Aller au contenu principal

Changelog

Quoi de neuf dans Tale ?

Les notes de version de Tale, de la plus récente à la plus ancienne — publiées depuis GitHub Releases.

Source : GitHub Releases.

Le tableau des tâches d’un projet Tale affiche les tâches, leur statut et la personne responsable.
Garde une vue d’ensemble du projet, du travail proposé à la vérification et à la finalisation.

v0.5.88

Highlights

  • A coding agent's sign-in now ends when it can no longer be trusted. A refresh token used a second time disconnects that app everywhere the person connected it. A changed or admin-set password, a removed passkey or a reset two-factor sign-in disconnects every agent the person connected, in every organization; the audit log names the act as the platform's, or as the admin's when an admin did it. An erasure deletes the person's connections in that organization and keeps only audit rows that have lost their personal data, and the erasure result counts Coding agent connections among the categories it removed. Coding-agent sign-in stays off unless an operator turns it on, so nothing changes for a deployment until then.
  • A person is told why an agent of theirs stopped working. When something other than themselves disconnects an agent they connected (a reused refresh token, a changed sign-in method, or an owner's or admin's disconnect, which names them), they get a notice in their bell, in the organization the connection was to, in English, German or French. It is never e-mailed, cannot be muted and is not shown to admins; their own disconnects and a seat that ended tell nothing.
  • Old coding-agent connections and unused apps are cleaned up every day. A connection that ended is removed 90 days later, an app that registered and was never connected a day after it appeared, and tokens a month after they expired or were retired. An app with a live connection and an organization's own sign-in application stay.
  • The screen where a coding agent asks to connect. The person sees who is asking, with a Verified domain or Unverified app badge and where the browser goes back to; chooses the organization (those they may not use stay listed with the reason) and how much the agent may do (read and test, edit automations, edit and deploy, and on top of those reading and changing settings and connecting AI subscriptions); and sees what that comes to. The server judges the choice again when they connect, and a sign-in older than a day asks for the password before more than reading is granted. It builds on two design-system components, ConnectionIdentity and PermissionSummary, and on a data route, and it appears only where an operator turns coding-agent sign-in on.
  • Releasing is no longer limited to one full release per UTC day: a version goes out whenever it is needed, a second one on the same day included.

Upgrade notes

  • From v0.5.87 to v0.5.88 no database migration, configuration change or manual step is needed, and the API contract stays at 3.43.0. The backend schedules one new daily sweep, maintenance.agent_connections_gc; it removes only the agent-connection records described above, and a deployment that never enabled coding-agent sign-in has none.
  • Where an operator has turned coding-agent sign-in on, a password change, a removed passkey or a two-factor reset now also disconnects the person's agents. Tell members to connect them again afterwards.
  • Upgrading from v0.5.86 or earlier also applies two migrations that the v0.5.87 notes leave out, 0203_agent_connections.sql and 0203_auth_factor_authorities.sql. Both only add tables and their indexes, and the previous image is safe to keep serving while they apply.
  • Since v0.5.87 a session that a trusted-header proxy or single sign-on opened for one organization can use only that organization; its other organizations answer ORG_FORBIDDEN, which the dashboard currently words as "not a member" (#4908). The new operator setting TALE_ORG_SESSION_CUTOVER_AT is for a second rollout, once every old writer has retired: leave it empty until then, and when you set it, give an exact UTC timestamp with milliseconds that has already passed (for example 2026-10-14T00:00:00.000Z), because the backend refuses to start on any other value.
  • Known limitation since v0.5.87: signing in on a browser that still holds another person's valid session is refused with ORG_FORBIDDEN, and the sign-in page words it as a wrong password. Sign out of the other account first (#4919).
  • This corrects an upgrade note of v0.5.87: the release workflow no longer admits one full cohort per UTC day. Adopt the service images and CLI of each version as one validated cohort, as before; recovery of a published tag is unchanged. Releasing Tale has the procedure.

API contract changes

None in this range. The contract stays at 3.43.0: 161 operations.

What's Changed

Voir sur GitHub

v0.5.87

Highlights

  • Automation tests now run from the editor with per-test failures and path coverage. Try the draft or an individual step, save a run as a regression test, and see test counts on saved versions and deployments.
  • MCP automation editing can page version history, compare saved versions and restore an older version as the newest. Deployment replies distinguish a rollback from another deployment. Agents can also run a draft on simulated outputs or one step at a time, and run the tests they name and read each run's output, effects and steps; both draw from the organization's budget for runs that are never stored.
  • Coding agents can read which writes an automation may make without a person approving each run and, with an owner's or admin's key, allow or withdraw them. The plan of an allowance names that it removes a person's approval, and the audit log records the change as made through MCP.
  • Release pipelines can read version history in pages, compare saved versions and read the deployment history over REST, then deploy a saved version — naming the one they expect to replace — or restore an older one. A deploy still passes the version's checks and tests. They can also run a saved version's tests, all or by name, and read the verdict the version carries and who made it.
  • Testing and version controls now keep selected tabs and keyboard focus visible in contrast themes. Code fields in dialogs let keyboard users reach the next control, and run playback fits phone screens. Automation authors can choose Save as test… directly from a Runs row menu.
  • Chat can make downloadable files and use skills from your library. Text files need no sandbox; documents, spreadsheets and charts use a light sandbox when code is needed.
  • Person-started standard agent tasks can make files and search knowledge on the shared chat engine without starting a full workspace. Work needing a repository, browser or hours hands over to the agent's full workspace with its brief and files.
  • OpenAI subscription credentials can carry the ChatGPT account identity Codex harness runs need in tasks and automations.
  • Resumed harness task turns receive their current execution identity, and resumed Codex turns receive the current organization and agent guidance. Retries and restarts open with this update before the task brief, even when the conversation retained older guidance.
  • SSO discovery skips unreadable saved connections in unrelated organizations. Signing in to an explicitly selected organization still reports its own connection errors.
  • Saved PII custom rules survive mode changes and disable/re-enable, including policies that use only custom rules. Unreadable rules show Retry and prevent edits from replacing the saved array.
  • After a successful deployment recovery, removing the takeover declaration allows configuration-only updates without an unnecessary runtime rollout when the retained receipt proves the unchanged basis.
  • Admins can set retention for automation versions, keeping versions needed by live deployments, runs, recent history and legal holds.
  • Historical package publication can fill missing ui-vX.Y.Z and marketing-ui-vX.Y.Z tags from their immutable release source without rewinding dist/ui or dist/marketing-ui. Publication uses maintained tooling and reads historical source as data.

Upgrade notes

  • Previously lost custom PII rules are not restored automatically. In Governance > Guardrails, review PII protection and re-enter any missing custom patterns before relying on the policy.
  • The backend applies migrations 0200_automation_test_reports.sql, 0201_provider_subscription_account_id.sql and 0202_agent_run_engine.sql at boot. The first two add nullable fields; earlier test verdicts remain readable without a detailed report. Migration 0202 adds a non-null engine defaulting to harness, keeping existing runs and old-image inserts on that route during a roll, plus nullable transcript/handoff fields and a native-run index. No manual data backfill is required. Include them in the normal deployment and database backup procedure.
  • To use an OpenAI Subscription key, provide its ChatGPT account ID; OpenAI broker tokens must supply the vendor's account_id. Restrict models to those your subscription supports and verify a Codex harness task. Subscriptions use their provider-bound harnesses in tasks and automations; the native engine requires an API-key or environment credential.
  • With Agent runtime on Automatic, person-started standard agents can use the native engine. Starts by automations, schedules or other agents use the full workspace, as do custom agents and runs with equipped connectors, platform tools or secrets, subscription models or pinned runtimes. To require it from the start, pin the standard agent's runtime. Each organization runs at most four native task runs concurrently; further native runs queue. Full-workspace workers keep their configured Agent workers limits.
  • Review Chat tools before rollout: Files, Skills and Code default to on and govern chat only. The new Light sandboxes quota defaults to two per organization, is shared by chat and native task code, and counts toward allocated capacity. Clients reading chat messages must accept the additive artifact part introduced in API contract 3.36.0.
  • The current API contract is 3.43.0. The MCP endpoint can answer the codes INSUFFICIENT_SCOPE and ORG_MISMATCH, and a sign-in challenge on its 401, but only where an operator lets coding agents sign in; no deployment does by default. MCP clients that read settings should accept the kind automation-approval, which has a read-only config and changes by the acts grant and revoke, the plan action unchanged for an act that changes nothing, and the code APPROVAL_ALLOWANCE_INVALID. They should also accept cursors beside get_settings' nextCursor, which is now set only after a read of one kind, and the codes USE_SUBSCRIPTION_FLOW (a subscription credential) and SETTINGS_NOT_FOUND (a provider Tale ships). MCP clients should accept the added version-history and comparison metadata, plus deployment kind values deploy, rollback and unchanged; test_automation and run_automation take new optional arguments, test_automation may answer recorded and detailOmitted, and both may answer RATE_LIMITED with data.retryAfterMs when the organization's budget for runs that are never stored is spent. REST clients should accept the added version fields; GET /api/v1/automations/{name}/versions also answers isDone and continueCursor, and still answers the whole history when no limit is named. POST and GET /api/v1/automations/{name}/versions/{version}/tests are new. Restoring an automation creates a new latest version; it does not rewrite saved history.
  • Automation-version cleanup stays off until an admin enables it. Review and apply retention bounds before enabling deletion; older organization files use bundled bounds for the new category until you add their own. Cleanup preserves the live version, the last ten deployments' versions, versions used by runs, the newest twenty, the versions those were based on or restored from, and legal holds.
  • After bundle recovery reaches Ready, remove supersedesPendingBundle and prepare again before a configuration-only apply. Preserve the Ready receipt and recovery snapshot. New pending rollouts allow at most 64 takeovers; legacy journals with 65 entries remain recoverable. At the limit, recover the exact pending bundle before selecting another rollout.
  • Adopt the service images and CLI as one validated cohort. Release publication admits one new full cohort per UTC day and permits same-version recovery. Retry or backfill package tags from main with both release_tag and the tag's full source_sha; identical existing tags are kept, and conflicting tags require reconciliation without moving them.

API contract changes

The contract moved from 3.35.0 to 3.43.0 (155 → 161 operations). Read the API reference's versioning section before upgrading a pinned client.

Added operations:

  • GET /api/v1/automations/{name}/compare
  • GET /api/v1/automations/{name}/deployments
  • GET /api/v1/automations/{name}/versions/{version}/tests
  • POST /api/v1/automations/{name}/deployments
  • POST /api/v1/automations/{name}/versions/{version}/restore
  • POST /api/v1/automations/{name}/versions/{version}/tests

Removed operations: none.

Changelog

3.43.0 — 2026-10-13: the MCP endpoint takes a coding agent's access token in place of an API key where the operator lets agents sign in (off unless set). Its 401 then carries the challenge RFC 9728 describes — resource_metadata and scope — and a credential that was presented and refused adds error="invalid_token". A token opens the MCP endpoint alone; on every other route it is read as an API key. A call made through a connection acts in the organization the connection was made for (naming another in X-Organization-Slug answers 403 ORG_MISMATCH) and holds the access its person allowed: a request that calls a tool needing access the connection lacks is refused whole with 403 INSUFFICIENT_SCOPE and the challenge error="insufficient_scope", which names what to ask for, and a call that the person's role could never be given the access for answers the same code as a tool error. tools/list is unchanged. The codes INSUFFICIENT_SCOPE and ORG_MISMATCH join the registry. No REST operation changes.

What's Changed

Voir sur GitHub

v0.5.86

Highlights

  • The CLI release lane now builds from the same source-bound workflow contract as service images, so native binaries and their release tag cannot silently come from different commits.
  • Release metadata and image identity checks now exercise the exact prepared source SHA used by the publication job.

Upgrade notes

  • Adopt v0.5.86 as one source-bound cohort through the normal Ops release workflow. It supersedes v0.5.85 for the managed deployment CLI and runtime; no database migration or manual configuration change is required.

API contract changes

None in this range. The contract stays at 3.35.0: 155 operations.

What's Changed

Voir sur GitHub

v0.5.85

Highlights

  • Release publication now binds every service image, CLI artifact and GitHub release tag to the exact immutable source commit used by the release workflow. A publication whose tag drifts to another commit fails before it can be adopted.

Upgrade notes

  • Deployments should adopt v0.5.85 as one source-bound cohort through the normal Ops release workflow. No database migration or manual configuration change is required; the managed deployment CLI and runtime remain paired.

API contract changes

None in this range. The contract stays at 3.35.0: 155 operations.

What's Changed

Voir sur GitHub

v0.5.84

Highlights

  • Subscription-backed fleet turns now start successfully when the organization reserves the request at no cost. The reservation reader accepts the zero-cost result only on the explicit subscription lane, while paid gateway turns continue to require a positive budget. #4793

Upgrade notes

  • No operator action or data migration is required. Deploy the platform image and its matching managed-deployment CLI together so the fleet manager and workers use the same reservation contract; the normal forward-only rollout has no database migration for this fix.

API contract changes

None in this range. The contract stays at 3.35.0: 155 operations.

What's Changed

Voir sur GitHub

v0.5.83

Highlights

  • Standing roles now wake when a project worker finishes, so a released slot can be refilled by the fleet manager without waiting for the next scheduled occurrence. Wake claims are scoped to the project and preserve the native task and review guards. #4608
  • Audit-log sealing runs asynchronously after the write path, while the integrity panel and verification checks continue to expose the exact source and chain state. The per-project work lock avoids serializing unrelated automation work behind the audit writer. #4700
  • Retained deployment observation is bound to the exact deployment reference before it reads runtime state, keeping recovery evidence tied to the operation it describes. #4755

Upgrade notes

  • Deploy the Tale platform and the managed-deployment CLI together. Tale CLI applies migrations 0168 and 0194 during the normal forward-only rollout; keep the existing backup, retained-release and configuration readback procedure until the new image is serving.
  • The autonomous-cycle schedules remain enabled and keep their native admission and per-task start limits. A worker or provider failure is recorded on its task and retried or rerouted according to the current project policy; no OpenRouter fallback is required for the OpenAI subscription workers.

API contract changes

The contract moved from 3.27.0 to 3.35.0 (143 → 155 operations). Read the API reference's versioning section before upgrading a pinned client.

Added operations:

  • GET /api/v1/projects/{id}/runs/{runId}/compare/{otherRunId}
  • GET /api/v1/projects/{id}/runs/{runId}/record
  • GET /api/v1/projects/{id}/runs/{runId}/record/items
  • GET /api/v1/projects/{id}/runs/{runId}/record/node
  • GET /api/v1/projects/{id}/runs/{runId}/replay
  • GET /api/v1/runs/{runId}/compare/{otherRunId}
  • GET /api/v1/runs/{runId}/record
  • GET /api/v1/runs/{runId}/record/items
  • GET /api/v1/runs/{runId}/record/node
  • GET /api/v1/runs/{runId}/replay
  • POST /api/v1/projects/{id}/runs/{runId}/replay
  • POST /api/v1/runs/{runId}/replay

Removed operations: none.

Changelog

3.35.0 — 2026-10-11: a notify.member connector node type tells members of the organization something from an automation — a title, words and a link (a page of the organization or an https address), in their notifications and by email — and answers who got it and who was skipped. StepFailure.reason gains NOTIFY_NOT_A_MEMBER and NOTIFY_RUN_LIMIT.

What's Changed

Voir sur GitHub

v0.5.82

Highlights

  • OpenAI gateway accounts now recover from revoked access-token generations without reusing rejected credentials. A late usage response cannot overwrite a terminal refresh refusal, so an affected account stays out of the shared pool until a new generation is available. Generic usage failures keep their existing retry behavior. #4664, #4667
  • tale deploy observe now keeps its retained-state checks and refusal phases bounded to read-only evidence, including portable lock and native-artifact validation. Observation never applies deployment changes or certifies readiness. #4648, #4655, #4657, #4660, #4662
  • CLI release checks now include every command suite, keep hashed workflow sources byte-stable on Windows, and reject receipt symlinks on platforms without O_NOFOLLOW. This prevents a platform-specific release from silently skipping a CLI suite or accepting a redirected receipt. #4668

Upgrade notes

  • Deploy the Tale platform and CLI together before adopting these fixes. The managed-deployments repository now pins its shared Tale CLI source to 13a41556bd017e4fade3165d988974e7a3bea184 in ops commit 1d932920.
  • No database migration is included in this release. Keep the existing backup, retained-release and configuration readback procedure, and verify the gateway account state after rollout before sending new task turns.

API contract changes

None in this range. The contract stays at 3.27.0: 143 operations.

What's Changed

Voir sur GitHub

v0.5.81

Highlights

  • Projects can run separate task workers for the same agent, wake a standing coordinator when capacity becomes available, and review a declared batch of up to twenty captured targets. Each review still requires independent evidence. The Tale CLI can provision pristine operational review contexts and manage agent model selections alongside the project's other configuration. #4603, #4608, #4616, #4632

  • Automation transforms have their own analysis budget, allowing bodies of up to 16,384 UTF-16 code units and 4,096 tokens. Ordinary expressions keep their smaller limits. The CLI compiler and runtime validator enforce the same bounds. #4639

  • A delayed failure while reading an organization's default provider credentials no longer removes a newer cached read. Changing the defaults still invalidates that organization's cache; credentials remain scoped to the organization. #4641

  • Maintainers can finish cancelled validation graphs across all seven supported CI workflows when their substantive jobs have ended and only unallocated summary jobs remain. Source, queue and job checks still refuse active work; cancellation never counts as successful validation. #4638

  • Sandbox commands no longer inherit the runtime daemon's token. The daemon retains it for its authenticated API. #4635

  • tale deploy observe reports retained deployment identity, configuration and runtime diagnostics without applying changes. Its evidence is read-only observation; it does not certify deployment acceptance or readiness.

Upgrade notes

  • v0.5.78, v0.5.79 and v0.5.80 were stopped before complete release publication. Preserve their existing tags, package snapshots and any architecture images; do not rebuild or retag them. This forward release includes their source changes. The retained v0.5.80 notes describe those features and their inherited upgrade requirements; they are not evidence that those versions finished deploying.
  • Back up the application and knowledge databases, configuration, keys and external stores. Follow the supported legacy runtime bridge and confirm zero unfinished legacy automation runs before managed protocol-2 cutover. Apply the complete ordered migration inventory, retain execution holds and reserve the vector-copy disk headroom described in the v0.5.79 upgrade requirements. Do not roll back by manually starting an older image.
  • Upgrade the CLI and runtime together before applying configurations containing larger transform bodies or managed review contexts. A configuration-only update cannot install the parser, worker or migration changes. A deployment that must cross the legacy bridge can preserve its installed packs with an explicitly empty configuration selection; verify the retained identities before and after that bridge, then apply the new packs with the compatible runtime.
  • Enable coordinator wakes explicitly with wakeOnSlotFreed: true through the Tale CLI. Only one schedule may claim a project's wake. Review contexts must be pristine operational tasks, and each captured target still requires a matching independent decision. Agent definitions, model selections and review contexts grant no credentials and do not guarantee twenty concurrent workers; provider, organization and host limits still govern admission.
  • Retain the v0.5.80 guidance for API contract 3.27.0, MCP clients, newly counted usage dimensions and private-route deployment acceptance. Finish the runtime rollout before relying on uniform behavior across processes.

API contract changes

None in this range. The contract stays at 3.27.0: 143 operations.

What's Changed

Voir sur GitHub

v0.5.80

Highlights

  • Project agents run each task in a separate worker. A busy agent can take another eligible task without another agent definition; organization and host capacity still limit concurrent work. Agent workspaces survive sandbox replacement, and active turns can recover from a temporary spawner restart. #4603, #4560, #4585
  • Managed schedules can wake a project's standing coordinator when an agent frees its slot. Concurrent completions coalesce into one pending wake, with normal admission checks and retry backoff. This reduces the wait for the next scheduled dispatch without bypassing provider or host limits. #4608
  • Operational review contexts can handle a declared batch of up to twenty captured reviews. Each target still needs its independent review decision and matching evidence. A review context does not create another review of its own report; requesting changes completes the review action while leaving the implementation unfinished. #4616
  • Coding agents can read and edit automation versions through MCP with version conflict checks, retained settings, project installation controls, saved-version tests and deployment history. Definition changes record their source and client in the audit trail. Discovery tools, resources and prompts help clients find available models, skills and automation inputs. #4592
  • Usage limits now cover transcriptions, knowledge embeddings, chat titles and later model rounds. Indexing that reaches a usage limit can wait and resume. Subscription-backed managed agent turns count as requests at no model cost; connector calls are counted separately from model requests. Agent runs started with an API key also count toward that key. #4525, #4541, #4520, #4555, #4550, #4558
  • Sandbox maintenance reduces repeated Docker and BuildKit work and adds a critical disk-pressure tier for reclaiming released idle sessions. Busy sessions use graceful shutdown, and Kubernetes placement and startup checks retain resource bounds. #4604, #4602, #4597
  • Private deployment acceptance can use a captured gateway container while keeping the canonical HTTPS hostname, certificate validation and serving-process checks. Its receipt records the gateway and network used. #4617
  • Maintainers can retire obsolete merge-group validation runs with source and queue checks. A separate finishing mode handles only an exact reviewed workflow whose real jobs have ended and whose remaining summary jobs are unallocated. Requests and uncertain outcomes are retained; cancellation never counts as passing CI. #4613

Upgrade notes

  • Use this release's reviewed CLI and matching runtime images. Back up the application and knowledge databases, configuration, keys and external stores. When upgrading from v0.5.77 or earlier, also follow the inherited v0.5.79 upgrade requirements: the supported legacy runtime bridge, zero unfinished automation census, managed protocol-2 cutover, vector-copy disk headroom and forward-repair restrictions still apply. Those retained notes are source history, not proof of a completed deployment. Do not clear execution holds or retag earlier versions.
  • Apply the complete ordered application migration inventory. This release adds direct-call holds, chat project stamps, agent-run key attribution, knowledge author attribution, agent handles and worker records, pending project wakes, operational review contexts, MCP activity and automation version attribution. Existing migration filenames remain immutable, including the two distinct 0164 files. The new runtime and its CLI must agree on the installed database floor; do not roll back by manually starting an older image.
  • Enable early coordinator wakes explicitly with wakeOnSlotFreed: true in the schedule's managed configuration and apply it through the Tale CLI. Only one schedule may claim a project's wake, and existing schedules retain their current behavior until configured. Enroll a pristine operational task before using review batches; an implementation task with retained work cannot be repurposed as a review context. Neither feature grants credentials or guarantees a particular number of concurrent workers.
  • Update integrations for REST/OpenAPI contract 3.27.0 and the published MCP schemas. MCP argument validation now returns structured tool errors, and reads offer structured content. Clients using the newer protocol revision must send its matching request metadata and headers; earlier supported protocol revisions remain available. Automations installed only in projects the caller cannot read are hidden on both MCP and REST reads.
  • Review request and token caps for subscription-backed agents, and allow for the newly counted transcription and embedding work. Earlier usage is not backfilled, and older processes do not write every new usage dimension during a mixed-version rollout. Finish retiring those processes before relying on uniform accounting. See Policies and limits.
  • A deployment adapter using --origin-container must supply the full running gateway ID and consume the optional serving.originRoute receipt. A gateway restart, identity change or ambiguous network makes acceptance fail. This proves the captured private route; it does not prove public DNS or off-host reachability. Keep ordinary monitoring for those paths.

API contract changes

The contract moved from 3.22.0 to 3.27.0 (143 → 143 operations). Read the API reference's versioning section before upgrading a pinned client.

Added operations: none.

Removed operations: none.

Changelog

3.27.0 — 2026-10-09: the MCP endpoint speaks protocol revision 2025-11-25 beside 2025-06-18 and 2025-03-26, answers an MCP-Protocol-Version it does not speak with -32022 naming the supported revisions, answers initialize with instructions, and reports this contract version as serverInfo.version. Tool arguments that miss a tool's schema are a tool result INVALID_ARGUMENTS listing every problem under data.issues (path, code, message) where they were a -32602 naming the first; a refusal a tool throws keeps its code, and an unexpected failure is INTERNAL_ERROR with data.requestId. Read tools declare an outputSchema and answer structuredContent; the text block is compact JSON. Tools carry _meta hints (anthropic/requiresUserInteraction on deploy_automation and set_trigger, anthropic/maxResultSizeChars). Tools that execute an automation, and answer_run_ask, draw from the rest:execute budget (RATE_LIMITED, data.retryAfterMs), and run_deployed, a live start_run, cancel_run and delete_trigger refuse a member before anything runs. A request whose Origin the deployment does not accept is logged, and refused with 403 ORIGIN_FORBIDDEN where the operator enforces the rule (TALE_MCP_ORIGIN_ENFORCE); clients that send no Origin are not judged. The MCP authoring tools reach the editor's parity: save_automation takes baseVersion (a version saved since is refused, AUTOMATION_VERSION_STALE with data.latestVersion), create, projectId and the version's settings, taskContract and presentation (left out = kept from the latest version, named in carried; null = none), answering carried and baseVersionChecked; get_automation answers the whole version beside meta and automation (latestVersion, deployedVersion, the three fields, createdVia, clientName, projectIds, trigger); a version records the door it was saved through (createdVia) and the client; test_automation tests a saved version by name and records its verdict; deploy_automation takes expectedDeployedVersion (AUTOMATION_DEPLOYMENT_STALE) and answers previousVersion; start_run takes mode — live, the default, or mock, any saved version and open to every member; list_runs filters by mode and statuses (the REST list's, quarantined included) and pages with cursor / nextCursor; list_versions answers createdVia, clientName and deployments. New tools: delete_automation, set_automation_projects, answer_run_ask (each asking the person first, like deploy_automation and set_trigger) and get_automation_metrics; get_run names the question a run waiting on a person asked under run.ask (askId, question, nodeId, expiresAt), which answer_run_ask answers, and takes detail (input, output, trace, effects; [] answers the status alone, for polling); run_automation advertises only mode: "mock", the one it ever ran here. Member reads tighten on MCP and REST alike: an automation installed only in projects the key holder cannot read is left out of list_automations and GET /api/v1/automations, and its reads answer AUTOMATION_NOT_FOUND (REST 404) — its runs by name too, unless runs of it are in the URL's scope — and an MCP save onto it is refused as AUTOMATION_NAME_TAKEN. Every change to an automation's definition is audited; one made through the REST API is recorded as the key's (actor type api, via: api-key, the key's id and the request id). validate_automation takes detail ([] answers the issues alone), and validation warns when a document names what the organization lacks: SKILL_UNKNOWN, CONNECTOR_NOT_CONNECTED, SECRET_UNKNOWN (told only to owners, admins and developers), HARNESS_UNKNOWN and EVENT_UNKNOWN, with params and a closest name; a save's warnings carry them too. New discovery tools list what a document may name: list_models, list_harnesses, list_skills, list_connectors, list_agent_secrets (names and masked previews for owners, admins and developers, an empty list for anyone else), list_projects and list_events. The endpoint serves resources and prompts: resources/list (the references tale://docs/{authoring,triggers,validation,skill}, the node kinds tale://catalog/{kind}, then every automation the key holder can see, 100 a page with nextCursor), resources/templates/list (tale://automations/{name}, …/versions/{version}, tale://runs/{runId}) and resources/read, which reads exactly as the matching tool does and answers a read that finds nothing with -32002 and the refusal's data.code; prompts/list and prompts/get (edit_automation, debug_failed_run, add_trigger, attaching what they are about). get_docs takes topic; get_catalog narrowed to a core kind answers the kind's reference. In a batch, a resource read or listing and a prompt draw from the request budget like a tool call. The session route GET /api/app/mcp/skill downloads the Tale skill (SKILL.md). The endpoint also speaks MCP 2026-07-28, request by request beside the 2025 revisions: a request whose params._meta names its revision (with the client's capabilities and, optionally, its name) needs no initialize; server/discover answers the revisions, capabilities and instructions; its MCP-Protocol-Version, Mcp-Method and Mcp-Name headers must say what its body says (else 400 -32020), a missing envelope is 400 -32602, a revision it does not speak is 400 -32022 (now listing 2026-07-28 too), a batch is refused (400 -32600), and initialize, ping and unknown methods are 404 -32601; every result adds resultType and the server under _meta, the lists, reads and server/discover add ttlMs and cacheScope: "private", and a read that finds nothing is -32602 with the same data. The client a 2026-07-28 request names is recorded with the call and on the versions and audit rows it writes. No other REST operation changes. A script that matched the -32602 argument error, or read the indented text, reads the tool result instead.

What's Changed

Voir sur GitHub

Dernière mise à jour le 13 octobre 2026.

Vois le prochain workflow de ton équipe dans Tale

Apporte une tâche à une démo guidée ou crée un espace et essaie avec tes propres agents.