhaustive matches also need updates. (Pull request)
  • Workspace source credentials: fresh workspace registration rejects Git origins containing embedded credentials; discovery redacts legacy values without rewriting registrations → use credential-free origins and a credential helper. (Pull request)
  • External-merge Rust API: ClaimMutation adds SettleExternalMerge; downstream exhaustive matches must handle the new variant. (Pull request)

Highlights

  • Crew-specific provider accounts: environment-backed crews have separate usage limits, holds, auth exclusions and budget targets. Crews without an environment retain the provider’s default account. (Pull request)
  • Scarcity-aware task admission: distributed and owner-local drains use a live executor census to prioritize work that fewer eligible hosts can run, with a workspace setting for affinity reach. (Pull request)
  • Drain controls: the dashboard composer covers local and pull drains, auto-run options, recorded settings and pull claim stages. (Pull request)
  • Task detail navigation: context files have a dedicated tab beside History, Runs, Artifacts and Comments, with the existing editor and keyboard navigation. (Pull request)

0.30.0

Breaking Changes

  • Distributed pull compatibility: the pull request fingerprint now covers browser-kit and forecast-fit admission fields, so an owner and followers on different builds refuse each other → upgrade every host together. (Pull request)
  • plugin.legacy_callback_identity removed: the legacy environment-token and process-ancestry callback authentication is gone, and the key now only warns → plugin backends must preserve the inherited callback descriptor. (Pull request)

Highlights

  • Read-only exploration: implementing and review-repair agents can ask an explorer crew bounded code-reading questions in the same worktree. The explorer runs read-only under a restricted tool policy and returns the paths it read with its answer. (Pull request)
  • Mid-run steering: an opt-in setting delivers operator comments added after dispatch into managed agent runs. Orbit reruns a bounded checkpoint and holds promotion until the run has consumed the new guidance, including on claimed work. (Pull request)
  • Agent scorecards: evidence-linked scorecards rank orchestrators, implementers, reviewers and pilots, with drill-down to the tasks and runs behind each score. The dashboard has role tabs, and orbit agents and orbit.agents.* expose the same scores. (Pull request)
  • Drain readiness and failure history: orbit run readiness and orbit run show name admitted and waiting tasks and why they wait, and orbit run history --failures reports each leaf failure once by root cause, with a link to its logs. (Pull request)
  • Host resource analytics: Analytics charts host CPU and memory profiles alongside forecast accuracy and measured delivery peaks, so operators can see where forecasts miss when tuning auto-concurrency. (Pull request)
  • Desktop fleet awareness: the desktop tray watches registered remote hosts and notifies when a drain ends or a task is held. Linux .deb and .AppImage bundles for x86_64 and arm64 are attached to the release. (Pull request)

0.29.0

Breaking Changes

  • orbit logs removed: the hidden top-level orbit logs <run-id> compatibility command is gone → use orbit run show for steps and orbit run logs for raw output. (Pull request)
  • orbit run ship-local removed: the hidden refusal-only alias now fails as an unknown command → use orbit run ship --mode local. (Pull request)
  • orbit tool list --include-hidden removed: the hidden alias no longer parses → use orbit tool list --all. (Pull request)
  • Scoreboard API reduced: GET /api/scoreboard now returns the GET /api/cost subset with deprecation headers, dropping agents, highlights, notable_completions and coverage, until its 2026-11-12 sunset → read /api/cost. scoring.enabled is ignored. (Pull request)
  • Forecast memory threshold must sit below the throttle: config load now rejects machine.auto_concurrency.admit_memory_percent (default 75) at or above workflow.resource_throttle.memory_high_percent, even with auto-concurrency off → keep the throttle above 75 or lower the admit percent. (Pull request)

Highlights

  • Muse Code executor: Meta’s Muse Code joins the built-in executors as the muse lane on macOS, with a seeded crew, orbit init detection and sandbox grants for its state directories. Existing crews and pools are unchanged. (Pull request)
  • Forecast-based admission: with machine.auto_concurrency.mode set to shadow or enforce, drains fit deliveries against measured CPU and memory profiles and shared host reservations. orbit run profiles --accuracy reports how good the forecasts were, and orbit run auto --auto-concurrency picks the mode per window. (Pull request)
  • Fleet drain overview: the dashboard’s All workspaces view shows every registered host and workspace’s drain, with phase, slots in use, eligible work and the most common wait reason. Clicking a row switches to that host and workspace. (Pull request)
  • Review for one run: orbit run auto|ship --review before-pr|before-landing adds a review layer to a single run without changing workspace policy. It can only strengthen the configured review, and run details show the layer used. (Pull request)
  • Auto-fix PR blockers: orbit run auto --auto-fix repairs a delivery PR’s failed required checks and branch conflicts through the existing recovery and validation path. It never merges; --complete stays the merge authority. (Pull request)
  • Post-deploy criteria stay owed: criteria that can only be measured on the deployed build are recorded as owed instead of failing the delivery, and repair handoffs or ordinary completion can’t drop them. (Pull request)

0.28.3

Highlights

  • Settings and Incidents redesign: Settings shows every key’s effective value, where it comes from and its built-in default, across the Effective, Workspace and Global views. Health › Incidents groups failures by a deterministic cause, with class tiles and a 24-hour timeline per incident. (Pull request)
  • Faster dashboard polling: polls reuse memoized scoreboard and diagnostics results, diagnostics payloads are bounded, JSON responses are compressed, and routine status no longer fetches from the remote, so typical polls drop from seconds to milliseconds. (Pull request)
  • Faster on Apple silicon: binary digests use hardware SHA-2 and the CPU baseline is shared, so each orbit process starts sooner and orbit run readiness drops from about 1.2 s to under 0.1 s. (Pull request)
  • Sturdier distributed claims: a failed claimed leaf keeps its uncommitted candidate, and Orbit builds the deferred-sandbox record itself. The owner captures its validation gates into the claim, and tests that also fail on the base, or pass when rerun alone, no longer block a candidate. (Pull request)
  • Review-finding verification: the finding_verification_pipeline job confirms or rejects review-filed bugs with file-and-line evidence. The opt-in review.verify_findings setting, off by default and toggleable in Settings, makes promotion and admission wait for a confirmed verdict. (Pull request)
  • System crew retired: system work (auto-task templates, the task pilot and recoveries) draws its crew from the complexity pools. workflow.system_crew and crew: system still load for one release with a deprecation warning → remove them. (Pull request)

0.28.2

Highlights

  • Extend a live drain: orbit run extend <run-id> --for <duration> moves a running drain’s deadline later without restarting it, so in-flight tasks, claims and authority carry over. --concurrency retunes it in the same call. It never shortens a window and refuses stopped or windowless drains. (Pull request)
  • Review crew pools: operation.review_crew accepts a pool such as ["sol", "grok"] (or weighted entries) as well as one crew. Each before-PR, before-landing or after-landing review draws one crew, preferring one that did not implement the work and skipping disabled or provider-limited crews. (Pull request)
  • Red base gets an owner: a baseline-red hold names the tests failing on the base in its hold text, verdicts and orbit run readiness. The owner’s clock tick files one deduplicated baseline-red-repair bug per failure set, and reviewer red-base claims replay only those tests. (Pull request)
  • One incident per failure: a gate or auto run that failed only because a child failed records that leaf as its root_cause, and orbit run history --incidents and the dashboard fold them into one row. Red-base holds now end held, and parents whose unsuccessful children were all cancelled end cancelled. (Pull request)
  • Dashboard redesign and codebase analytics: tasks show as status cards whose readiness dot says why backlog work waits, and drain status and controls get their own cards. Audit › Codebase replaces Audit › Policy and ranks files, directories or crates by regression rate, lock contention, footprint misses and blocked runs. (Pull request)
  • Linux worker containment: machine.worker_cpu_quota caps contained worker scopes with systemd CPUQuota=, and orbit doctor reports the limit. Pass-listed credentials reach workers through their environment instead of systemd-run arguments, so a process listing no longer shows provider tokens. (Pull request)

0.28.1

Highlights

  • Provider-limit-aware routing: drains and crew-pool draws skip crews whose provider is near or at its usage limit and resume them after the reset. orbit run readiness shows the limits and gated crews, and operators can declare rolling budgets for providers that report no usage. (Pull request)
  • Before-landing review: review.before_landing reviews a task’s open PR while hosted CI runs, and the PR lands only on an accepted review. It is an alternative to before-PR review and is off by default. (Pull request)
  • Dashboard host switcher: pick a registered host above the workspace picker, and every panel, action, log tail and resource chip follows it. The selection is URL-addressable as ?host=. (Pull request)
  • Run history by task: orbit run history shows each run’s task and duration and filters by --task, --state and --since, so you can find the run that delivered a task and how long it took. (Pull request)
  • Declared reclaim paths: worktree.reclaim globs declare which build outputs in kept run worktrees may be swept, replacing the hard-coded target/. (Pull request)
  • macOS .env deny hardening: renaming a containing directory out of the workspace no longer bypasses a **/.env sandbox deny on macOS. (Pull request)
  • Claude worker token on macOS: a Claude activity that receives neither CLAUDE_CODE_OAUTH_TOKEN nor ANTHROPIC_API_KEY is now refused before claude starts, instead of falling back to the Claude Desktop login, which the Desktop revokes mid-run (HTTP 401). Create a token with claude setup-token and list it in [execution.env] pass. For scheduled runs, the clock tick now exports credentials from ~/.orbit/clock.env (mode 0600) under the names pass admits, and orbit doctor’s claude-worker-token row warns when a routed Claude crew would run without the token. (Pull request)

0.28.0

Breaking Changes

  • Distributed pull protocol: the revision goes from 5 to 10, and a newer owner refuses older followers → run matching owner and follower versions. (Pull request)
  • Job-run state migration: SQLite migration 38 moves run state into job_run_states and drops job_runs.pipeline_state_json. It is declared Breaking, so older binaries refuse the migrated store. (Pull request)
  • terra is no longer a built-in crew: a config that names terra without defining [crews.terra] must define it. Existing [crews.terra] tables keep working. (Pull request)

Highlights

  • Host registry and routing: orbit host add|list|show|remove keeps a registry of named hosts. Task IDs route to the host that owns their prefix, --host selects one explicitly, and Settings › Hosts manages the registry from the dashboard. (Pull request)
  • Linux sandbox setup: Linux installs can use Orbit’s signed, pinned Bubblewrap when the host wrapper is missing, and package-manager discovery follows host capabilities. orbit init finishes with a readiness warning if preparation fails, and execution stays fail-closed until the sandbox is ready. (Pull request)
  • Faster run, dashboard and scoreboard queries: run listings no longer scan every run’s pipeline state, and scoreboard failure aggregation and cold dashboard paths are optimized, taking pages that took many seconds down to a fraction of that. (Pull request)
  • Distributed claimed review and recovery: claimed PR work runs the owner’s before-PR review contract and gets one automatic repair when landing stops on a stale or conflicting base. Operators can reconcile a merged PR whose final head differs from the candidate. (Pull request)
  • Proposed-task intake: orbit run auto --approve-proposed promotes qualifying proposed tasks and holds uncertain ones. The dashboard auto-drain panel offers it when starting a window, and proposed and backlog tasks show a readiness chip. (Pull request)
  • Built-in Haiku crew: fresh configs ship a Claude Haiku crew, and orbit init seeds complexity pools from the provider families it detects. (Pull request)

0.27.0

Breaking Changes

  • Review config split: before-PR review is review.before_pr; after-landing review is the delivery-code-review auto-task’s enabled flag and now covers follower landings. operation.review_policy and review_minutes still load with a warning, and review.minutes now limits one candidate’s review. (Pull request, Pull request)
  • delivery-qa auto-task removed: it is no longer seeded, and the integrated_qa_v1 coverage is refused for new or updated schedules. Existing copies still load. QA is covered by qa-sweep and qa-full-sweep. (Pull request)

Highlights

  • Final recovery for blocked tasks: a task that ends up blocked after its step recovery is used up now gets one last recovery agent, drawn from workflow.final_recovery_crews (default sol:100, opus:20). That agent can repair, requeue or resume the task instead of leaving it for an operator; [] turns it off. (Pull request)
  • Before-PR reviewer fixes what it finds: the reviewer commits its fixes as a second commit on the candidate instead of sending the task back for rework. Each candidate gets one review, limited by review.minutes. (Pull request)
  • Re-runs resume from a preserved candidate: re-running a task that has a [BLOCKED] candidate rebases and re-validates that candidate instead of implementing it again. (Pull request)
  • Host resource throttle: Orbit samples CPU, memory and disk, shows them as KPI chips and in a Settings › System tab on the dashboard, and holds drain and ship admissions while the host is under pressure. It is on by default, with warnings over the CLI and MCP. (Pull request)
  • Agent invocations you can wait on: orbit run agent --wait returns the agent’s structured answer (summary, findings, next steps), orbit run logs --follow streams output, and orbit_workflow_run_show reports progress while it runs. (Pull request)
  • Orbit closes its own PRs: when a task lands, is rejected or is archived, Orbit closes the task’s open delivery and [BLOCKED] PRs that it authored, keeping the PR that landed open. Human PRs are never touched. Turn it off with pr.close_on_terminal = false. (Pull request)

0.26.0

Breaking Changes

  • Public MCP tool surface consolidated: orbit_crew_list, orbit_friction_list, orbit_friction_rehome, orbit_auto_task_toggle, orbit_auto_task_delete, and orbit_workflow_run_workers are gone. Use orbit_workspace_list with include: ["crews"], orbit_search with kind: "friction", rehome_to on orbit_friction_update, enabled on orbit_auto_task_update, and orbit_workflow_auto with action: "resize". (Pull request)
  • Distributed-drain tools removed from MCP: orbit_task_pull and the orbit_drain_* probe, receipt, bind, and settle tools are no longer public, and calls to them are refused. orbit run auto --pull is unchanged. (Pull request)
  • Older-client desktop compatibility removed: the desktop MCP server no longer answers MCP peers that predate the guarded desktop contracts. Upgrade the desktop client to this release.
  • Desktop controls moved to guarded domain contracts: desktop task, auto-task, and drain controls go through guarded MCP contracts, and guarded orbit_task_* fields are refused on older MCP peers. Upgrade MCP peers with the server.

Highlights

  • Orbit mod for Claude Code: the Claude Code plugin adds a band above the prompt, a status line, and an Orbit pane with Board, Ship, and Map tabs, so you can act on tasks and watch a ship without leaving the session. (Pull request)
  • Desktop control center: the desktop app gains a control center with guarded task review, and its Tasks list groups tasks by status, defaults to active ones, and carries the task controls. (Pull request)
  • No bare agent spawns: dispatch now refuses when a shipped agent executor has no sandbox backend for the host OS instead of running it unsandboxed. (Pull request)
  • Disabled external tools stay manageable: a disabled external tool remains listed and inspectable, and can be re-enabled.
  • Workspace-safe dashboard: task edits, approvals, run actions, friction results, and plugin toggles made around a workspace switch no longer land on, or display under, the wrong workspace.

0.25.1

Highlights

  • Cancel stops the agent: orbit run cancel now stops the provider CLIs a run spawned, not just its owner, and orbit run ship refuses a task the pipeline cannot start from instead of reporting it submitted.
  • Symlinked homes and checkouts: an Orbit home or workspace reached through a symlink (macOS /tmp and /var, a symlinked ~/workspace) no longer breaks auto-task edits, plugin enablement, routine disables or workspace remove, and the plugin write-root containment check resolves .. as the kernel does.
  • Tighter sandboxing and secret masking: CLI workers inherit the sandbox on Linux and macOS, onboarding prepares the Linux sandbox prerequisites, and delivered plugin secrets and credential-bearing dotenv files are masked. (Pull request)
  • MCP annotations and field projection: tools/list advertises read-only, destructive and open-world hints so clients can auto-approve read-only tools, and orbit_task_list accepts fields to return only the fields you need.
  • Faster listings, search and dashboard: filtered task list, search, the dashboard scoreboard and auto-drain readiness read task metadata instead of whole task bundles, cutting them from hundreds of milliseconds to tens at about 1,500 tasks.
  • Clearer runs and drains: orbit run show summarizes a drain’s leaf outcomes and names the holder of a waited-on lock, blocked tasks name their next step, and every confirmation command honours --format json.

0.25.0

Breaking Changes

  • Plugin source root moved: plugin sources must live under .orbit-plugin/; a top-level plugin.yaml is refused by orbit plugin add, validate, test, sync, and upgrade, with no fallback. Move plugin.yaml and its tree into .orbit-plugin/ → orbit plugin scaffold creates the new shape. (Pull request)

Highlights

  • Follower pull: orbit run auto --pull lets a follower machine pull one owner-admitted task at a time, run it, and hand the result back; the owner keeps landing authority. (Pull request)
  • Plugin secrets and Python backends: plugins declare secrets that the host stores, delivers to backends, and rotates, and exec backends can be written in Python. (Pull request)
  • Brokered plugin calls and per-workspace enablement: sandboxed agents call plugin tools through a broker, and plugins are enabled per workspace from the CLI or dashboard. (Pull request)
  • New recurring auto-tasks: disabled-by-default documentation, run-failure, and backlog-hygiene auto-tasks, plus delete and restore for shipped defaults with a durable opt-out. (Pull request)
  • Deny-list agent policy: shipped agent activities now name the operations they deny instead of allow-listing tools; custom allow-list activities keep working, and strict worker containment is opt-in. (Pull request)
  • Safer upgrades: upgrade admission checks store and layout compatibility, so an older binary can read but no longer write a store migrated by a newer one. (Pull request)

0.24.0

Breaking Changes

  • Activity, executor, and policy commands removed: orbit activity, orbit executor, and orbit policy are gone with no aliases and now fail as unknown subcommands. job, tool, and plugin only moved in help. (Pull request)
  • Operation mode removed: the orbit operation CLI group, orbit run auto --grant, and the orbit.operation.* tools are gone. Store migration v26 drops operation grants, so older binaries fail against an upgraded store. (Pull request)
  • Docs command and tools removed: orbit docs and the five orbit.docs.* tools are gone. A leftover [docs] config section is ignored with a one-time warning → delete it. (Pull request)
  • Host renamed to machine: orbit host is removed; ~/.orbit/host.toml moves into [machine] in global config.toml (copied automatically this release), and host fields become machine_name / job_run_machine. Migration v23 renames audit columns, so older binaries fail against an upgraded store. (Pull request)
  • Semantic and hybrid search removed: orbit semantic, orbit search similar, --hybrid, and the semantic / hybrid inputs on orbit.search are gone. Lexical orbit search / orbit.search is unchanged. (Pull request)

Highlights

  • Plugin standard: plugin.yaml v2 plugins install and manage through orbit plugin, run under enforced grants in a sandboxed exec or mcp backend, and can ship tools, jobs, routines, auto-tasks, skills, and a schema-derived orbit <ns> <verb> CLI. (Pull request)
  • Distributed drain v1: drains across machines admit work atomically through durable claims with replayable receipts, pull and bind leaves uniquely, and hand results back to the owning workspace. (Pull request)
  • Complexity crew pools: a new xhard tier, weighted name:weight pool entries, and pool routing for ordinary orbit run ship admission; the drawn crew is fixed when the task is created. (Pull request)
  • State-triggered task-pilot: task-pilot runs as a preparation_eligible state routine and batches due tasks into one run instead of polling on a cron. (Pull request)
  • Dashboard Config tab and runnable jobs: the dashboard shows effective config grouped with provenance and inline edits, and the Jobs pane can start runs. (Pull request)
  • Bounded worker runs: on Linux, pipeline worker runs execute inside a bounded cgroup, so a runaway run cannot run the host out of memory. (Pull request)

0.23.0

Breaking Changes

  • MCP task-write response projections: orbit.task.add, orbit.task.update, and orbit.task.reject no longer include comments and history by default; request those fields explicitly with fields or field. (Pull request)

Highlights

  • Federated semantic search: search fans out concurrently across workspaces, reuses one query embedder, and batches embedding and reindex work for faster large-corpus queries. (Pull request)
  • Workflow failure recovery: agent-declared failed envelopes now enter step recovery instead of sending otherwise recoverable tasks directly to blocked. (Pull request)
  • Safer process supervision: execution cleanup validates process-group identity before signaling, preventing PID reuse from killing an unrelated process group. (Pull request)
  • Reliable worktree cleanup: garbage collection can continue cleaning stale worktrees whose large build directories exceed a single cleanup window. (Pull request)
  • Actionable job diagnostics: strict job catalog reads now surface malformed job files instead of silently hiding them from listing, inspection, and execution. (Pull request)

0.22.1

Highlights

  • orbit run task-pilot: named entrypoint for task_pilot_pipeline, matching orbit run triage for on-demand discovery or explicit task IDs. (Pull request)
  • make install targets ~/.orbit/bin: the source-checkout install now lands in the same directory as install.sh (override with INSTALL_BIN_DIR), so a stale ~/.cargo/bin/orbit can no longer shadow the deployed binary on PATH.
  • Dashboard: force any task status: a human can force any task to any status from the dashboard, matching the CLI’s --force escape hatch. (Pull request)
  • Forward-compatible store opens: Orbit no longer hard-fails when a workspace’s store schema or layout version is newer than the running binary supports; it opens read-only or auto-migrates instead. (Pull request)

0.22.0

Breaking Changes

  • Task start and approve tools consolidated: orbit task start CLI command and orbit.task.start / orbit.task.approve MCP tools are removed in favor of orbit.task.update. (Pull request)
  • Operation tools withdrawn from MCP: all five orbit.operation.* tools are removed from the MCP surface; operations are managed via the Orbit CLI. (Pull request)
  • MCP task listing response shape updated: orbit.task.list returns { tasks, total, truncated } instead of { items }; update MCP callers to consume the tasks array. (Pull request)
  • Routine hosts and source role configuration removed: routine hosts: pins and [routines] role = "source" configuration schemas are removed in favor of host clock consolidation. (Pull request)

Highlights

  • Dashboard inline task editors: the expanded task detail edits complexity, description, tags, acceptance criteria, and context files in place, each saved as a single-field PATCH /api/tasks/:id with inline errors and an allow-missing-context escape. (Pull request)
  • Unified host clock consolidation: orbit clock tick evaluates routines and auto-tasks in a single host loop, retiring separate auto-task schedulers and simplifying multi-host execution. (Pull request)
  • Owner-wins cross-host task sync: cross-host synchronization safely imports foreign-prefix task bundles while preserving local task state and preventing overwrites. (Pull request)
  • Delivery auto-task reset and self-healing: added auto-task reset and automated replay proof for diverged history recovery instead of silent per-tick deferrals. (Pull request)
  • Effective provider sandbox visibility: orbit.agent.invoke surfaces the active execution sandbox and warns when running under full-access mode. (Pull request)

0.21.0

Breaking Changes

  • Layout-v3 checkout task projections removed: .orbit/tasks symlink projections and projection fields are removed in favor of canonical storage; stop older binaries before running layout migration v3 to prevent recreating obsolete links. (Pull request, Pull request)
  • Worktree GC activity schema updated: custom worktree_gc invocations and automations must supply target_run_id instead of run_id for single-run scoping. (Pull request)
  • Public config path Result API: OrbitRuntime::config_path now returns Result<PathBuf, OrbitError> instead of an infallible PathBuf; callers must handle config validation errors. (Pull request)
  • Linux runtime sandbox requires descriptor mounts: Linux runtime write sandboxes now require Bubblewrap --bind-fd support; environments lacking descriptor-binding cannot mount writable runtime roots. (Pull request, Pull request)
  • Task-pilot admission contract enforced: automatic and explicit implementation admission now withhold unassessed tasks until complexity is evaluated, while no-diff-expected tasks are exempt. (Pull request, Pull request, Pull request)
  • Task complexity required on update: human and agent task.update surfaces now reject complexity: unassessed with the create-time validation error; only automated system pipelines may assign unassessed status. (Pull request)
  • Strict new-file delivery declarations: task delivery now enforces explicit new-file intent declarations and rejects unknown untracked paths before index mutation; declare new files via file: selectors. (Pull request)

Highlights

  • Read-only observational Orbit modes: SQLite databases can be safely observed without write locks even when uncheckpointed WAL frames exist, and workspaces can start without requiring semantic indexing. (Pull request, Pull request, Pull request)
  • Rust path-validation and descriptor hardening: file inspection and sandbox directory creation hold file descriptors through verification and mount operations, closing race conditions across runtime roots. (Pull request, Pull request, Pull request, Pull request)
  • Task-pilot automated repair assessment: code-scanning tasks are minted without speculative selectors, while task-pilot evaluates repair complexity and persists modification targets atomically before implementation. (Pull request, Pull request)
  • CLI and MCP surface reliability: fixes across tool execution, workspace filtering, MCP caller authorization, routine naming, credential redaction, and CLI output consistency improve agent ergonomics. (Pull request, Pull request, Pull request, Pull request)
  • End-to-end full-QA coverage: comprehensive QA verification suites, deterministic process fixtures, and browser capability evidence retention ensure stability across supported platforms. (Pull request, Pull request, Pull request, Pull request, Pull request)

0.20.0

Breaking Changes

  • Task workspace flags clarified: task add --workspace now selects the target workspace; use --workspace-path for task workspace-path metadata. (Pull request)
  • Tool-run flags removed: orbit tool run no longer accepts --timeout or --output; remove those flags from existing invocations. (Pull request)
  • Operation grants upgraded: operation-policy version 2 introduces independent review and repair controls; replace version-1 grants before delivery. (Pull request)
  • Task-registry schema upgraded: task action keys require schema version 6; older binaries that only support version 5 cannot open an upgraded registry. (Pull request)
  • Semantic-index layout migrated: semantic search now uses chunk storage with external-content FTS5; coordinate runtime upgrades because older index layouts are incompatible. (Pull request, Pull request)

Highlights

  • Independent delivery review: configure review timing and repair budgets separately from implementation, with direct repairs and explicit delivery admission. (Pull request)
  • Flexible task lifecycle: reopen completed work and make other explicitly authorized status transitions while retaining the audit trail. (Pull request)
  • Shared-root workspace isolation: multiple repositories sharing an Orbit root keep task writes and bundles in the selected workspace. (Pull request)
  • Usable dashboard operations: authorized sessions can mint and toggle automation, with compact controls that work on narrow screens. (Pull request, Pull request)
  • Filesystem policy for child processes: Linux proc.spawn applies the declared filesystem policy to indirect child-process access. (Pull request)

0.19.0

Breaking Changes

  • Task-pilot activity schemas migrated: custom prepare_task_pilot, task_pilot, and apply_task_pilot_results payloads must use the current required fields and succeeded|failed status contract. (Pull request, Pull request)
  • CI-failure filing schema migrated: custom file_ci_failure_tasks producers must supply pilot partitions and match evidence under the current activity contract. (Pull request)

Highlights

  • More provider lanes and effort controls: crews can forward validated effort settings, and Pi, OpenCode, and Antigravity join the supported executor catalog. (Pull request, Pull request, Pull request, Pull request, Pull request)
  • Task artifacts render safely: retrieve task attachments through the new read-only tool, with bounded presentation for validated raster images. (Pull request)
  • Safer source-remote changes: inspect or rebind a workspace source remote while preserving workspace, task, and checkout identity. (Pull request)
  • Operator-controlled agent investigation: authorized operators can invoke a bounded host-side agent run for diagnostics without changing task delivery state. (Pull request)
  • Constellation-works distribution cutover: active distribution references and the pre-release checklist now point to the current project owner. (Pull request)

0.18.0

Breaking Changes

  • SSH MCP caller authorization requires a protected launcher: orbit mcp callers authorize now requires --launcher and no longer falls back to the current executable; configure the validated protected launcher before authorizing callers. (Pull request)

Highlights

  • Accurate task-flow lifecycle reporting: orbit task flow now derives outflow and boundary openness from recorded transitions, including rejected tasks that later reopen. (Pull request)
  • Safer direct lock reservations: direct task-lock reservations canonicalize in-workspace selectors and reject outside-workspace paths, so they reliably conflict with task-scoped locks. (Pull request)
  • Selected roots reach sweep execution: orbit sweep now honors an explicit --root or ORBIT_ROOT, including when the process HOME points elsewhere. (Pull request)

0.17.1

Highlights

  • Managed workspace convergence: orbit workspace sync safely refreshes shipped routines and other managed artifacts while preserving local edits, identity, registration, and automation choices. (Pull request)
  • Security finding automation: the disabled-by-default security sweep can collect Dependabot, Code scanning, and secret scanning findings into deduplicated, evidence-complete remediation tasks without exposing credentials to agent sandboxes. (Pull request)
  • Corrective work prioritized automatically: bounded run ship and run auto discovery now selects critical, bug, code-review, and security-review tasks ahead of ordinary feature work. (Pull request)
  • Actionable friction triage: the dashboard defaults to active frictions, adds status filtering, and provides keyboard-operable responsive detail views. (Pull request)

0.17.0

Breaking Changes

  • session_log tools withdrawn: orbit.session_log.append, orbit.session_log.list, and orbit.session_log.resolve are gone from the public MCP/CLI agent surface → no public replacement. (Pull request)
  • ci-failure-remediation auto-task removed: the shipped ci-failure-remediation auto-task, catalog entry, and seed file are gone; existing workspace copies stay until removed by hand. (Pull request)
  • code-review auto-task renamed: the seeded code-review-sweep auto-task is now code-review; the old name no longer loads. (Pull request)
  • Remote MCP --operator ignored: destination-side caller authorization no longer honors caller --operator argv; a missing ~/.orbit/mcp-callers.toml yields agent only. (Pull request)

Highlights

  • Agent plugins restored: Claude Code, Codex, and Cursor plugin distribution ships again after the 0.15.0 retirement. (Pull request)
  • Federated MCP includes local workspaces: federated serve now includes the accepting machine’s local workspaces automatically, so a missing destinations file is a valid local-only server. (Pull request)
  • Task publication V1: operators can bind a private Git publication repo, publish owner-only snapshots, inspect labelled read-only state, and restore without renumbering. (Pull request)
  • Managed workspace routing: managed CLI and MCP calls now keep the catalog workspace selector instead of inferring from a linked-worktree cwd. (Pull request)
  • Richer task.show projections: orbit.task.show can project every stable public task field, including relations, job run, and attribution. (Pull request)

0.16.0

Breaking Changes

  • Replica task writes report capability refusal: orbit.task.add on a replica now returns capability_refused instead of invalid_input; update clients that branch on the old error code. (Pull request)
  • Replica MCP access is capability-gated: workspace-scoped control-plane reads and writes now return capability_refused on replica checkouts; route those calls to the owner selector. (Pull request)

Highlights

  • Federated multi-host MCP: register an opt-in federated server, discover remote workspaces over configured SSH destinations, and route workspace-scoped calls with host-qualified selectors. (Pull request)
  • Cross-workspace search: one CLI or MCP search can query multiple registered workspaces, fuse ranked results, and attribute every hit to its source workspace. (Pull request)
  • Official MCP Registry launch: versioned registry metadata and the npm proxy now provide a validated, fail-closed path from registry installation to an unbound Orbit MCP server. (Pull request)
  • Explicit docs-root overrides: configured docs roots can opt out of gitignore filtering while retaining Orbit’s .git and .orbit exclusions. (Pull request)
  • Policy-bound process spawning: activity proc.spawn calls now honor the child-environment allowlist and the activity’s filesystem sandbox. (Pull request)
  • Private SQLite state: Orbit creates and repairs its state databases, WAL files, sidecars, and sensitive directories with private Unix permissions. (Pull request)

0.15.0

Breaking Changes

  • Agent plugin channel retired: Orbit no longer ships Claude, Codex, or Cursor agent plugins; install and use the standalone Orbit CLI instead. (Pull request)

Highlights

  • Built-in code-review sweep: fresh workspaces include a disabled code-review auto-task that teams can enable to schedule recurring reviews. (Pull request)
  • Reliable concurrent task updates: task writes now coordinate safely across simultaneous CLI, MCP, and dashboard activity. (Pull request)
  • Simpler binary-first onboarding: the Quick Start now leads with installing the standalone CLI and brings your existing repository conventions into Orbit. (Pull request)

0.14.0

Breaking Changes

  • Fresh installs no longer seed the legacy qa crew: new orbit init workspaces use the supported system crew; update prompts that name the seeded qa crew. Existing user-authored qa configuration remains compatible. (Pull request)
  • proc.spawn activities now require an allowlist: v2 activities that grant proc.spawn must declare proc_allowed_programs; add the allowed programs explicitly. (Pull request)

Highlights

  • Cursor Agent Plugin support: Orbit now ships as a Cursor-compatible Agent Plugin alongside the Claude and Codex plugin formats. (Pull request)
  • More managed CLI executors: GitHub Copilot CLI and Cursor Agent CLI are first-class workflow executors, with improved installed-agent detection. (Pull request)
  • Operator-authorized MCP setup: orbit workspace init --mcp can generate MCP configuration for a workspace. (Pull request)
  • Dashboard workspace shell: the dashboard gains a command rail, two-mode dock, and log status bar for faster operator navigation. (Pull request)
  • Security-review automation: fresh workspaces include a disabled security-review auto-task that teams can enable when ready. (Pull request)

0.13.0

Breaking Changes

  • Store schema 15 → 16: audit actor identity splits role into actor_kind plus a normalized actor. Workspaces migrate on open; a newer store still refuses an older binary. (Pull request)
  • Store schema 16 → 17: unauthenticated MCP calls persist a self-reported actor. Same migrate-on-open / no-downgrade rule. (Pull request)
  • Task complexity is required: CLI task add, web POST /api/tasks, and orbit.task.add reject a missing complexity — pass low, medium, or hard. (Pull request)

Highlights

  • Dashboard delivery controls: inspect, toggle, and mint auto-tasks; operate routines and the sweep clock; the Tasks view is clearer with safer live-log and inline edits. (Pull request)
  • Complexity is first-class: every new task must set it, and the dashboard charts it as a dimension. (Pull request)
  • Governed MCP tools need an operator: unauthenticated sessions can no longer invoke those tools. (Pull request)
  • Claude Keychain OAuth under sandbox: Claude can read the user Keychain for OAuth; an explicit denyRead still wins. (Pull request)
  • System crew is seeded: jobs can name a portable system crew; qa stays for compatibility. (Pull request)

0.12.1

Highlights

  • Release cut: version files and the GitHub tag aligned at 0.12.1. (Pull request)

0.12.0

Breaking Changes

  • Seeded skill directories retired: orbit-task, orbit-search, orbit-workflow, and the plugin orbit-task-pilot skill are gone — use the orbit router skill and its references/. (Pull request)

Highlights

  • One Orbit skill: four seeded skills collapsed into a single orbit router with on-demand references. (Pull request)
  • Stale skill links cleaned on upgrade: workspace init removes dangling Claude/Codex skill symlinks left after skill retirement. (Pull request)
  • Doctor reports leftover skill directories: orphaned skill dirs without SKILL.md are a warning, not a clean bill of health. (Pull request)
  • Safer artifact redaction: env-value redaction no longer silently eats ordinary words in task prose. (Pull request)

0.11.0

Breaking Changes

  • Store schema 11 → 15: workspaces migrate on open (friction SQLite, workspace-claim scope, learning-projection removal, invocation audit context). A newer store still refuses an older binary. (Pull request)
  • Native project learnings removed: orbit learning, orbit.learning.*, and learning search are gone — use tasks, docs, feature ADRs, and frictions. (Pull request)
  • Native ADR store and tools removed: orbit adr / orbit.adr.* and .orbit/adrs/ are gone — author decisions in each feature’s 4_decisions.md. (Pull request)
  • Duel, independent review, and role: retired: planning-duel and review CLI/MCP/jobs are gone; activity YAML role: no longer loads — use crew: or the run-resolved crew, and ship via task_pr_pipeline. (Pull request)
  • Jobs are asynchronous by default: orbit run job returns a durable run immediately; --backend http|cli|auto is gone — use agent_cli and pass --wait when you need to block. (Pull request)
  • MCP and fs.* builtins trimmed: auto-task MCP advertises only list+mint; fs.read, fs.delete, and remaining fs.* tools are gone — use CLI/orbit tool run for auto-task authoring and provider-native file tools. (Pull request)

Highlights

  • Asynchronous job execution: job submission returns a durable run immediately, with --wait and terminal diagnostics when you need to block. (Pull request)
  • Epic-owned continuous delivery: an epic run drains descendant work through one owned worktree and lands the epic branch only after the epic’s own destination checks pass. (Pull request)
  • One workspace selector grammar: CLI and MCP accept the same registered name, logical ws_* id, or absolute checkout path. (Pull request)
  • Network MCP transports: orbit mcp listen and related tunnel/loopback paths serve isolated sessions with explicit workspace routing and claim-gated remote execution. (Pull request)
  • Fail-closed Linux sandbox: Bubblewrap plus policy-derived grants give Linux CLI agents and task-pilot a read-only, fail-closed execution path. (Pull request)

0.10.0

Breaking Changes

  • Task locks returned to task administration: orbit locks is now orbit task locks. This partially reverses ORB-10000: unlike the other removed task verbs, which were consolidated into existing verbs, locks were only relocated to a more prominent top-level surface. Moving them back completes the original surface-reduction intent without changing the MCP tool surface. (Pull request)
  • Learning vote and comment surfaces removed: orbit.learning.upvote and all orbit.learning.comment.* tools and CLI subcommands are gone — use priority + search rank for ranking, update/supersede for corrections, and evidence for provenance. (Pull request)
  • orbit task trimmed from 19 to 12 subcommands: seven state-transition and read-projection verbs removed. approve/reject/unarchive/delete → orbit task update --status …; locks → orbit locks; prune-context → orbit task lint --fix; templates dropped. MCP tool surface unchanged. (Pull request)

Highlights

  • Graph-tool guidance lives only in the MCP surface: removed graph-tool invocation instructions from activity prompts and deleted the orbit-graph skill; graph tools self-describe when present, and essential selector/usage knowledge now lives in the orbit.graph.* tool descriptions. (Pull request)
  • QA sweep is an auto-task: the checked-in six-hour definition routes hands-on validation to crews.qa, dedupes open runs, and files findings through Orbit; no-diff-expected exempts side-effect-only tasks from workflow diff gates. Legacy QA code, config, CLI, routines, and deployment units are removed. (Pull request)
  • Auto-task primitive: recurring chores are data, not code. orbit auto-task add/list/show/update/toggle (+ orbit.auto_task.* tools) define .orbit/auto_tasks/*.yaml templates with a cron/interval schedule and dedupe policy; one seeded scheduler routine mints tasks from the due ones with catch-up collapse and skip_if_open dedupe. Provider-neutral (ADR-0217). (Pull request)
  • Routine scheduler health over HTTP: GET /api/routines on the dashboard exposes each routine’s last fire (timestamp, ok/error outcome, duration) and next-due slot, so a stopped scheduler is visible remotely as a stale last_fire without box ssh. (Pull request)
  • ADR create over HTTP: POST /api/adrs records a Proposed ADR (mirroring orbit.adr.add), so remote orchestrators can author decisions without an on-box run; malformed payloads get a structured 400. Friction/learning create routes stay absent. (Pull request)
  • Learning + ADR update routes on the HTTP API: PATCH /learnings/:id (summary/scope/tags/body/evidence/priority) and PATCH /adrs/:id (status/tags and mutable metadata) delegate to the CLI tools, preserving supersede-don’t-delete and invalid-transition rejection; no create routes added. (Pull request)
  • Default failed-run triage: a seeded task_triage_pipeline (+ hourly routine, orbit run triage) diagnoses tasks blocked by failed runs, re-backlogs environmental failures with a bounded retry budget, and leaves the rest blocked with a diagnosis attached. See ADR-0215/ADR-0216. (Pull request)
  • Orphaned pending job runs now reconcile to interrupted: workers claim their queued run’s pid, the open-time orphan scan and orbit doctor cover pending runs, and orbit run cancel <run_id> terminalizes stuck runs. (Pull request)
  • Remote dashboard task actions now reach the selected workspace: approve/reject/archive used a raw fetch() that skipped the workspace-routing helper, so they silently no-op’d against a non-default remote workspace; the workspace selector also no longer renders its filesystem path beneath the dropdown. (Pull request)
  • Codex init default avoids nested sandboxing: fresh Orbit configurations now seed Codex with danger-full-access, leaving Orbit and the host as the execution boundary. (Pull request)
  • Crews are one provider-model assignment: [crews.<name>] now uses flat model, provider, and backend fields shared by every activity role; legacy three-role crews load through their implementer assignment. See ADR-0213. (Pull request)
  • orbit web serve defaults to global (multi-workspace) mode: the all-workspaces dashboard dropdown is always present; --global is now a deprecated no-op. (Pull request)
  • Versioned .orbit/ upgrade path — orbit migrate: the whole workspace layout is now versioned via a state/layout.version marker and an ordered migration registry; older layouts auto-migrate on open, newer ones refuse to open. orbit migrate [--dry-run] reports and inspects. (Pull request)
  • Policy deny rules can no longer be dodged by respelling a path: glob-path normalization rebuilds paths from their normal components so secret/./key.txt, secret//key.txt, and trailing-slash spellings all match rules for secret/key.txt; adds Linux enforcement integration tests and a CI sandbox/policy job. (Pull request)
  • Task migration between machines: new orbit task export/import pack and restore task bundles as a portable tar.zst with integrity validation and collision renumbering; orbit task reindex rebuilds the index from bundles, and --task-id-start seeds disjoint id ranges. (Pull request)
  • orbit web connect <ssh-host>: view a remote workspace’s dashboard over an auto-managed SSH tunnel — picks a free port, starts remote serve, forwards, opens the browser, and tears down on Ctrl-C. (Pull request)
  • Global, multi-workspace dashboard: orbit web serve outside any workspace (or with --global) serves every registered workspace via a workspace-keyed runtime map, with a header selector and an aggregate “All workspaces” task view. (Pull request)
  • Dashboard shows each workspace’s filesystem location: the global dashboard surfaces where each workspace lives on disk (home-abbreviated) in the header selector and task Details box, disambiguating same-named workspaces. (Pull request)
  • Aggregate dashboard view no longer 400s: an isAggregateView() predicate gates per-workspace fetches so the “All workspaces” view fetches only the aggregate task list and shows a placeholder for workspace-specific panels instead of erroring. (Pull request)
  • Versioned store-schema migrations: the store SQLite DB tracks numbered migrations in a schema_meta ledger, applies each atomically, adopts pre-ledger DBs via an idempotent baseline, and refuses to open a DB written by a newer orbit. (Pull request)
  • Job-run checkpoint/resume for hard interruptions: the v2 executor persists a recovery checkpoint after each completed step; dead-owner runs finalize to a new interrupted state, and orbit job resume <run_id> replays only the remaining steps. (Pull request)
  • Always-on operability: detailed healthz and orbit doctor: /healthz?detailed=true runs per-check probes, and orbit doctor runs workspace pass/warn/fail health checks. (Pull request)

0.9.2

Highlights

  • orbit graph is back as a thin CLI wrapper: the v2 graph is now reachable from the single orbit binary via orbit graph {sync, search, show, refs, callees, impact, trace, overview, implementors, deps, version, db-path, clean}, not only the standalone orbit-graph-cli binary or the in-process MCP adapter. orbit-graph-cli is lib-ified (lib + bin) so both front ends share one command layer with no duplication; the agent-facing graph surface is unchanged (still MCP-only). This amends the ADR-0198 consequence that dropped the subcommand. (Pull request, ADR-0199)
  • Task store crash durability hardened: the task registry SQLite connection now runs with synchronous=FULL, and task-bundle creation fsyncs the bundle directory’s parent before returning — so a power loss or OS crash mid-write can no longer leave the task store or a freshly-created bundle in a torn, unrecoverable state. (Pull request, Pull request)

0.9.1

This release completes the orbit-graph v2 migration — v1 (orbit-knowledge) is fully decommissioned and v2 is now the only graph surface — and adds a stable protocol for plugging in out-of-process executors.

Highlights

  • orbit-graph v2 is now the sole graph surface: the orbit-knowledge (v1) crate and its orbit.graph.* builtin tools are removed, the orbit graph CLI command is dropped (the v2 graph ships as the standalone orbit-graph-cli binary plus an always-on in-process MCP adapter), and orbit init no longer builds a graph up front — v2 syncs on demand. The MCP graph surface is now search / show / refs / callees / impact / trace / overview / implementors / deps; pack is gone and callers is folded into refs. (Pull request, Pull request)
  • External Executor Protocol v1: register a homegrown agent or CLI as an Orbit executor without forking core, through a documented stdin/stdout envelope contract and dynamic out-of-process registration. (Pull request)
  • Graph reads are faster and more accurate: queries are index-backed and decoupled from a full-worktree rescan, show returns UTF-8 text instead of a raw byte array, trace resolves far more command handlers, and refs/impact fall back to fuzzy-name matches when the precise floor is empty. (Pull request, Pull request, Pull request, Pull request, Pull request)

0.9.0

This is a security-hardening release: the bulk of the changes close secret-leak, sandbox-escape, network-exposure, and workspace-containment issues found in a focused audit (ORB-00355–ORB-00374).

Breaking Changes

  • Shell activity/job type removed: the v2 type: shell activity surface is deleted end to end — a workspace activity or job declaring type: shell now fails to deserialize at load instead of executing. The shell program allowlist was self-asserted within the same workspace-supplied YAML (the gate was a tautology) and run_shell spawned with no OS sandbox, cwd confinement, or policy consultation. The shell_reference asset and the pure-shell job fixtures/demos are removed; use the sandboxed agent-CLI execution path instead. (Pull request, Pull request)
  • [execution.env] inherit config key removed: an auto-discovered workspace config.toml could set inherit = true and — because workspace config replaces global — flip agent subprocess spawning to full environment inheritance, leaking the orbit process’s API keys and tokens into spawned agents. Env inheritance is now fixed to false and is no longer configurable; only the pass allowlist remains tunable. Stale inherit keys are silently ignored. (Pull request)

Highlights

  • Secret redaction hardened across audit and logs: redaction now covers non-UTF-8 audit blobs (previously bypassed entirely), error/bytes fields recorded via tracing, and Gemini’s HTTP transport no longer places the API key in the URL query string. The pattern redactor also learned major Google / SCM / cloud key formats it previously missed. (Pull request, Pull request, Pull request, Pull request)
  • Sandbox and execution confinement tightened: the macOS sandbox profile no longer grants unrestricted file-read* plus network*, an SBPL (?i) inline-flag bug that disabled sandbox-exec for all macOS CLI runs is fixed, the run_command automation action no longer shell-interpolates rendered templates, and unbounded child stdout/stderr capture (a memory-exhaustion DoS) is now bounded. (Pull request, Pull request, Pull request, Pull request)
  • Dashboard network exposure closed: orbit web serve now refuses to bind any non-loopback host (only 127.0.0.0/8 and ::1) — the unauthenticated read/write API can no longer be exposed to the network via --host 0.0.0.0; bind loopback behind an authenticated tunnel/reverse proxy for remote access. Stored XSS from untrusted task/learning/ADR/friction content rendered in the dashboard is also fixed. (Pull request, Pull request)
  • Secret-file policy matching fixed: denyRead/denyModify globs now match dotenv-variant secret files (.env.local, etc.), policy glob matching is no longer case-sensitive on case-insensitive filesystems, and audit blobs and JSONL logs are no longer written world-readable. (Pull request, Pull request, Pull request)
  • Workspace containment enforced: a malicious workspace can no longer override seeded .orbit/resources/{jobs,activities}/ assets, and graph operations stay confined to the worktree even when an MCP session omits the workspace. (Pull request, Pull request)

0.8.2

Fixes

  • Intel-macOS semantic companion dropped from the release: orbit-search-companion cannot link on x86_64-apple-darwin — ort-sys (2.0.0-rc.12) ships no ONNX Runtime prebuilt for Intel macOS. The Intel-mac release leg now builds the orbit CLI only, and the release publishes companion binaries for macOS-arm64, Linux x86_64, and Linux arm64 only. The Intel-mac CLI tarball still ships; semantic search is unsupported on Intel Macs (it never successfully shipped). Together with the glibc fix in 0.8.1, this completes recovery of the v0.8.0 release, whose GitHub Release never published while @orbit-tools/cli@0.8.0 was already on npm. (Pull request)

0.8.1

Fixes

  • Linux release build of the semantic companion: orbit-search-companion failed to link on the release runners — its prebuilt ONNX Runtime (via ort / fastembed) references glibc 2.38+ symbols (__isoc23_*) that don’t exist on ubuntu-22.04 (glibc 2.35), failing both Linux build legs and blocking the GitHub Release. The release workflow now builds the Linux companion on ubuntu-24.04 (glibc 2.39) while keeping the orbit CLI on ubuntu-22.04, so the CLI’s glibc-2.35 compatibility floor is preserved. This fixes forward after the v0.8.0 GitHub Release failed to publish while @orbit-tools/cli@0.8.0 had already been published to npm. (Pull request)

0.8.0

Breaking Changes

  • Agent MCP surface slimmed; orbit.review-thread.* alias dropped: six tools moved off the agent-facing MCP surface to register_inactive — orbit.task.reject, orbit.friction.list, orbit.friction.show, orbit.friction.resolve, orbit.learning.comment.list, and orbit.learning.upvote. They stay reachable via the CLI (orbit tool run) and runtime.run_tool, but tools/list over MCP no longer advertises them and orbit mcp serve rejects agent calls to them — these are operator/triage actions, while agents file friction via orbit.friction.add and discover learnings via orbit.search. Separately, the duplicate orbit.review-thread.* alias (added in ORB-00273) is removed; the canonical orbit.task.review_thread.* form that the README, the orbit-review-task skill, and the audit/redaction policy reference is unchanged. Seeded activity allowlists and skill docs were updated to drop the stale references. EXPECTED_INACTIVE_TOOL_NAMES length canary moves 21 → 27. (Pull request, Pull request)

Highlights

  • Dashboard and scoreboard redesign: the dashboard adopts the new visual mock and the scoreboard gets a phase-1 reskin plus a window-aware backend — a ?window= selector (1h / 24h / 7d / 30d / all) now scopes scoreboard metrics, on an additive snapshot schema bump (v5 → v6). (Pull request, Pull request, Pull request)
  • Crew and duel composition adapt to detected agents: [workflow].default_crew and [duel].candidates are now derived from the agent families detected at orbit init rather than hard-coded, so a workspace’s defaults match the agents actually installed. (Pull request)
  • Semantic-companion install hardened: the release workflow now publishes the standalone orbit-search-companion-* binaries alongside the CLI tarballs, so orbit semantic install can resolve signed companion assets from the GitHub Release. (Pull request)

0.7.1

Fixes

  • Rotate release signing key and recover package installs: adds orbit-release-key-3 to the shell and npm installer trust sets, makes it the canonical plugin/npm/release-signing.pub, and documents the new current signing path. This patch fixes forward after the v0.7.0 GitHub Release publish failed before assets were created while @orbit-tools/cli@0.7.0 had already been published. (Pull request)

0.7.0

Breaking Changes

  • Agent MCP surface trimmed: six admin/destructive tools moved off the agent-facing MCP surface — orbit.task.delete, orbit.task.lint, orbit.semantic.uninstall, orbit.adr.list, orbit.learning.prune, orbit.learning.comment.delete. They remain registered (register_inactive) and reachable via the CLI (orbit task delete, orbit adr list, etc.) and runtime.run_tool; tools/list over MCP no longer advertises them and orbit mcp serve rejects calls to them. Agents that need ADR enumeration use orbit.search --kind adr. A new orbit adr list CLI subcommand routes through the tool to preserve filter parity (--include-remote, --status, --owner, etc.). EXPECTED_INACTIVE_TOOL_NAMES length canary in crates/orbit-cli/src/command/mcp/tests/mod.rs bumps 15 → 21. (Pull request)
  • Semantic companion lookup no longer searches $PATH: normal runtime lookup uses the managed ~/.orbit/embed/bin/ install only. Dev/CI overrides must set ORBIT_SEARCH_COMPANION=<absolute path> plus ORBIT_SEARCH_COMPANION_ALLOW_UNSAFE=1; managed downloads now require a release-signed checksum manifest before replacement or execution. (Pull request)
  • Envelope and embedding rebuild verbs renamed: orbit docs reindex / orbit.docs.reindex is now orbit docs index / orbit.docs.index for doc-corpus embeddings, and orbit learning reindex / orbit.learning.reindex is now orbit learning sync / orbit.learning.sync for YAML-to-SQLite envelope reconciliation. No compatibility aliases are retained. (Pull request, Pull request)
  • orbit search mode split and per-kind status syntax: the query surface now has three visible CLI forms: orbit search <query>, orbit search similar <id>, and orbit search path <path>. The old orbit search --semantic <id> and orbit search --path <path> forms are removed. --status now requires kind:value tokens such as task:open,doc:active,adr:proposed; bare tokens like --status open are rejected. CLI --field / --model and MCP field / embedding_model are removed from orbit search; MCP model remains provenance-only. ADR-0179 supersedes ADR-0175. (Pull request)
  • Per-domain search subcommands removed; cross-kind filters on orbit search: orbit task search, orbit docs search, orbit learning search (and the matching orbit.task.search, orbit.docs.search, orbit.learning.search MCP tools) are removed. Replacement: orbit search --kind {task,doc,learning,adr,all} <query> for content similarity and kind-specific list commands for structural filters. orbit search --all, --tag, per-kind --status, and orbit search path <path> cover the previous include-superseded/path/tag use cases. orbit learning list --path now uses glob containment rather than exact string match. (Pull request)
  • orbit search flag rename: free-text vector ranking is now --hybrid / hybrid: true; task-neighbor lookup is now --semantic <id> / semantic: "<id>". The old --semantic boolean and --related <id> surfaces are hard-removed; JSON mode values are now hybrid and neighbor. Historical phase-1 audit payloads carrying semantic: true are orphaned by the no-shim rename. (Pull request)
  • Design-doc tooling retired: orbit design check, orbit design init/list/show, orbit.design.*, the orbit-design skill, the wrapper script, and make check-design-docs are gone. Use orbit-docs, orbit.docs.*, and the same-PR doc update rule; orbit workspace init --design still seeds docs/design/CONVENTIONS.md. (Pull request, Pull request, Pull request)
  • Search namespace split: orbit.semantic.search, orbit.semantic.related, and the orbit-semantic skill are removed in favor of orbit.search and orbit-search; orbit semantic reindex is now orbit semantic index. Historical semantic.search / semantic.related audit event names are orphaned by this hard break because there are no external audit-history consumers yet. (Pull request)
  • Low-use CLI commands retired: orbit scoreboard, orbit metrics, and orbit adr migrate are removed. Their supported replacements are the dashboard scoreboard and metrics views, dashboard HTTP metrics endpoints, and the already-migrated orbit.adr.* artifact tools. (Pull request, Pull request, Pull request)
  • Legacy executor models: compatibility removed: executor YAML must use model_pair_override:; the one-release models: alias and deprecation warning path are gone. (Pull request)

Highlights

  • Unified knowledge search: lexical and hybrid search now spans tasks, docs, learnings, and ADRs, with doc/learning/ADR embeddings, task-neighbor lookup, path applicability, and a single orbit search mental model. (Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request)
  • Worktree-aware artifact foundations: Orbit now distinguishes shared and local .orbit/ roots, adds a global artifact ID allocator, and lays the groundwork for worktree-local ADR/Learning artifacts while keeping existing stores on the shared root. (Pull request, Pull request, Pull request)
  • Docs and learning context in the agent loop: PreToolUse learning reminders, orbit-docs, indexed design docs, and task.show --with-context make project knowledge easier for agents to retrieve before editing. (Pull request, Pull request, Pull request, Pull request)
  • Dashboard and steering improvements: Audit / Diagnostics side panels, scoreboard grouping, a task resolver, metrics panels, review-thread message panels, and hook-driven async steering make the dashboard a fuller operations surface. (Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request)
  • Release and runtime hardening: installers authenticate release archives, semantic-companion binaries require signed checksums, key rotation is documented in the installer trust set, and the companion freshness check closes a TOCTOU window. (Pull request, Pull request, Pull request, Pull request)
  • Safer agent writes and execution: artifact-write redaction, ambient MCP session context, and restricted v2 proc.spawn execution reduce the odds of leaking secrets or running in the wrong workspace. (Pull request, Pull request, Pull request)

0.6.0

Breaking Changes

  • Branching model flipped: main is now the release/production branch; agent-main is the dev integration branch where task PRs land. Each release tags on agent-main then promotes to main via merge commit; hotfixes branch from main and back-merge to agent-main. Install URLs in README.md and the website now point at main. Retired stub crates/orbit-core/assets/activities/examples/promote_agent_main.yaml removed. See RELEASING.md §10b and §Hotfix flow. (Pull request)
  • Crew registry replaces role-keyed agent config: [agent.planner], [agent.implementer], [agent.reviewer] blocks are removed in favor of named [crews.<name>] registries selected via [workflow].default_crew or per-task crew. Workspaces with stale schema are rejected at load. (Pull request)
  • ship-auto and ship-local aliases removed: orbit run ship is now the unified async-by-default command — empty task IDs trigger auto-backlog mode, explicit IDs queue-and-wait instead of fail-fast. (Pull request)
  • Agent identity schema collapsed to family: PlanningRoleAssignment drops model; planning-duel artifact paths and signatures rename to {slot}.md and *authored by: {family} / {slot}*; resolve_agent_model_pair* helpers and alias canonicalization removed; scoreboard by_model → by_family. Includes a read-side migration; downstream consumers indexing on by_model keys break. (Pull request)
  • Family-identity migration script required after upgrade: scripts/migrate_family_identity.py rewrites persisted task/friction/run/audit/scoreboard records and four SQLite stores to family-keyed shapes. Dry-run by default; back up before applying. (Pull request)
  • Audit task.locks.* and job run-pipeline-worker events populate task_id/job_run_id semantically: previously these overloaded target_id. Downstream analytics joining on target_id for these events need updating. (Pull request)
  • FrictionStatus::Default derive swap: enum default impl moved from a hand-rolled impl to #[derive(Default)] to satisfy a clippy-deny lint. Behaviorally a no-op; listed for completeness. (Pull request)
  • Task relation enum gains Produces/Resolves; frictions auto-close on Review → Done: relation enum extension is additive but the auto-friction-close on task completion is a semantics change. A resolved_by_task back-pointer is added to friction records. (Pull request)
  • Learning storage layout: flat → per-entity directories: .orbit/learnings/<L-id>.yaml moves to .orbit/learnings/<L-id>/learning.yaml. Legacy-layout load returns a typed error directing operators at orbit learning migrate-layout. (Pull request)
  • ADR allocation policy: new ADR headings must first allocate globally via orbit.adr.add; local 3-digit ADR headings under 4_decisions.md are grandfathered. docs/design/CONVENTIONS.md §4 updated. (Pull request)
  • Knowledge-graph workspace_root attribution fixed: graph refs are now keyed on the worktree’s actual branch rather than falling back to the main repo’s agent-main; the missing-ref case rebuilds rather than silently reading the default branch. Cached selectors keyed off the old behavior may need refresh. (Pull request, Pull request)
  • orbit.task.update persists source_task_id: the property was previously silently dropped on writes; clients that depended on the drop now see persistence. Empty-string clears the field. (Pull request)
  • MCP schema emitter no longer degrades object_list params to "string": evidence and similar fields now emit array-shaped schemas. MCP clients that worked around the bug by string-encoding payloads must send arrays. (Pull request)
  • Ship batch commit message template: feat: parallel batch [ORB-id] is replaced by a deterministic template — <type>: <truncated title>… [ORB-id] [EXT-id]… with an optional full-title line, execution-summary paragraph, and Planned-By / Implemented-By trailers. Release-note builders and git log --grep 'parallel batch' workflows break. (Pull request)

Highlights

  • Grok onboarded as fourth agent family: sandbox state dir + SBPL allowances; CLI runner + executor YAML; orbit mcp init --client grok writes .grok/config.toml; duel / scoreboard / friction-stats render zero-grok rows; design folder, commit identity, and docs updated. (Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request)
  • Knowledge tab in dashboard: Learnings subtab (list, supersede, stats), Frictions subtab (triage, resolve, stats, tag-picker sourced from YAML), ADRs subtab (accept, supersede, related-task deep-links), task-detail enrichment, Locked Files panel, per-task crew selector, markdown rendering. (Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request)
  • Duel-plan agent pool configurable via [duel] candidates + [duel.models] in config.toml; runtime-host trait methods for candidate/model resolution; preserves fallback for non-duel callers. (Pull request)
  • Cross-artifact task relations: Produces / Resolves variants enable typed task ↔ friction / learning links; auto-resolves frictions on Review → Done with a resolved_by_task back-pointer. (Pull request)
  • Learning enrichments: per-learning comments.jsonl with push-injection rendering; decay-weighted upvotes with task-anchored idempotency; learning-creation wired into agent activity loops with a checkpoint. (Pull request, Pull request, Pull request)
  • Direct-agent runtime model injection: ExecutorDef.model_flag enables data-driven -m / --model flag dispatch per step. (Pull request)
  • Ship command unified to async-by-default: orbit run ship empty-task-IDs → auto mode, explicit IDs → gated; waiting-reason fields surfaced through orbit run history / show. (Pull request, Pull request)
  • Backlog dependency gating: list_backlog_tasks filters by task_dependencies_ready, fixing out-of-order auto-pipeline execution. (Pull request)
  • Skill quality nudges: orbit-create-task now teaches optional complexity, dependencies, parent_id, and cross-artifact relations; orbit.friction.add description enumerates the tag taxonomy from YAML. (Pull request, Pull request, Pull request)

0.5.4

Features

  • Project-learnings push-injection (L1/L2/L3): relevant learning summaries are now injected into agent context at three layers — engine pre-prompt before runtime spawn, MCP sidecar on path-bearing tools (orbit.graph.show, orbit.graph.refs, orbit.task.show), and a Claude Code PreToolUse hook on Edit | Write | Read. Summary-only payloads with per-session dedup, per-call caps, and an ORBIT_SESSION_ID envelope for cross-process dedup. (Pull request)
  • First-class design-docs surface (orbit.design.* + orbit design check CLI): four MCP tools (init, list, show, check) plus a Rust port of the design-doc decay checker, with orbit workspace init --design seeding docs/design/CONVENTIONS.md when absent. make check-design-docs and scripts/check_design_doc_decay.py now wrap the Rust path. (Pull request)
  • Default-seeded skills aligned across asset, registry, plugin, and router catalogs: orbit-learning and orbit-design onboarded; orbit-review-task and the semantic-search skill brought into the default seed and the plugin’s skills/ symlinks; three drift-detection unit tests guard the four catalogs against recurrence. Default seed bumped 7 → 11. (Pull request, Pull request)
  • orbit workspace init --inject-agent-rules: opt-in flag writes an idempotent Orbit-rules block into CLAUDE.md and AGENTS.md at the workspace root, delimited by <!-- orbit-managed:start/end --> markers. Block content sourced from an editable asset; malformed marker pairs refuse to write. (Pull request)
  • Inline task status transitions in the dashboard: per-task actions row gains a status selector wired through the existing PATCH /tasks/:id backend, ordered by STATUS_ORDER with done last and excluding rejected/archived/friction. Surfaces a “no longer shown in dashboard list” notice when transitioning to done. (Pull request)
  • orbit.learning.* exposed over MCP: the full eight-tool learning surface (add, list, search, show, update, supersede, prune, reindex) is now reachable from every MCP client, not just orbit tool run. Restores parity with the orbit-learning skill instructions. (Pull request)

0.5.3

Features

  • Claude Code plugin SessionStart hook: the Orbit plugin now ships a SessionStart hook (plugin/hooks/check-workspace.sh) that detects an uninitialized workspace via a pure filesystem walk and surfaces a systemMessage instructing the user to run orbit init / orbit workspace init. Closes the silent-no-op gap where orbit mcp serve would attach with zero tools and no in-session signal. (Pull request)

Fixes

  • No-diff PR handoffs: pr_open now treats branches with zero commits ahead of the configured base as successful repository-noop handoffs, moves completed tasks to review, returns pr_created: false, and avoids creating an empty GitHub PR. (Pull request)

0.5.1

Fixes

  • Deprecation warning routes through tracing: the knowledge.task_id_pattern deprecation notice was emitted via writeln! directly to stderr from orbit-core, contradicting the CLAUDE.md “tracing for diagnostics; eprintln! only in orbit-cli” rule and leaking into cargo test output. Now goes through tracing::warn! with a structured config field.

0.5.0

Breaking Changes

  • Legacy task artifact store removed: the v1 single-file task store and its migration helpers were deleted in favor of the v2 bundle store. [task] artifact_store accepts "v2" as a no-op for forward compatibility and rejects "legacy" with an explicit migration error. Workspaces still on legacy storage must migrate under v0.4.0 before upgrading. ([commit e9582eba], [commit 222f6020], [commit 123f89f7])
  • Task.workspace_path / Task.repo_root dropped from the update path: TaskAutomationUpdate, TaskRecordUpdateParams, and TaskDocumentUpdateParams no longer accept these fields, and the v2 document layer rejects them at load. Worktree setup and parallel-batch dispatch no longer thread them through. The fields remain on the public DTO via projection from workspace metadata, but write-path inputs that include them now fail. ([commit 6beb14a2])

Features

  • ADR-artifact subsystem (orbit.adr.*): ADRs lift out of per-feature 4_decisions.md markdown into first-class artifacts at .orbit/adrs/<status>/<id>/{adr.yaml,body.md} with globally-unique ADR-NNNN IDs and a three-state lifecycle (proposed/accepted/superseded). Ships domain types, a SQLite envelope index, a file store with per-ADR fs2 locks, five tools (add, show, list, update, supersede) with lifecycle audit rows, an orbit adr migrate one-shot, and a parser/sweeper hardened against rollup-bullet form and code-fenced examples. Migration imported 142 ADRs (97 accepted / 39 proposed / 6 superseded) across the existing corpus. The envelope’s legacy_id: String field was renamed to legacy_ids: Vec<String> mid-development per the ADR-002 amendment; pre-GA churn, no external consumers yet. (Pull request, Pull request, Pull request, Pull request, Pull request, Pull request)
  • Project learnings subsystem (orbit.learning.*): workspace-scoped LearningFileStore under .orbit/learnings/<id>/learning.yaml with a SQLite learnings_index for fast scope-glob lookups on the injection hot path. Eight MCP tools (add, list, search, show, update, supersede, reindex, prune) plus matching orbit learning <verb> CLI; orbit learning migrate-layout upgrades legacy flat workspaces. Learnings travel with the repo via a carved-out .gitignore entry. End-to-end dispatch latency p95 ≈ 0.04 ms at 500 records. (Pull request, Pull request, Pull request)
  • Plugin install contract locked down: make release-check enforces version lockstep across plugin/npm/package.json, plugin/.claude-plugin/plugin.json, npm view @orbit-tools/cli, and gh release list -L 1; scripts/smoke-plugin-install.sh exercises the published @orbit-tools/cli@latest postinstall + MCP tools/list handshake; .github/workflows/smoke-plugin-install.yml runs the smoke weekly and on every v* tag on macOS and Linux; docs/RELEASE.md codifies the chain. v0.5.0 switched the npm publish step to manual after the workflow ran into account-level 2FA. (Pull request, Pull request)
  • Typed error discriminators: KnowledgeError.kind is now the typed KnowledgeErrorKind enum, and ten per-kind OrbitError::*NotFound(String) variants consolidate into NotFound { kind: NotFoundKind, id: String } with exhaustive matches in every error-code translator. JSON wire shape is preserved via #[serde]; internal Rust matchers move to the typed kind. (Pull request)
  • Property + snapshot tests on protocol boundaries: proptest-backed Selector display/parse roundtrip (256 cases per variant), multi-threaded GraphLockGuard/LockStore concurrency with a 5-second deadline, and committed JSON-snapshot coverage for AuditGuard success/failure/denied events. (Pull request)
  • Knowledge graph workflow relocation: build/show/search/history workflows moved from orbit-tools into a new orbit_knowledge::workflows module, so orbit-knowledge owns both the tool surface (commands/) and the host application surface (workflows/). ([commit 5fc2b72c])
  • Forward-only YAML migration framework: orbit_common::migration::Plan chains Value → Value steps keyed on schema_version; OrbitError::Migration separates chain failures from store/parse errors; the v2 task-bundle envelope read path now flows through task_migrations::envelope_plan() (empty chain today — next schema bump adds one add_step call). ([commit 01928e76])
  • Semantic companion install hardening: companion installation is version-aware, --force reinstalls supported, and background-companion stderr is suppressed during task-mutation indexing. (Pull request)

0.4.0

Breaking Changes

  • Friction reports relocated: friction is no longer a task type or status. Records live as append-only markdown under .orbit/frictions/{yyyy}-{mm}/F{nnn}.md and are managed through orbit.friction.add/list/show/stats. orbit.task.add rejects type: friction / status: friction; web API and scoreboard JSON drop friction_bounty. (Pull request)
  • Task type taxonomy reduced: the task | feature | epic | issue | bug | chore | refactor | friction enum collapses to feature | bug | refactor | chore. orbit.task.add and orbit.task.update reject the removed values; existing tasks were migrated. (Pull request)
  • Attribution narrowed to model: the agent field is removed from Actor; orbit.task.add rejects an agent parameter and Orbit infers the agent family from model via agent_from_model. MCP orbit_task_list, orbit_task_search, and orbit_task_review_thread_list responses are now object-shaped (previously top-level arrays) so Cursor and VS Code accept them. (Pull request)
  • Knowledge-graph leaf IDs unified across extractors: Python, Rust, Java, and TypeScript leaf selectors now use a single canonical form so SQL and JSON paths return set-equivalent results. GRAPH_SQLITE_INDEX_SCHEMA_VERSION bumps; consumers caching selectors must rebuild. (Pull request)
  • Semantic search requires a companion binary: orbit-embed-companion is installed separately via orbit semantic install; orbit semantic * and the matching MCP tools fail until it is present. (Pull request, Pull request)
  • JobV2Step rejects multi-body shapes: YAML steps that previously parsed silently with both target and parallel (or any other body combination) now fail at load. (Pull request)
  • orbit-locks skill removed: the seeded orbit-locks/SKILL.md and the ad-hoc orbit.task.locks* instructions in the seeded orbit skill are gone — the gate pipeline still owns reservations. External agent prompts referencing the skill must be updated. (Pull request)

Features

  • Knowledge graph latency wins: lazy source hydration via GraphReadOptions, a bounded default-ranking work cap on search, and a BinaryHeap top-K in overview.top_files. (Pull request, Pull request, Pull request)
  • Knowledge command surface: graph business logic — ranking, classification, fast-path orchestration — relocated into orbit_knowledge::commands::* so non-tool consumers share canonical behavior. (Pull request)
  • Semantic search subsystem: orbit-embed client, orbit-embed-companion binary, embeddings and tasks_fts SQLite schema, paragraph chunker, BLAKE3 dedup, task-mutation index hooks, and orbit semantic install/uninstall/reindex/stats/search/related CLI plus MCP surface. (Pull request, Pull request, Pull request, Pull request)
  • Task tags: first-class tags: Vec<String> field with normalized SQLite index and --tag filtering on orbit task list/search. (Pull request)
  • Activity/job runtime polish: wildcard-aware tool allowlists honored at dispatch and HTTP-loop schema advertisement, asset-load-time allowlist validation, agent-loop on_denial: continue, literal-boolean condition atoms, and exclusive locking on duel scoreboard appends. (Pull request, Pull request, Pull request, Pull request, Pull request)
  • Recovery role configurability: seeded step-failure recovery activity uses role: reviewer and resolves agent/model from [agent.reviewer] config instead of hardcoded Codex. (Pull request)
  • Done-task sync cap: website task sync caps generated pages to the 100 most recent done tasks. (Pull request)
  • Debug-job-failure skill: seeded orbit-debug-job-failure SKILL.md teaching agents how to investigate failed/stuck/cancelled job runs across run state, audit events, blobs, and live processes. (Pull request)

0.3.1

Features

  • Pipeline dispatch reliability: hardened parallel, gate, and epic pipelines with failed-child completion handling, longer task lock coverage, epic timeout/convergence fixes, resolved workspace subprocess cwd, and per-step agent log/error surfacing. (Pull request, Pull request, Pull request, Pull request, Pull request, Pull request)
  • Metrics and public docs: split the public metrics surface into Operations and Scoreboard views, added done-task sync pages for orbit-cli.com, refreshed positioning/reference docs, and refined the website UI. (Pull request, Pull request, Pull request, Pull request, Pull request)
  • Registry and benchmark tooling: added the orbit-registry crate and identity-key benchmark harness for exercising knowledge graph selector stability. (Pull request, Pull request)

Fixes

  • macOS sandbox and CLI execution: allowed Claude’s $HOME/.claude.json lock/tmp siblings, re-allowed the active job-run worktree after global deny rules, and demoted successful CLI exits when the inner Orbit envelope reported failure. (Pull request, Pull request)
  • Workflow defaults and links: made workflow base branches resolve from [workflow] base_branch when CLI flags are omitted, and fixed task-ID links in generated PR bodies with an opt-in URL template. (Pull request, Pull request)
  • CI clippy guardrails: grouped macOS sandbox spawn inputs into a request struct so strict workspace clippy passes under -D warnings. (Pull request)

0.3.0

Release scope

  • Stable surface: CLI agent backends. v1 supports backend: cli as the stable agent invocation path, running Codex, Claude Code, Gemini CLI, and other official CLIs as supervised subprocesses. backend: http (LoopTransport) and the Groundhog checkpoint runner remain preview-only for v1; they are exercised in tests but can change before v2.

Breaking Changes

  • Activity/job schema v1 removed: loaders now reject schemaVersion: 1 activity/job assets, the v1 reconcile/runtime/store paths are gone, and schemaVersion: 2 is the canonical activity/job surface. (Pull request, Pull request)
  • Workflow commands reorganized: stable entrypoints are orbit run ship <TASK_ID>..., orbit run ship --mode local <TASK_ID>..., orbit run ship-auto, orbit run duel-plan <TASK_ID>, and orbit run job <JOB_ID>. The direct orbit run <JOB_ID> shorthand and workflow-specific run ship list/show and run duel list/show commands were removed; use orbit run history and orbit run show for job-run inspection. (Pull request, Pull request, Pull request, Pull request)
  • Task attribution history moved to orbit graph history: selector history is graph-owned, so the query now lives next to orbit graph search/show, and rebuilds use orbit graph build. Both orbit graph build and orbit graph history accept --task-id-pattern <regex>; workspace config knowledge.task_id_pattern is the steady-state setting, with CLI flag > config > Orbit default precedence. The selected pattern is recorded in manifest.json, and mismatches emit a stderr warning. orbit.graph.history exposes the same surface to MCP clients. (Pull request)

Features

  • Activity/job v2 runtime: added schema v2 activities and jobs with typed DAG blocks (parallel, fan_out, loop, retry, when), activity name resolution, backend: auto normalization, backend: cli dispatch, HTTP agent loops, session-bound loop steps, and a v2 audit envelope with workspace provenance. (Pull request, Pull request, Pull request, Pull request, Pull request, Pull request)
  • Seeded task pipelines: added load-bearing seeded workflows for PR, local, gate, auto-dispatch, and epic shipment, including task reservations, backlog bundling, admission-controlled dispatch, and session-backed epic orchestration. ([T20260419-0622-3], Pull request, [T20260419-0623-2], Pull request)
  • Knowledge graph: added the Rust orbit-knowledge graph, orbit graph build/update/search/show, graph MCP tools, compact overviews, callers/implementors/dependency navigation, edit buffering, shared locks, auto-refresh, branch-scoped refs, task-ID attribution metadata, and markdown/config/table extraction. (Pull request, [T20260412-0645-2], [T20260412-0645-3], Pull request, Pull request, Pull request)
  • MCP integrations: added the orbit-mcp crate, orbit mcp serve, safe default graph/task tool exposure, external MCP/plugin tooling, and orbit mcp init/remove setup for Claude, Codex, and Gemini clients. (Pull request, Pull request, Pull request, Pull request)
  • Dashboard and observability: added orbit web serve; task, job, audit, scoreboard, and dashboard APIs; diagnostics and recent-runs views; task actions; copyable task IDs; connection health; skeleton/loading states; markdown rendering; and live-data animations. (Pull request, Pull request, Pull request, Pull request, Pull request, Pull request, Pull request)
  • Task planning and search: added structured task plans, dependency support, epic task type support, selector-first task context, agent task search, and richer task field projection for agent/tool callers. (Pull request, [T20260420-0509-2], Pull request, Pull request, Pull request)
  • Groundhog execution model (preview): added Groundhog chronicle serialization, workspace snapshots, verb tools, checkpoint verification, and a dedicated Groundhog v1 activity runner. (Pull request, [T20260420-0509-3], [T20260420-0509-4], Pull request, [T20260420-0510-2])
  • Provider and evaluation support: added Gemini support, configurable agent/model selection, provider invocation traces, HTTP LoopTransport implementations for Anthropic/OpenAI-compatible/Gemini providers, planning duels, scoreboard attribution improvements, and versioned knowledge-graph benchmark harnesses. ([T20260411-1937-2], [T20260412-0457-2], Pull request, Pull request, Pull request, Pull request, Pull request)

0.2.0

Features

  • Parallel batch execution: dispatch and execute multiple tasks in parallel with file-level conflict detection and shared worktrees
  • Auto-cleanup on merge: ship workflow now deletes the remote branch after a successful PR merge

Fixes

  • --parallelism flag: serialized as JSON integer instead of string, fixing schema validation failure on orbit run ship --parallelism N
  • Stale default artifacts: orbit workspace init now always refreshes default skills, activities, and jobs to their latest embedded versions (custom artifacts are preserved)
  • Clippy warning: resolved unused-mut warning and removed clippy from CI

0.1.0

Initial release of Orbit.

Core

  • Task lifecycle: propose, approve, implement, review, and archive tasks with full history tracking
  • Activity system: reusable operations with defined input/output schemas and three spec types (agent_invoke, cli_command, automation)
  • Job engine: composable multi-step pipelines with conditional execution, retry logic, nested jobs, and parallel dispatch
  • Workflow aliases: orbit run ship, orbit run ship-local, orbit run review as ergonomic entry points over raw job invocation
  • Multi-agent orchestration: parallel task workers with file-level locking in shared worktrees
  • Multi-model strategy: configurable agent/model per job step (e.g., Opus for planning, Codex for implementation)

CLI

  • Grouped command surface: run workflows, manage work, configure and inspect
  • JSON and table output modes across all commands
  • Audit event logging for every CLI invocation

Infrastructure

  • Layered Rust crate architecture (types, policy, exec, tools, store, agent, engine, core, cli)
  • Two-root workspace model: global (~/.orbit/) and workspace-local (.orbit/)
  • File-based (YAML) and SQLite persistence
  • RBAC policy evaluation engine
  • Process sandboxing and timeout handling
  • Skill system for agent prompt composition