Release history

What changed, and when.

Generated from the repository’s own CHANGELOG.md. Each released entry links to its GitHub release, which carries the exact artifacts, checksums and attestations.

v0.6.8

2026-09-10

Fixed

**The documented loop refused the word the documents teach.** `tasq setup` and `tasq onboard` take `--space`, and so does every guide, contract and AGENTS.md block. The next line those same instructions tell a new user to run answered `Unknown flag: --space`: `claim`, `attempt`, `evidence`, `done`, `whoami`, `add` and `list` read `--tenant` only. The two are one flag now, whichever a command reads, and passing both with different values is refused rather than resolved to a guess about whose ledger to write to. `onboard` still teaches `--space` alone.
**A flag value that begins with a dash was silently dropped.** Any token starting with `-` was taken for the next flag, so `--summary "--space is refused everywhere"` set `--summary` to true and then reported the sentence itself as an unknown flag. Filing evidence about a flag is exactly where that bites. A flag is one word; a sentence is not, and is now read as the value.
**A missing required flag answered with a raw schema dump.** A first `tasq agent install claude-code` printed a zod `invalid_type` array that named neither the flag that was missing nor the command that wanted it. It now names the flag and the usage line, on `agent install`, `agent instructions`, `mcp` and `web`.
**The npm publish job was cancelled mid-release.** Its 20 minute budget had to cover install, build, attestation and then a serial publish loop that waits for the registry to serve each package as `latest` before starting the next one. npm took over two minutes on one package during v0.6.7, the job hit its timeout with a package still unpublished, and the GitHub release was skipped. The budget now covers the whole loop.

Artifacts and attestations

v0.6.7

2026-09-10

Fixed

**`tasq backup list` wrote a 1.4 MB database into a file named `list`.** Every other noun in this CLI takes `list`, so that is what a new user types; `backup` takes a bare path, so it took the word as the path, wrote the file into whatever directory they were standing in, and reported success. A backup nobody asked for, in a repository somebody may commit. A target that is really a mistyped subcommand is now refused, and `--target list` or `./list` still means the file.
**An unknown command named nothing to try.** `unknown command: tasks` and `run tasq help for usage` is not something a first-time user can act on, and `tasks`, `claime` and `evidenc` are exactly what someone types before reading the help. The nearest command is now suggested when there is an obvious one, and nothing is guessed when there is not.
**Every MCP host logged `0.1.0` whichever release it was talking to.** The handshake is the only place a host learns which Tasq is on the other end, and the version it reported was the workspace package version, hardcoded. So the one field that would say what actually runs where said the same thing everywhere, forever. It is now the version the same binary reports for `--version`, proven against the built artifact.
**`tasq mcp` refused `--space`.** Every other command that names a space takes `--space`; this one took `--tenant` only, which its own help calls the rare override. Both work now. The space is still required and never inferred from the working directory, which the agent contract forbids on purpose: a server that guesses its space guesses whose ledger an agent writes to.
**`tasq agent install claude-code` was refused, by the id the machine contract itself uses.** `AGENT_INTEGRATIONS.json` names that host `claude-code`, so that is what an agent has in hand when it reaches for the deterministic fallback, and the CLI took `claude` only. The refusal now also names what is accepted.
**One refusal could make `tasq usage` unreadable.** A command called with no arguments answers with its whole usage banner, and the journal stored it whole, so the report reprinted a dozen lines of argument syntax inside an indented column and the counts were lost in it. Only the first line is recorded now, and the reader repairs anything older, because records already on disk would otherwise break the report forever.
**A release stopped dead when PyPI certified a wheel it had just published.** PyPI serves two indexes that do not update together: the JSON API carries a new release before the simple index `pip` resolves against does. v0.6.6 read the wheel digest from the JSON API, dispatched the certification two minutes later, and the certifier's `pip download` failed with `from versions: 0.4.0, 0.6.3, 0.6.4, 0.6.5` - the wheel was published, and the index it installs from had not caught up. The release now waits until the simple index lists the exact file the certifier will ask for.
**A failed certification forced the rest of the release to be done by hand.** No registry accepts a version twice, so a rerun of `release:publish-surfaces` failed on the publication step it had already completed, and the only way forward was dispatching the remaining workflows one by one - which is how v0.6.2 shipped without two of its surfaces. Publication is now decided by the registry rather than by rerunning it, so a resumed release re-certifies and continues.

Artifacts and attestations

v0.6.6

2026-09-10

Added

**A private local record of what every `tasq` command was asked to do.** The ledger records mutations that SUCCEED. A read left no trace and a refusal left none at all, so the two signals that say whether the product fits the hand using it were invisible: both defects found on 2026-09-09 existed in the ledger as nothing whatsoever. `~/.tasq/commands.jsonl` now takes one line per invocation, successes included, and `tasq usage [--since 30d] [--all]` reports refusals, reads, harnesses and versions, and every space on this machine at once. The file is `0600`, bounded, and leaves the machine only when someone runs `tasq feedback push`. It records the SHAPE of a command - verb, allowlisted subcommand, flag NAMES - and never a positional, a flag value or an actor label.
**Agents are told where to send what the tool got wrong.** The managed `AGENTS.md` block now names `tasq feedback`, and says how it differs from a capture: a capture belongs to the project's work, feedback belongs to the tool. Blocks written before this are version 1.
**`tasq doctor` says when this project's managed block is older than the executable.** Upgrading the binary does not touch a block a project already carries, so an adopted project kept teaching its agents the previous release's rules, and nothing announced the drift. Reported and never fatal: failing `doctor` in every project that has not re-run `setup` would turn a routine upgrade into an outage.

Fixed

**`tasq capture` refused every commitment carrying a planning scope.** 97 of 100 live commitments could not take a capture, because a discovery copied its source's area, goal and project into the kernel, which refuses planning vocabulary without an injected planning-profile policy. Inherited scope is now dropped rather than refused: the kernel still owns no planning vocabulary, and the observation is still filed.
**`release:publish-surfaces` dispatched on the tag by default, and no certifier could accept what that built.** Every provenance verification in the server, Python and Fly workflows expects `refs/heads/main`, and a workflow's provenance names the ref it was dispatched on. v0.6.4 passed only because it was run with `--workflow-ref main` by hand; v0.6.5 failed on the default. The default is now `main`; the bytes stay bound to the tag by `source_commit`.
**A publish dispatched on the release tag deadlocked the release.** Every certifier demands `refs/heads/main`, and provenance names the ref a run was dispatched on, so a tag-dispatched build produced an image no certifier could ever accept, which `ensure-oci-tag.sh` then pinned to the release tag where no rerun could replace it. v0.6.5 hit exactly this and only a registry deletion broke it. Both publish workflows now refuse that dispatch before anything is pushed, and the reuse refusal names the deletion that is the only exit rather than reading as a transient registry error.
**Every Dependabot pull request failed on its first step.** With no configuration, Dependabot rewrote `apps/site/package.json` alone while the pnpm lockfile sits at the workspace root, so all seven jobs died on `ERR_PNPM_OUTDATED_LOCKFILE`. The npm ecosystem is now pointed at the workspace root.
**The manifests agent hosts read to acquire Tasq pinned v0.4.0.** Nothing advanced `AGENT_INTEGRATIONS.json`, its Markdown companion or the mirrored `integration.json` at release time, so a host following the documented "Executable acquisition" path installed a CLI two releases old and never learned anything newer existed. Recording a release now advances them, and the publication gate refuses a repository where they name anything but the published version.
**`tasq usage` could not see a read that lives under a verb.** `attempt list` and `evidence list` are reads, a read leaves no ledger event, and the journal was the only place they could be counted - but the counter matched on the top-level verb alone, so both were invisible. Reads are now keyed by the shape actually invoked.
**Two agents rotating the command journal at once could lose records.** Rotation read the whole file and renamed a rewritten copy over it, so the second rename dropped whatever the first had appended. Rotation now takes an exclusive lock and skips rather than races; the record is appended either way. The lock names its holder's pid, so a writer killed mid-rotation is reclaimed rather than wedging it forever - an expiry would need a clock, and only `systemClock` may read the host clock.
**The CLI JSON contract described a `tasq demo` output that no longer existed.** It documented `tasq.isolated-demo.v1` with a `before` key while `demo` had moved to v2 with `claimed`, `refusals` and `evidence`, and `CommitmentInspectionV1` omitted `signedStatementProofs`. Nothing read that reference, which is why it drifted; `docs:check` now does.
**The README stated a recipe count `tasq onboard` does not return** (45, against 47) and its `setup` transcript omitted the global-default line the command prints. Both are now checked against the running CLI.
**Five reads printed nothing at all on an empty space.** `attempt list`, `evidence list`, `wait list`, `observation list` and `signature bindings` exited 0 with no output, which is indistinguishable from a command that silently failed - and an empty space is exactly what a first-time user has. They now say what is missing and how to create the first one, while `--json` still returns `[]`.
**`tasq usage --all` could not tell a real project from test residue.** It listed every space in the store, and on a working machine 16 of 18 were left by tests, so the first cross-project read was mostly noise. Each space now reports `boundDirectories`, and the human report says how many spaces no directory is bound to.
**House punctuation is the plain hyphen everywhere.** 851 em-dashes and en-dashes had accumulated across 219 files; `pnpm docs:check` now refuses them. Applied migrations are exempt and must be: their bytes are checksummed, so editing a comment in one makes every existing store refuse to open.

Artifacts and attestations

v0.6.5

2026-09-08

Fixed

**The managed AGENTS.md block failed at its fifth step as written.** It prescribed `attempt succeed <task-id>` and the command wanted an attempt id, so every agent following the block hit `attempt not found`. No test executed the block; they read its prose. A task id now resolves to the task's single open attempt, ambiguity and absence are both named, and a test runs the block's own lines in order and fails without the fix.
**`tasq setup` taught Codex and left Claude Code blind.** Claude Code reads `CLAUDE.md` and not `AGENTS.md`, so the block setup wrote was never loaded by a Claude Code session. Setup now writes a `CLAUDE.md` that imports the block, the recipe the Claude Code documentation gives, appends the `@AGENTS.md` line once to an existing file, and leaves a symlink alone.

Artifacts and attestations

v0.6.4

2026-09-06

Fixed

**The server image could not be published for a fresh tag.** `publish-server` runs the full handoff on the tagged commit, and the publication record there necessarily still names the previous release, so the record check refused it. The window between a tag and its record is what the policy calls a release in state `authorized`: the check now reports that tag as `inFlight` instead of a lag, and refuses again once the release is recorded and a newer tag appears. A surface published without its certification is recorded as `published`, never as certified.

Artifacts and attestations

v0.6.3

2026-09-06

Added

**A context link is found from the thing it points at.** `tasq context-link list --system … --resource-type … --external-id …` lists every commitment an external thing is linked to across the space, and `--purpose` narrows a listing to one purpose. This is how "has this message already become a commitment" and "who reported it" are answered. The chat-provenance vocabulary is documented in `docs/guides/CHAT_PROVENANCE.md`.
**`tasq usage` counts what actors actually do.** Per actor and per event type over a window, mapped onto the ritual the managed `AGENTS.md` block prescribes, with the commands nobody ran named. The block prescribed ten commands; this project's own ledger used five. Reads that leave no event are reported as unobservable, never as zero. Contract `tasq.usage-report.v1`.
**The two hand-done halves of a release are commands.** `release:prepare` advances the release authorization, the TQ-616 program and every candidate surface to the version about to be tagged and opens the changelog entry; `release:publish-surfaces` publishes and certifies the server image and the Python wheel and deploys the Fly private beta; `release:record` advances every public surface from what is actually public and verifies the record. Every surface ships by default: v0.6.2 went out without its server image and Python wheel because nobody had authorized them before the tag.

Fixed

**A release stopped after its first npm package.** npm now acknowledges a publish before the version is readable ("Your package is being processed and may take a few minutes to become available"). The verification asked for the version in the same second, got 404 and failed the job, with six packages unpublished and the GitHub release skipped. It now waits for the registry, with backoff up to a deadline (`--wait-seconds`, ten minutes by default); the pre-check that expects "missing" still answers at once.

Artifacts and attestations

v0.6.2

2026-09-06

Added

**`tasq doctor` reads the configuration, not only the store.** A new `config` section, also available alone as `tasq doctor --config` without opening any store, reports `binding_drift` (the `AGENTS.md` block names a space commands here would not use), `dangling_binding` and `temporary_binding`, `default_space_unbound` and `projection_outside_bound_tree`, each with the command that repairs it. `--prune-bindings` removes bindings to directories that no longer exist and journals the removal. `pnpm agent:preflight` runs the check and stops on drift. Contract `tasq.config-doctor.v1`.

Fixed

**`tasq-dev` refused every git worktree.** The shim written by `pnpm dev:link` tested for a `.git` directory, and a worktree's `.git` is a file, so a dev build linked from a worktree was reported as a checkout that is gone. The shim now tests that the checkout exists, whichever shape git gave it.
**A fresh machine wrote into one person's ledger.** The CLI's built-in defaults named a personal space and actor, so any command run before `tasq setup` on any machine landed in that person's space. A machine with no config file now has no space: an unbound directory refuses with the two commands that fix it, `tasq init` creates a neutral `local/default` space and writes it to the file, and the default actor is the account name. The same personal name was the implicit workspace in more than a hundred kernel and service fallbacks; it is now one named constant, `LEGACY_DEFAULT_WORKSPACE_ID`, documented as a defect to remove by requiring the workspace everywhere. Column defaults in the store are unchanged.
**One projection file rendered whichever space was effective.** The global `projectionTarget` was written after every mutation with the effective space, so a life instance's `TASKS.md` received a software project's backlog. A directory binding now renders only its own projection, registered with `tasq use <space> --project-to <file>` and required to live inside the bound directory; the global target renders only the global default space. `tasq doctor --config` reports `global_projection_ignored_here` when the global target is set but the directory in effect renders nothing.
**`tasq setup` rewrote the global default on every run.** Setting up one project made its space the fallback for every unbound directory on the machine, so an agent setting up a scratch project redirected every other checkout. The global default now moves only with `--default`, or when no config existed yet; `setup` in an already bound directory is set up again for its own space (`spaceSource: inherited-from-directory`). The result contract is `tasq.human-setup.v3` and reports `globalDefault {space, changed, source}`.
**A config write could drop bindings another session had made.** Every writer saved its own copy of `config.json` whole. Writes now merge directory bindings with the file on disk, remove a binding only when the command names the directory it unbinds, and append a `tasq.config-change.v1` record to `~/.tasq/config-journal.jsonl` saying which command changed what.
**A repository's managed block could name a space no command used.** The block in `AGENTS.md` is digest-bound, so its text could not drift; its binding could, silently. `tasq use --json` now reports `managedBlock` and `drift`, and `tasq use --from-instructions` binds the project as its repository declares.

Artifacts and attestations

v0.6.1

2026-09-02

Fixed

**The refusal suggestion could not be pasted.** Refusing a completion printed the resolved `.local/lib/tasq/0.6.0/darwin-arm64/index.js` path. It is meant to be copy-pasted, and it pasted the versioned internal layout instead of the `tasq` on PATH, which stops working the moment the managed symlink moves. The onboard recipes deliberately keep the absolute path for the opposite reason: an agent executes the returned vector verbatim.
**The Console had a live feed and asked a human to press a button.** Genuine SSE, an `EventSource` held open, cursor pagination and recovery, and then a badge reading "Changes available" beside a Refresh button. Somebody watching their agents work saw a static page. It re-fetches and swaps the record list now, keeping the filter that was typed, and still degrades to the badge for a history gap where a reload genuinely is required.
**The installer refused after creating things, not before.** Installing over a pre-existing unmanaged `bin/tasq` extracted the archive and wrote the install record first, leaving two directories behind; the next run then failed with a raw `EEXIST`. The check happens before the network is touched, `--dry-run` reports `BLOCKED` and exits non-zero instead of printing a successful plan for the command that would refuse, and an error reaches a human as prose before the contract document.
**The migration receipt did not record who performed the migration.** It captured the source path identity, the format, the cursor and the snapshot digest, and not the writer. When a source-only dev build migrated a live ledger to a format no published binary writes, that was invisible afterwards and had to be inferred. `writtenBy` now carries the version, whether it was a source build, the executable and the pid.

Added

**`pnpm dev:link`** puts a working-tree build on PATH as `tasq-dev`, beside the published `tasq`, so your own build never displaces the one that answers "does this work for somebody who installed it". It refuses to replace a `tasq-dev` it did not write.

Documentation

**A getting started that matches the product.** The README taught `tasq onboard`, which is the agent bootstrap, and never mentioned the command that puts Tasq in a project. It also named `v0.4.2` in five places, including the `npx` line that is the first command anyone runs.
**A skill that can set a project up.** An agent handed `SKILL.md` could join an existing space and could not put Tasq in a project at all: the skill correctly forbade inferring the space and offered no path from "nothing here yet" to "working". Proposing an id and having a human confirm it once is not inferring, and the difference is load-bearing.
**[ADR-025](docs/decisions/ADR-025_WHAT_WOULD_END_THIS.md)** states what evidence would end this project, written before that evidence exists.

Release machinery

The publication record is measured against the newest release tag rather than against itself, and a tag that published nothing must be recorded as retired with a reason. It failed the moment v0.6.0 published, which is the gate working.
Both version-pinned `state` fields are guarded. After v0.5.1 shipped partially certified on one of them, the guard was applied to that block only; the other kept the state of the release it had already been consumed by.
Two vendors on one ledger is an eval that runs in CI, so the claim stays true rather than being made once.

Artifacts and attestations

v0.6.0

2026-08-28

Added

**`tasq contention`.** What the ledger refused. Thirty-one event types and a hundred and thirty-four claims acquired, and until now nothing recorded a single refusal: the ledger kept a complete account of everything it ALLOWED and no trace of anything it PREVENTED. The refusal is the product - `tasq demo` exists to show three of them - and nobody could answer "how many collisions did this stop for me last week". All four refusals a second worker can hit are counted, and a contention is a SITUATION rather than an instant: a polling agent turned away four hundred times is one record with a count of four hundred. A refusal is never a mutation - no event, no revision, no claim change - because everything downstream of the event journal describes work that happened, and a refusal is work that did not.
**`tasq fleet`.** Who is holding what, right now. The expiring lease is what makes this possible without owning any process: a holder that dies stops appearing when its claim lapses. Live claims are grouped by client identity AND working directory, because one label on two machines is not one worker.
**`tasq whoami`.** The actor, the principal it resolves to, this installation's device key, and any OTHER device that has written under the same actor label here. It authenticates nobody and says so, in the prose and in the JSON contract: an identity printed without what it proves reads as authentication.
**Device identity.** Each installation gets an Ed25519 key it did not choose, written `0600` on first `setup`. Every domain mutation records the device behind it. The principal is derived from (space, alias), so two machines using one label were ONE principal and the ledger merged them without a word; on a shared store that silence is the whole problem, and it cannot be reconstructed after the fact.
**Decomposition in the kernel.** `tasq add --parent`, `tasq tree`, `parentCommitmentId` on MCP create, and `tasq_commitment_tree`. Decomposition answers what a commitment is MADE OF; it is a column, not a relation, and it gates nothing. See [ADR-023](docs/decisions/ADR-023_DECOMPOSITION_IS_A_COLUMN.md).
**Relations over MCP.** `tasq_relation_add`, `tasq_relation_end` and `tasq_relation_list` serve the capability the server already advertised and never implemented.
**Four argv recipes** for decomposition and relations, and a `decompose-and-sequence` journey, so an agent on the CLI can discover what an agent on MCP can.
**What a completion just opened.** Finishing a commitment reports the work it unblocked, so the next agent does not have to poll for it.

Changed

**`tasq setup` is one command.** Bringing Tasq into a project took three - `setup`, `use`, `agent instructions --write` - and `setup` mentioned neither of the other two. It now joins the space, binds this directory and its descendants, writes the managed AGENTS.md block, and says which of the three it did. `--no-bind` and `--no-instructions` skip either half. It refuses to set a project up in the home directory or at the filesystem root.
**The first run teaches the agent path.** It printed `add`, `list`, `done` - a single-player todo app, under a headline promising a tracker you share with your agents. It now points at `tasq fleet` and `tasq demo`.
`RELATION_TYPES` drops `parent_of` and `supersedes`. Nothing ever wrote either, and two ways to say one thing is how a ledger starts disagreeing with itself.

Fixed

**A space bound to another project is refused**, naming the directory that owns it. Inheriting a space silently is how work lands in somebody else's ledger.
**Claiming a commitment with unresolved blockers is refused**, naming them, with `--force` for a deliberate override.
`tasq_discovery_capture` on the MCP surface, under the `propose` capability. An agent working over MCP could not report a bug, a missing capability or an inconsistency at all. Capturing never widens, renews or releases the caller's claim, so it is safe mid-task. See [ADR-020](docs/decisions/ADR-020_KERNEL_DISCOVERY_CAPTURE.md).
`captureCommitmentDiscovery` exported from `@tasq-run/core`.

Release machinery

**The publication record** answered `ok: true` while it named a release one version behind reality and the documented installer 404'd. It is now measured against the newest release tag, which the commit that gets it wrong cannot edit. A tag that published nothing must be recorded in `policy.retiredReleases` with a reason - v0.5.0 is the first entry.
**That gate had never run against this repository.** Every case built its own tree, so it was fully covered and never looked at the thing it protects.
**`pnpm verify:release-rehearsal`** builds a real installable release from the working commit and runs every replay the certification job runs, before a tag exists. The last two releases were each broken by something only observable after tagging, and tag protection means a version number does not come back.
**The certification replay** proves the refusal first, then upgrades through `tasq store upgrade`. v0.5.1's own consent gate correctly refused to migrate unasked, and the test was describing behaviour that release removed.
**Every package's tests typecheck**, and the debt list is empty. A test could call an API that does not exist and pass; 66 errors were hiding there, including six input types published as parsed types, a type no value could satisfy, a fixture missing five fields, and four more silently dropped keys.

Artifacts and attestations

v0.5.1

2026-08-27

Added

**Shared assumptions.** One immutable sentence that work rests on, shared by every commitment that depends on it and matched by its normalised text, so two agents phrasing the same belief differently land on one record. Withdrawing it pauses the open commitments resting on it. Three limits are load-bearing and tested: **one hop** - the effect never traverses `depends_on` or `parent_of`; **never terminal** - paused, never cancelled, and `tasq resume` recovers; **never required** - a commitment with no assumption behaves exactly as before. `tasq add --because`, `wrong`, `why`, `resume`, `because list|attach`, and four MCP tools. See [ADR-021](docs/decisions/ADR-021_SHARED_ASSUMPTIONS.md).
**A store safety envelope.** `tasq store status`, `upgrade`, `recovery-points`, `restore` and `clone`. The rollback rule `restore-matching-verified-pre-migration-snapshot-and-binary` was named in three places in the release policy and had no command behind it; now it has one, refusing a snapshot whose bytes no longer hash to its receipt and refusing to discard work written after a recovery point unless forced. `tasq store clone` uses `VACUUM INTO` and rewrites every path inside the clone, because copying `db.sqlite` by hand is wrong twice: wrong paths, and silently empty when the content is still in the WAL.

Changed

Crossing a store format is a decision on **every** build, not only on unreleased ones. Tasq is a shared ledger, so two machines on one store with different versions means whoever runs first silently locks the others out. `tasq store upgrade` is the consent: typing the verb is the decision, so there is no prompt to script around and no flag to set blindly.
A diagnosis no longer mutates what it diagnoses. `tasq doctor` inspects the store format first and stops with an actionable report instead of applying an irreversible upgrade.
Claiming a commitment with unresolved blockers is refused, naming them, with `--force` for a deliberate override. Blocking previously lived in the prioritizer alone, so a blocked commitment could be claimed, started and completed by asking for it directly.

Fixed

`TASQ_HOME` can no longer be overridden by an absolute `dbPath` in the config it loads. Copying a Tasq home to rehearse a migration on it drove the original instead, which cost this project's own ledger a store-format migration under a binary the operator never intended to run there.

Artifacts and attestations

v0.4.2

2026-08-26

Added

A release preflight refuses a tag while any version-pinned policy block still names an older release. Publishing v0.4.0 left several blocks naming v0.4.0, and each only failed when the next release reached it; the one that escaped review could not be fixed afterwards, because the certification workflow reads the policy from the immutable tagged commit. That is why v0.4.1 shipped byte-verified but only partially certified.
ADR-020 proposes discovery capture as a kernel operation reachable from every agent surface. The relation table is already kernel storage and `discovered_from` is already a first-party relation type, but no kernel API writes a relation, so MCP clients cannot report a defect at all.

Changed

No product behaviour changes from v0.4.1. This release exists so the published line ends on a version whose certification is complete rather than partial. Store format 32 is unchanged.

Artifacts and attestations

v0.4.1

2026-08-25

Changed

**Behaviour change.** A terminal transition (`done`, `cancel`) by an actor other than the one holding an active claim is now refused, naming the holder and the expiry. Previously the exclusivity guard existed on `claim` and `attempt` but not on completion, so a third actor could close a claimed task and silently force-release the holder's claim at the decisive moment. Expiry still ends ownership, and `--force` records a deliberate takeover, exactly like `release --force`. Unclaimed tasks are unaffected.
**Behaviour change.** An explicit task `--priority` now replaces the importance inherited from its area or goal, in both directions. Importance previously acted as a floor, so `priority` could only raise it: every task in an important area scored identically and `tasq next` silently degraded to creation order. Tasks with no explicit priority are unaffected and still inherit importance. Existing ledgers will see `tasq next` reorder where a deliberately low priority was being discarded.
`tasq doctor` no longer reports healthy ledgers as missing completion records. The check required a receipt whose revision equalled the task's CURRENT revision, so any edit after closing orphaned it: on a real ledger all twelve findings were false positives. It now requires one receipt per completion, which still catches a reopen and re-close that produced none.
Under `--json`, every non-zero exit now writes a `tasq.command-problem.v1` envelope to stdout alongside the existing stderr message, so an agent-driven caller can act on a refusal instead of receiving an empty machine channel. See [`docs/reference/CLI_JSON_CONTRACT.md`](docs/reference/CLI_JSON_CONTRACT.md).
Per-command help (`tasq help <cmd>` and `tasq <cmd> --help`) now lists the flags accepted by every command, including `--actor`. Argument-error output is unchanged.

Added

`tasq mcp --completion assertion|evidence` sets the completion policy for commitments that MCP server creates when the caller states none, and `tasq agent install` now registers `--completion evidence`. Work an agent proposes through the documented integration therefore states what done looks like and closes only against an inspectable receipt. An explicit `completionPolicy` on the call always wins, and creating an evidence-backed commitment without `successCriteria` is refused rather than silently downgraded.
`tasq list --priority 1-5` and `tasq next --priority 1-5` filter by explicit priority. Filtering happens in the query rather than after the limit, so a match beyond the limit is never hidden.
`tasq evidence add` prints a note when the filing actor differs from the actor holding the active claim, naming the `--actor` value that would attribute the receipt to the holder.

Artifacts and attestations

v0.4.0

2026-08-11

Changed

Use the controlled product-aligned `@tasq-run/*` namespace for every public package, workspace import, release artifact and SBOM identity; explicitly prohibit the unrelated `tasq` package and `@tasq/*` scope.
Make `https://tasq.run` the canonical public website, documentation and pre-executable agent-acquisition entrypoint while retaining `https://github.com/gwendall/tasq` as the source authority.

Fixed

Keep the autonomous onboarding `audit.list` recipe unfiltered across the workspace. `event list --actor` is an event-producer filter, so the recipe now omits it and preserves lossless multi-actor cursor resume.

Added

Add private directory-scoped space selection with `tasq use`, preserving explicit flag/environment precedence and the operator's global defaults.
Add digest-bound `tasq agent instructions` generation with idempotent atomic writes, hand-edit protection and distinct CI exits for missing, stale and modified blocks.
Add offline-first `tasq feedback` capture with secret-free failure context, bounded private storage and explicit idempotence-marked GitHub batch push.
Add a revision-guarded, atomic TQ-607 dogfood tracker for baselines, active use, consumer journeys, resilience drills, friction, critical failures and the final `go`, `extend` or `no_go` decision.
Add machine-readable coding-agent preflight, one-command handoff verification, a documentation map, an executable onboarding eval and a safety-focused pull request template.
Add a concise standalone `SKILL.md` that delegates exact agent workflows to the versioned `tasq onboard` guide and remove installed-help references to repository files that may not exist beside the executable.
Make `packages/tasq-core` the single neutral source authority, replace the Local compatibility mirror with forwarding modules and build the public `@tasq-run/core` candidate from its real source directory.
Add TQ-607, a machine-tracked private dogfood gate requiring at least 30 days across the personal life-pilot, Kami Robotics and an interactive agent runtime before an explicit public-launch decision.
Add standalone human/agent development onboarding, audience routing and a local README for every workspace, plus an executable documentation contract that rejects broken links, stale checkout commands, missing ownership docs and public/private package-metadata drift.
Add canonical human and machine-readable public backlogs so a fresh human or agent can distinguish the next executable checkpoint, external publication gates and unimplemented remote products without private-repository context.
Add a deterministic target release installer with side-by-side versions, atomic activation and data-preserving uninstall, plus a clean-room lifecycle certificate covering onboarding, contention, Console, backup, upgrade and matching snapshot/binary restore.
Add transport-neutral bounded Console overview, work, actor, claim, resource, wait, effect, redacted audit and honest operational-health read contracts with workspace-bound keyset cursors and injected authority time.
Add loopback polling and SSE over one redacted Console event-batch contract, with exclusive reconnect, typed cursor recovery, one-frame backpressure, deterministic overflow fallback and injected time/scheduling.
Add the responsive, keyboard-accessible Local operator Console with seven canonical views, bounded filters, explicit live/stale states, audit timeline and preview-before-download redacted support bundles.
Add installed Local Console lifecycle contracts: a versioned foreground listener announcement, proof-of-life `web status`, private crash-safe registration and full standalone/npm candidate upgrade coverage without checkout-relative assets or install-created listeners.
Add a statically exportable Next.js public product and documentation app with consumer-specific guides, machine-derived support status, exact `/product-truth.json`, synthetic-only visuals and adversarial browser gates.
Add the fail-closed `/adopt.json` pre-executable contract and a candidate human-plus-agent adoption certificate covering installed-byte onboarding, typed contention/recovery and same-ledger Console inspection.
Add cross-platform Local Console browser certification for empty, mature, hostile, corrupt and 2,501-commitment ledgers using fixed injected time, bounded pages, safe errors and real Chromium on Linux and macOS.
Add the pure hosted-authority foundation: strict verified identity, binding, grant, delegation and decision contracts; 16 digest-bound actions; and a deny-by-default injected-clock evaluator with clean-room client evals.
Add the durable hosted-authority control plane with checksum-pinned migrations, CAS/idempotent lifecycle writes, append-only decisions/audit and a host-configured opaque router that opens no foreign ledger before allow.
Add a host-integrated authenticated read-only REST handler with RFC 9728 discovery, strict verifier and live-authority boundaries, bounded commitment reads, payload-free event metadata and one injected clock snapshot per request.
Add registered guarded mutation REST with a state-free operation catalog, mandatory caller-scoped durable idempotency, live revocation serialization, bounded portable envelopes and exact recovery for unknown cross-database commit outcomes.

Security

Upgrade `drizzle-orm` to 0.45.2 for corrected SQL identifier escaping. Public package manifests now derive external dependency versions from their source manifests, and wrapped driver errors retain safe contention classification.

Artifacts and attestations

v0.3.0

2026-07-23

Added

Opt-in independent completion resolution across Core, embedded client, CLI, local MCP and Console: frozen policies, proposals, challenges and explicit decisions. Validated tasks can no longer be completed by evidence alone.
Append-only evidence trust, proposal, challenge and validation records.

Artifacts and attestations

v0.2.0

2026-07-23

Added

`createLocalTasq`, the embedded TypeScript client, published as compiled ESM with declarations and certified on Node 22 and Bun.

Artifacts and attestations

v0.1.1

2026-07-23

Added

Public acquisition and universal-agent entrypoints: the versioned installer, `setup`, `demo` and the deterministic `agent install` helper.

Changed

Made the public adoption examples executable rather than illustrative.

Artifacts and attestations

v0.1.0

2026-07-23

Added

First published `@tasq-run/*` packages and checksummed native assets for macOS arm64 and Linux x64, built from protected GitHub Actions OIDC with npm provenance.

Artifacts and attestations