Skip to content

CHANGELOG.md

Latest commit

 

History

History
1396 lines (1027 loc) · 67.4 KB

CHANGELOG.md

File metadata and controls

1396 lines (1027 loc) · 67.4 KB
Raw

@objectstack/service-ai

10.3.0

Patch Changes

  • @objectstack/spec@10.3.0
  • @objectstack/core@10.3.0
  • @objectstack/types@10.3.0
  • @objectstack/formula@10.3.0

10.2.0

Patch Changes

  • Updated dependencies [b496498]
    • @objectstack/spec@10.2.0
    • @objectstack/core@10.2.0
    • @objectstack/formula@10.2.0
    • @objectstack/types@10.2.0

10.1.0

Patch Changes

  • Updated dependencies [49da36e]
  • Updated dependencies [ac79f16]
    • @objectstack/spec@10.1.0
    • @objectstack/core@10.1.0
    • @objectstack/formula@10.1.0
    • @objectstack/types@10.1.0

10.0.0

Minor Changes

  • e411a82: feat(ai): split ask/build agents by surface + tool scoping (ADR-0063/0064).

    Two kernel agents bound by surface, not a per-turn classifier. SkillSchema gains surface: 'ask'|'build'|'both' and AgentSchema gains surface: 'ask'|'build' (ADR-0063 §3); an agent's tools are exactly the union of its surface-compatible skills' tools — incompatible binding is a load error in resolveActiveSkills (ADR-0064 §3). The ask agent is now data-only (the ADR-0040 unified "INTENT FIRST" classifier and the buildRegisterActive degradation shim are removed); a new schema_reader (surface:'both') owns the shared reads describe_object/list_objects/query_data so the build agent reuses them without dual-listing. *.agent.ts is closed to third parties: the agent metadata-type is allowRuntimeCreate:false, allowOrgOverride:false and the runtime catalog lists only platform agents (ADR-0063 §2). Renames data-chat-agent.ts→ask-agent.ts, DEFAULT_DATA_AGENT_NAME→ASK_AGENT_NAME (the data_chat/metadata_assistant aliases stay resolvable).

  • be07ce7: Rename the built-in data agent data_chat → ask (Path A: friendly console URL == real id). Back-compat preserved via a new process-wide alias registry: AgentRuntime.loadAgent normalizes legacy names, so /agents/data_chat/chat and persisted agent_id='data_chat' keep resolving. registerAgentAlias() is exported so other packages register their own renames (cloud AI Studio: metadata_assistant→build). The plugin prunes the stale legacy agent record on upgrade so the catalog isn't doubled.

Patch Changes

  • Updated dependencies [d7ff626]
  • Updated dependencies [2a1b16b]
  • Updated dependencies [e16f2a8]
  • Updated dependencies [cfd86ce]
  • Updated dependencies [e411a82]
  • Updated dependencies [a581385]
  • Updated dependencies [d5f6d29]
  • Updated dependencies [220ce5b]
  • Updated dependencies [3efe334]
  • Updated dependencies [feead7e]
  • Updated dependencies [6ca20b3]
  • Updated dependencies [5f875fe]
  • Updated dependencies [b469950]
  • Updated dependencies [48a307a]
  • Updated dependencies [25fc0e4]
    • @objectstack/spec@10.0.0
    • @objectstack/formula@10.0.0
    • @objectstack/core@10.0.0
    • @objectstack/types@10.0.0

9.11.0

Patch Changes

  • Updated dependencies [e7f6539]
  • Updated dependencies [2365d07]
  • Updated dependencies [6595b53]
  • Updated dependencies [fa8964d]
  • Updated dependencies [36138c7]
  • Updated dependencies [a8e4f3b]
  • Updated dependencies [4c213c2]
  • Updated dependencies [2afb612]
    • @objectstack/spec@9.11.0
    • @objectstack/core@9.11.0
    • @objectstack/formula@9.11.0
    • @objectstack/types@9.11.0

9.10.0

Patch Changes

  • Updated dependencies [db02bd5]
  • Updated dependencies [641675d]
  • Updated dependencies [1f88fd9]
  • Updated dependencies [94e9040]
  • Updated dependencies [1f88fd9]
  • Updated dependencies [1f88fd9]
    • @objectstack/spec@9.10.0
    • @objectstack/formula@9.10.0
    • @objectstack/core@9.10.0
    • @objectstack/types@9.10.0

9.9.1

Patch Changes

  • @objectstack/spec@9.9.1
  • @objectstack/core@9.9.1
  • @objectstack/types@9.9.1
  • @objectstack/formula@9.9.1

9.9.0

Patch Changes

  • Updated dependencies [84249a4]
  • Updated dependencies [11af299]
  • Updated dependencies [d5774b5]
  • Updated dependencies [134043a]
  • Updated dependencies [90108e0]
  • Updated dependencies [9afeb2d]
  • Updated dependencies [6bec07e]
  • Updated dependencies [601cc11]
  • Updated dependencies [d99a75a]
  • Updated dependencies [575448d]
    • @objectstack/spec@9.9.0
    • @objectstack/core@9.9.0
    • @objectstack/formula@9.9.0
    • @objectstack/types@9.9.0

9.8.0

Patch Changes

  • Updated dependencies [c17d2c8]
  • Updated dependencies [97c55b3]
  • Updated dependencies [1b1f490]
    • @objectstack/formula@9.8.0
    • @objectstack/spec@9.8.0
    • @objectstack/core@9.8.0
    • @objectstack/types@9.8.0

9.7.0

Patch Changes

  • Updated dependencies [82c7438]
  • Updated dependencies [417b6ac]
  • Updated dependencies [ff0a87a]
    • @objectstack/formula@9.7.0
    • @objectstack/spec@9.7.0
    • @objectstack/core@9.7.0
    • @objectstack/types@9.7.0

9.6.0

Patch Changes

  • Updated dependencies [d1e930a]
  • Updated dependencies [71578f2]
  • Updated dependencies [bb00a50]
  • Updated dependencies [5e3a301]
  • Updated dependencies [5db2742]
    • @objectstack/spec@9.6.0
    • @objectstack/formula@9.6.0
    • @objectstack/core@9.6.0
    • @objectstack/types@9.6.0

9.5.1

Patch Changes

  • Updated dependencies [ee72aae]
    • @objectstack/spec@9.5.1
    • @objectstack/core@9.5.1
    • @objectstack/formula@9.5.1
    • @objectstack/types@9.5.1

9.5.0

Patch Changes

  • Updated dependencies [d08551c]
  • Updated dependencies [707aeed]
  • Updated dependencies [7a103d4]
  • Updated dependencies [4b01250]
    • @objectstack/spec@9.5.0
    • @objectstack/core@9.5.0
    • @objectstack/formula@9.5.0
    • @objectstack/types@9.5.0

9.4.0

Minor Changes

  • b678d8c: feat(service-ai): open framework AI is query-only, declines app-building

    The unified data_chat persona (ADR-0040) advertised that it can BUILD or CHANGE the application, but that capability is supplied entirely by the cloud AI Studio plugin's metadata_authoring/solution_design skills. On the open single-env framework those skills are not registered, so the authoring tools never resolve — yet the LLM, still reading the "you can build" persona, role-played designing a whole system (emitting design docs it had no tools to execute).

    Fix: in buildSystemMessages, when no authoring (build-register) skill is active, append a deployment-capability note constraining the assistant to data/query and instructing it to decline build requests instead of pretending. Keyed off actual skill presence, so cloud/EE (AI Studio loaded) keeps the full build UX with zero extra wiring.

  • b678d8c: feat(service-ai): steer the agent to visualize_data for chart requests

    Small models sometimes answered a "draw a bar chart" request with a markdown TABLE instead of calling visualize_data — a tool-selection problem where the chart preference was buried as a low-priority guideline competing with "format with markdown tables".

    • data-explorer-skill.ts — adds a prominent "Choosing the right tool" section above the guidelines: chart intent (incl. CN terms 图表/柱状图/折线图/饼图/画图) MUST call visualize_data, never substitute a table; reconciles the table-formatting guideline and fixes duplicate guideline numbering.
    • visualize-data.tool.ts — strengthens the tool description to be imperative ("the ONLY tool that draws a chart… if you already fetched the numbers, still call visualize_data to render them").

    Prompt-only tuning — no behavior/contract change.

  • b678d8c: feat(service-ai): visualize_data tool — return charts from AI data queries

    Adds a visualize_data AI tool so the data-query assistant can answer with a CHART instead of plain text/markdown. The tool runs an aggregation through the existing analytics service (auto-inferred cube), maps the result into the SDUI <chart> contract, and emits it to the client as a data-chart custom stream part (the same onProgress channel data-build-progress already uses). It also returns a compact textual summary so the model narrates the answer alongside the rendered chart.

    • tools/visualize-data.tool.ts — tool definition, handler, and register fn (function+field → analytics measure key; single dimension → x-axis; measures → series; chartType bar/line/pie/…).
    • plugin.ts — registers the tool when an analytics service is present and persists it as tool metadata in lockstep (Studio visibility).
    • skills/data-explorer-skill.ts — exposes visualize_data plus chart trigger phrases and guidance to prefer it for "chart/plot/trend/breakdown" requests.

Patch Changes

  • b678d8c: fix(service-ai): resolve the current object for AI chat across languages

    The console assistant reported "can't find the X object" when asked to analyse the object on the current page — most visibly for non-English prompts. Three compounding gaps fixed:

    • SchemaRetriever.tokenise() dropped all CJK text, so a Chinese request yielded zero terms; it now emits CJK single-char + bigram terms.
    • Nothing fed the current object's schema to the agent, so "this object" could not be resolved without a lucky keyword hit. AgentRuntime.buildContextSchema Messages() now injects the current object's schema into the system prompt and both chat routes call it.
    • ToolExecutionContext (and the ai-service spec contract) gains currentObjectName/currentViewName; routes thread them through and query_data falls back to the current object when keyword retrieval is empty (so the open edition, which lacks describe_object/list_objects, still resolves the page's object).

  • Updated dependencies [060467a]

  • Updated dependencies [0856476]

  • Updated dependencies [b678d8c]

  • Updated dependencies [b678d8c]

  • Updated dependencies [b678d8c]

    • @objectstack/spec@9.4.0
    • @objectstack/core@9.4.0
    • @objectstack/formula@9.4.0
    • @objectstack/types@9.4.0

9.3.0

Minor Changes

  • d100707: AI provider misconfiguration is now visible, rejected at save time, and recoverable from the UI. Background: a half-saved ai settings row (provider=cloudflare, empty key) silently overrode env auto-detection and the only symptom was a bare "Bad Request" in chat.

    • GET /api/v1/ai/status — active adapter provenance: source (explicit/env/settings/fallback), provider, model, plus settingsError explaining why saved settings were NOT applied. AIServicePlugin tracks this through boot detection, settings rebuilds, and resets.
    • Save-time validation in SettingsService.setMany (fulfilling the spec promise that required is enforced server-side): visible+required fields and pattern mismatches reject the whole batch with field-level errors (400 SETTINGS_VALIDATION). Visibility expressions (${data.provider === '…'}) are evaluated server-side by a restricted-grammar parser; unparseable expressions and all-null patches (resets) stay lenient. gateway_model / cloudflare_model gain provider/model patterns.
    • Built-in reset settings action for every namespace (SettingsService.resetNamespace), overridden for ai to also re-run env adapter detection immediately; the AI manifest ships a "Reset to environment defaults" button — no more hand-editing sys_setting.
    • Chat/agent/assistant stream errors are enriched with the active adapter description and actionable hints (400 → model-id format, 401/403 → credential, 404 → unknown model, 429 → rate limit) instead of a bare HTTP status.

Patch Changes

  • Updated dependencies [1ada658]
  • Updated dependencies [3219191]
  • Updated dependencies [290f631]
  • Updated dependencies [50b7b47]
  • Updated dependencies [f15d6f6]
  • Updated dependencies [f8684ea]
  • Updated dependencies [b4765be]
    • @objectstack/spec@9.3.0
    • @objectstack/core@9.3.0
    • @objectstack/formula@9.3.0
    • @objectstack/types@9.3.0

9.2.0

Patch Changes

  • Updated dependencies [2f57b75]
  • Updated dependencies [2f57b75]
    • @objectstack/spec@9.2.0
    • @objectstack/core@9.2.0
    • @objectstack/formula@9.2.0
    • @objectstack/types@9.2.0

9.1.0

Patch Changes

  • Updated dependencies [b9062c9]
    • @objectstack/spec@9.1.0
    • @objectstack/core@9.1.0
    • @objectstack/formula@9.1.0
    • @objectstack/types@9.1.0

9.0.1

Patch Changes

  • Updated dependencies [1817845]
    • @objectstack/spec@9.0.1
    • @objectstack/core@9.0.1
    • @objectstack/formula@9.0.1
    • @objectstack/types@9.0.1

9.0.0

Patch Changes

  • f533f42: Settings namespace environment overrides now use the canonical ObjectStack OS_<NAMESPACE>_<KEY> form, with no unprefixed aliases. For example, ai.openai_base_url is now OS_AI_OPENAI_BASE_URL, and feature_flags.ai_enabled is now OS_FEATURE_FLAGS_AI_ENABLED.

    The AI service now treats a stored or env-locked provider=memory setting as an explicit override, while the manifest default still leaves boot-time provider auto-detection intact.

    The auth plugin now binds the auth settings namespace to better-auth runtime configuration, exposes an extension hook for provider packages, and includes a basic Google sign-in implementation configured either in Setup → Authentication or by deployment-level GOOGLE_CLIENT_ID and GOOGLE_CLIENT_SECRET.

  • Updated dependencies [4c3f693]

  • Updated dependencies [0bf39f1]

  • Updated dependencies [f533f42]

  • Updated dependencies [1c83ee8]

    • @objectstack/spec@9.0.0
    • @objectstack/core@9.0.0
    • @objectstack/formula@9.0.0
    • @objectstack/types@9.0.0

8.0.1

Patch Changes

  • @objectstack/spec@8.0.1
  • @objectstack/core@8.0.1
  • @objectstack/types@8.0.1
  • @objectstack/formula@8.0.1

8.0.0

Patch Changes

  • a46c017: feat(ai): actions opt in to being AI tools via an ai: block (ADR-0011)

    Realigns ADR-0011 with its original opt-in design. An Action becomes an AI-callable tool only when its metadata sets ai.exposed: true, which requires an explicit, LLM-facing ai.description (≥40 chars, distinct from the UI label). There is no heuristic auto-exposure and no description derived from the label — a clean break from the first implementation's opt-out aiExposed flag, which is removed (no compatibility shim; the platform has not shipped).

    The ai: block also carries category, paramHints (per-parameter JSON-Schema refinement), outputSchema (summarised into the tool description for chaining), and requiresConfirmation (overrides the destructive-action HITL default). AIToolDefinition is extended to carry category / outputSchema / objectName / requiresConfirmation. The @objectstack/service-ai bridge (action-tools.ts) now gates on opt-in, merges paramHints, and emits a lint warning when an exposed destructive-looking action asserts itself safe via ai.requiresConfirmation: false.

  • Updated dependencies [a46c017]

  • Updated dependencies [b990b89]

  • Updated dependencies [99111ec]

  • Updated dependencies [d5a8161]

  • Updated dependencies [5cf1f1b]

  • Updated dependencies [9ef89d4]

  • Updated dependencies [3306d2f]

  • Updated dependencies [c262301]

  • Updated dependencies [bc44195]

  • Updated dependencies [9e2e229]

    • @objectstack/spec@8.0.0
    • @objectstack/core@8.0.0
    • @objectstack/formula@8.0.0
    • @objectstack/types@8.0.0

7.9.0

Patch Changes

  • ac1fc4c: fix(ai): authoring tools can see their own drafts; blueprint surfaces the package to bind to

    Two gaps that broke the multi-step "build app → author a flow for it" path (found while verifying the new solution_design guardrail):

    1. The agent couldn't discover its own draft objects. list_objects / list_metadata read getMetaItems active-only, so a brand-new object the agent had just drafted (never published) was reported as "not found" when it then tried to author an approval flow against it. They now pass previewDrafts: true, overlaying pending drafts on the active list (older runtimes ignore the flag → stay active-only). describe_metadata was already draft-first.

    2. The auto-authored flow had no package to bind to. apply_blueprint already homes its artifacts in an app package, but its result only nested the id under package. It now also surfaces a top-level packageId and a bindingHint telling the agent to pass that packageId to create_metadata when it drafts follow-up automation (e.g. the approval flow) — so the flow lands in the app package instead of becoming an orphan draft.

    Together with the solution_design process guardrail, this makes the "model the data, then proactively draft the approval flow bound to the app" flow actually executable end-to-end.

  • 4705fb8: fix(ai): solution_design no longer models a process/approval as a table

    Asking the assistant to "design expense reimbursement" made it, by default, invent an approval_record TABLE to represent the approval process — a non-functional "process-as-data" anti-pattern. It only switched to a flow after the user pushed back.

    Hardens the default in two places:

    • propose_blueprint generation prompt (the "metadata architect" system message): status/lifecycle is modeled as a select field, never a table; it must NOT create approval / approval_record / approval_step / workflow / process objects (a process is a flow, its trail comes from platform history); the people a process references (approver/reviewer/owner) are lookup fields; and if the goal implies a process it adds an assumption that the approval flow is a separate step.

    • solution_design skill instructions: the same modeling rules, plus — when the goal involves a process — the agent now PROACTIVELY drafts the approval flow after apply_blueprint (call get_metadata_schema('flow'), then create_metadata(type:'flow', …) with the approval node(s), bound to the same app package) instead of waiting for the user to ask "now create the flow". Optionally adds a state_machine rule to block illegal status transitions.

    Regression-tested: solution-design-guardrail.test.ts asserts the skill instructions carry the no-process-as-table rule, status-as-select, and the proactive-flow step.

    • @objectstack/spec@7.9.0
    • @objectstack/core@7.9.0
    • @objectstack/types@7.9.0
    • @objectstack/formula@7.9.0

7.8.0

Minor Changes

  • 6b82e68: feat(ai): zero-package app building — auto-home a blueprint's app in a writable package

    When the AI blueprint flow builds an app, it now silently gives that app a writable "home" package (one app ⇒ one app.<name> package) and binds every drafted artifact (objects, views, dashboards, the app) to it — so a business user never has to create a "package" to start building (the mainstream AI-builder UX: Power Apps' default solution, Salesforce orgs). Packaging/versioning stays an opt-in, later concern.

    • apply_blueprint ensures the app package up front (idempotent: reuse if it exists, else create via the runtime package service) and threads its packageId through every stageDraft → sys_metadata.package_id. The result envelope gains package: { id, name, created }.
    • The package service is resolved lazily (per call, not at plugin-init time) so it works regardless of service-init order and picks up the opt-in marketplace capability when present.
    • Best-effort, non-fatal: if no package service is wired, drafting proceeds package-less exactly as before — the build never fails on packaging.

    Scope/caveats: this stamps the legacy sys_metadata.package_id (a real grouping + the foundation for later version/export/promote), not the sealed sys_package_version model — full cross-environment promotion and Studio package-selector visibility depend on finishing the runtime package subsystem (ADR-0027), tracked separately. (The showcase example enables the marketplace capability to exercise this.)

  • 4888ea2: feat(ai): add get_metadata_schema tool so the agent can read a type's contract before authoring

    The metadata-authoring agent never sees the real spec Zod schemas — it works against a simplified blueprint or sends a free-form definition and only learns the true shape from post-hoc validation errors. For complex types (view, dashboard, flow, …) that means guessing, e.g. a kanban view's required kanban: { groupByField, columns } block.

    New get_metadata_schema tool returns the JSON Schema (via Zod v4's toJSONSchema) derived from the SAME live schema saveMetaItem validates against (getMetadataTypeSchema). The metadata_authoring skill now instructs the agent to call it before authoring a non-trivial type, so it conforms first time instead of trial-and-error. Read-only; resolves plural type names; returns a graceful error for types that can't be serialized (e.g. object, which the dedicated create_object tools cover anyway).

  • 36719db: fix: AI-built apps are usable immediately — sync new object tables on publish + emit valid kanban config

    Two gaps found by end-to-end testing of an AI-built app:

    1. A freshly-published object couldn't accept records until a server restart. Publishing a drafted object registered it in the in-memory registry but never created its physical table (table sync only ran at boot), so inserts failed with object_not_found ("no such table"). Added ObjectQL.syncObjectSchema(name) (a targeted, idempotent single-object schema sync) and call it from the publish paths (protocol.publishMetaItem and saveMetaItem mode:'publish', via ensureObjectStorage). Best-effort + non-fatal. New objects are now CRUD-able the moment they're published.

    2. AI-generated kanban views rendered as plain lists (and sometimes failed validation). The blueprint viewBody emitted list.type:'kanban' with no kanban config; KanbanConfigSchema requires groupByField and columns. Added an optional groupBy to the blueprint view schema (lenient + strict) and have apply_blueprint set list.kanban = { groupByField, columns } — using the view's explicit groupBy when given, else inferring the object's first select field. AI-built kanban views now validate, publish, and carry a real group-by field.

Patch Changes

  • 06f2bbb: fix(ai): make ADR-0033 blueprint authoring work with OpenAI structured outputs

    Two bugs surfaced by a live end-to-end run (Studio chat → blueprint → draft → review → publish) against a real model (OpenAI via the Vercel AI Gateway) — both invisible to the existing unit tests:

    1. propose_blueprint failed against OpenAI strict structured outputs. SolutionBlueprintSchema uses optional fields and a free-form seedData record; OpenAI's strict mode requires every property listed in required and rejects open additionalProperties, so generateObject errored ('required' … must include every key in properties) and the agent silently fell back to free-text. Adds SolutionBlueprintStrictSchema — a strict-compatible mirror (optional → nullable, no z.record) used only as the generateObject output contract. The lenient SolutionBlueprintSchema (and every existing consumer/test) is unchanged; the blueprint tools strip the nulls the strict contract emits so downstream stays clean.

    2. Tool-only assistant turns failed to persist. ai_messages.content is required, but an assistant turn that only calls a tool has no text, so the insert failed, the turn was dropped, and the next turn lost context (the agent re-proposed instead of applying the confirmed blueprint). ObjectQLConversationService.addMessage now synthesizes a readable placeholder from the tool names ((called propose_blueprint)) plus a defensive non-empty fallback.

    With both fixes the full plan-first loop runs end-to-end on OpenAI models: propose → confirm → batch-draft objects/views/dashboards/app → review/diff → publish.

  • 4fbb86a: feat(packages): consolidate the package subsystem so AI-built app packages surface in Studio

    The package subsystem was split across two stores that never met: the in-memory SchemaRegistry (what the dispatcher's /api/v1/packages list/detail and getMetaItems({type:'package'}) read — i.e. Studio's package selector) and the durable sys_packages table (where the AI's auto app package, and any package-service publish, were written). Nothing reconciled the two, so an AI-created app.<name> package never appeared in Studio.

    This unifies them around one write primitive and one read source:

    • protocol.installPackage is now implemented (it was declared-but-missing). It is the single canonical write path: it registers the package in the in-memory registry and best-effort persists it to sys_packages via the package service. Non-fatal when no package service is wired (registry write still succeeds).
    • Dispatcher POST /api/v1/packages routes through protocol.installPackage (falling back to the bare registry write when the protocol is unavailable), so HTTP installs are durable too.
    • @objectstack/service-package reconciles sys_packages back into the registry on boot, without clobbering filesystem-registered packages — so persisted packages survive a restart and stay visible in the registry-backed read paths.
    • @objectstack/service-ai apply_blueprint now homes an app via protocol.installPackage (falling back to the legacy package-service publish), so the app package lands where Studio reads it.

    Still the legacy package_id plane — sealed sys_package_version versioning and cross-environment promotion remain ADR-0027 follow-ups.

  • 328a7c4: fix(ai): AI-authored views now bind to their object and render (kanban as a board, not a list)

    An AI-built app's views (including kanban) appeared only as the default list and never as selectable tabs. Diagnosis (vs the working showcase kanban) showed it was a metadata-shape bug in the blueprint's viewBody, not the renderer or skill: it emitted a bare { list: {…} } fragment instead of the canonical view record. Three things were missing/wrong:

    • no top-level name → getMetaItems only surfaces overlay rows whose body has name, so every AI view was silently dropped from the object's view list;
    • no top-level object / viewKind → the console couldn't bind the view to its object;
    • the view name wasn't <object>.<key>-prefixed (the convention the console keys view tabs off).

    viewBody now emits { name: '<object>.<key>', object, viewKind: 'list'|'form', config: <ListView|FormView> }, matching the shape the showcase's own views use (verified against the real ViewSchema). End-to-end verified: an AI-built kanban app surfaces 看板 + 列表 as tabs and renders the kanban as a board grouped by status.

  • Updated dependencies [06f2bbb]

  • Updated dependencies [f01f9fa]

  • Updated dependencies [36719db]

  • Updated dependencies [424ab26]

    • @objectstack/spec@7.8.0
    • @objectstack/formula@7.8.0
    • @objectstack/core@7.8.0
    • @objectstack/types@7.8.0

7.7.0

Minor Changes

  • b391955: feat(ai): blueprint app-building — propose/draft the navigation app, not just the data model

    The plan-first blueprint (ADR-0033 §4) now also designs the app (the navigation shell end users open in the App Launcher), so "build me a project-management application" yields an openable app — not just its objects, views, and dashboards.

    • SolutionBlueprintSchema (@objectstack/spec/ai) gains an optional app: { name, label?, icon?, nav? }, where each nav entry targets a created object or dashboard. nav may be omitted to auto-surface every object (then dashboard).
    • apply_blueprint expands the app into an AppSchema body (single-level navigation of object/dashboard items) and drafts it last — through the same draft-gated, per-type-validated stageDraft path as everything else. It never sets isDefault.
    • propose_blueprint now asks the agent to include the app and reports counts.app.

    Still draft-gated: nothing is live until the human publishes. Scope is basic app-building (one app, flat nav); areas/groups/mobile-nav remain author-it-later via update_metadata.

  • 984ddff: feat(service-ai): ADR-0033 Phase A — draft-gate AI metadata authoring

    AI metadata mutations no longer publish straight to the live schema. Every write now routes through the ADR-0027 draft workspace via protocol.saveMetaItem({ mode:'draft' }) — nothing an agent authors goes live until a human reviews the diff and publishes. The draft is the approval gate (the never-enforced requiresConfirmation flag is retired).

    Adds a type-agnostic apply surface — create_metadata / update_metadata / describe_metadata / list_metadata — that works for any metadata type (view, dashboard, flow, …), validated against each type's Zod schema with errors fed back to the agent for self-correction. The existing object/field tools become thin draft-writing wrappers. Tool results return { status:'drafted', type, name, summary, changedKeys }.

  • f06b64e: feat(ai): ADR-0033 Phase C — plan-first blueprint authoring

    For high-level goals ("build me a project-management system") the metadata assistant now designs before it builds. Adds a SolutionBlueprintSchema (@objectstack/spec/ai) describing proposed objects, fields, relationships, views, dashboards, and seed data with stated assumptions, plus two tools:

    • propose_blueprint(goal) — emits a structured blueprint via structured output. Nothing is persisted; the agent presents it for conversational confirmation and asks at most 1–2 structure-deciding questions.
    • apply_blueprint(blueprint) — only after the human approves, batch-drafts every artifact through the Phase A draft path (protocol.saveMetaItem({mode:'draft'})), validated per-type and partial-tolerant (a bad item is reported, the rest still draft). Seed data is reported as proposed, not auto-applied (no runtime dataset type).

    A new solution_design skill carries the plan-first instructions and is bound to metadata_assistant alongside metadata_authoring. The shared draft-write primitive is exported from the metadata tools as stageDraft and reused, keeping one draft-write path.

Patch Changes

  • Updated dependencies [b391955]
  • Updated dependencies [f06b64e]
  • Updated dependencies [825ab06]
  • Updated dependencies [023bf93]
    • @objectstack/spec@7.7.0
    • @objectstack/formula@7.7.0
    • @objectstack/core@7.7.0
    • @objectstack/types@7.7.0

7.6.0

Minor Changes

  • c4a4cbd: ADR-0032 (phase 1): validate-by-default expression layer — no silent failure.

    Kills the #1491 class where a malformed predicate (e.g. the {record.x} template-brace-in-CEL mistake) silently evaluated to false and made a flow "fire" with no effect:

    • service-automation: flow evaluateCondition no longer swallows CEL failures to false — it throws an attributed, corrective error; and registerFlow now parse-validates every predicate (start/decision/edge condition) at registration, failing loudly with the offending location + source + the fix.
    • formula: new shared validator — validateExpression(role, src, schema?), introspectScope, CEL_STDLIB_FUNCTIONS — with schema-aware field-existence
      • did-you-mean. The {{ }} template engine gains a formatter whitelist (currency/number/percent/date/datetime/truncate/upper/lower/ default/…) with defined value→string semantics; arbitrary logic in holes is rejected. Plain {{ path }} stays back-compatible.
    • cli: objectstack compile validates every flow / validation-rule / field-formula predicate against the resolved object schema and fails the build with located, corrective messages.
    • service-ai: new agent-callable validate_expression tool so authoring agents self-correct before committing.
    • spec: fix the FlowSchema JSDoc example that taught the bad condition: "{amount} < 500" single-brace form.

Patch Changes

  • 3377e38: fix(release): stop the fixed-group major cascade caused by internal @objectstack/* peerDependencies.

    These packages declared workspace peerDependencies on other framework packages in the changesets fixed group. Inside a fixed group, changesets rewrites those peer ranges on every release and treats a peer-range change as breaking → major, which cascaded to all 69 packages → 8.0.0 on any minor changeset. Required internal peers are now regular dependencies; optional ones move to devDependencies (kept for in-workspace tests, no longer a published peer edge). Releases now bump correctly (patch/minor) instead of a spurious major.

  • Updated dependencies [955d4c8]

  • Updated dependencies [c4a4cbd]

  • Updated dependencies [b046ec2]

  • Updated dependencies [2170ad9]

  • Updated dependencies [02d6359]

  • Updated dependencies [7648242]

  • Updated dependencies [8fa1e7f]

  • Updated dependencies [55866f5]

  • Updated dependencies [60f9c45]

    • @objectstack/spec@7.6.0
    • @objectstack/formula@7.6.0
    • @objectstack/core@7.6.0
    • @objectstack/types@7.6.0

7.5.0

Patch Changes

  • @objectstack/spec@7.5.0
  • @objectstack/core@7.5.0
  • @objectstack/types@7.5.0

7.4.1

Patch Changes

  • @objectstack/spec@7.4.1
  • @objectstack/core@7.4.1
  • @objectstack/types@7.4.1

7.4.0

Minor Changes

  • 2faf9f2: External Datasource Federation (ADR-0015) — Phase 4: AI awareness.

    SchemaRetriever.renderSnippet now annotates federated objects in the auto-injected schema context, e.g. ### wh_order — Warehouse Order [external, read-only, datasource=warehouse], so the LLM knows an object comes from a customer's production database and must not propose schema changes or unsafe writes. ObjectShape gains datasource

    • external (read from object metadata). Managed objects are unannotated.

Patch Changes

  • Updated dependencies [23c7107]
  • Updated dependencies [c72daad]
  • Updated dependencies [f115182]
  • Updated dependencies [2faf9f2]
  • Updated dependencies [2faf9f2]
  • Updated dependencies [2faf9f2]
  • Updated dependencies [58b450b]
  • Updated dependencies [82eb6cf]
  • Updated dependencies [13d8653]
  • Updated dependencies [ff3d006]
  • Updated dependencies [5e831de]
    • @objectstack/spec@7.4.0
    • @objectstack/core@7.4.0
    • @objectstack/types@7.4.0

7.3.0

Patch Changes

  • Updated dependencies [5e7c554]
    • @objectstack/spec@7.3.0
    • @objectstack/core@7.3.0
    • @objectstack/types@7.3.0

7.2.1

Patch Changes

  • 9096dfe: OS_ env-var prefix migration (issue #1382).

    All ObjectStack-owned environment variables now use the OS_ prefix. Legacy names still work for one release and emit a one-shot deprecation warning via the new readEnvWithDeprecation() helper in @objectstack/types.

    Renamed (with legacy fallback):

    New Legacy (deprecated)
    OS_AUTH_SECRET AUTH_SECRET, BETTER_AUTH_SECRET
    OS_AUTH_URL AUTH_BASE_URL, BETTER_AUTH_URL, OS_AUTH_BASE_URL
    OS_PORT PORT
    OS_DATABASE_URL DATABASE_URL
    OS_ROOT_DOMAIN ROOT_DOMAIN
    OS_MULTI_ORG_ENABLED OS_MULTI_TENANT
    OS_CORS_ENABLED CORS_ENABLED
    OS_CORS_ORIGIN CORS_ORIGIN
    OS_CORS_CREDENTIALS CORS_CREDENTIALS
    OS_CORS_MAX_AGE CORS_MAX_AGE
    OS_AI_MODEL AI_MODEL
    OS_MCP_SERVER_ENABLED MCP_SERVER_ENABLED
    OS_MCP_SERVER_NAME MCP_SERVER_NAME
    OS_MCP_SERVER_TRANSPORT MCP_SERVER_TRANSPORT
    OS_NODE_ID OBJECTSTACK_NODE_ID
    OS_METADATA_WRITABLE OBJECTSTACK_METADATA_WRITABLE
    OS_DEV_CRYPTO_KEY OBJECTSTACK_DEV_CRYPTO_KEY
    OS_HOME OBJECTSTACK_HOME

    Migration: rename in your .env. Legacy names continue to work this release and will be removed in a future major. Industry-standard names (NODE_ENV, HOME, OPENAI_API_KEY, TURSO_*, OAuth *_CLIENT_ID/SECRET, RESEND_API_KEY, POSTMARK_TOKEN, AI_GATEWAY_*, SMTP_*) are NOT renamed.

  • Updated dependencies [9096dfe]

    • @objectstack/types@7.2.1
    • @objectstack/spec@7.2.1
    • @objectstack/core@7.2.1

7.2.0

Patch Changes

  • @objectstack/spec@7.2.0
  • @objectstack/core@7.2.0

7.1.0

Patch Changes

  • Updated dependencies [47a92f4]
    • @objectstack/spec@7.1.0
    • @objectstack/core@7.1.0

7.0.0

Patch Changes

  • Updated dependencies [74470ad]
  • Updated dependencies [d29617e]
  • Updated dependencies [dc72172]
    • @objectstack/spec@7.0.0
    • @objectstack/core@7.0.0
    • @objectstack/embedder-openai@7.0.0

6.9.0

Minor Changes

  • e9bacda: Auto-generate concise titles for AI conversations.

    AIService now exposes summarizeConversation(id) and fires it once per conversation after the first assistant turn lands. The generated title (≤ 16 chars by default) is PATCHed onto the ai_conversations row so the sidebar shows a meaningful label instead of "New conversation". Failures are silently swallowed — title generation is purely cosmetic and never blocks chat.

    Plumbing:

    • New AI settings (in the ai Settings namespace):
      • title_generation_enabled (toggle, default on for non-memory providers)
      • title_max_length (number, 8–80, default 16)
    • AIService.setTitleGenerationConfig({ enabled, maxLength }) — called by AIServicePlugin.bindSettings() whenever the ai namespace changes, so admins can toggle the feature live from Setup without a restart.
    • AIService calls summarizeConversation() fire-and-forget at the natural end of chatWithTools and streamChatWithTools. Idempotent per service instance — a single titling attempt per conversation per process.

    Defaults are conservative: memory provider stays untouched (no LLM call is made), and any per-test AIService that doesn't explicitly call setTitleGenerationConfig({ enabled: true }) behaves exactly as before.

Patch Changes

  • @objectstack/spec@6.9.0
  • @objectstack/core@6.9.0

6.8.1

Patch Changes

  • @objectstack/spec@6.8.1
  • @objectstack/core@6.8.1

6.8.0

Minor Changes

  • 6e88f77: Auto-persist chat history when a conversationId is supplied.

    • AIService.chatWithTools and streamChatWithTools now write the inbound user turn, each intermediate assistant/tool round, and the final assistant turn to ai_messages whenever toolExecutionContext.conversationId is set. Persistence is best-effort: failures are warned and never break the chat response.
    • Add IAIConversationService.update(conversationId, { title?, metadata? }) and a matching PATCH /api/v1/ai/conversations/:id route so clients can rename conversations and edit metadata.
    • ObjectQLConversationService and InMemoryConversationService both implement the new update method.

Patch Changes

  • 50ccd9c: Fix peer-dependency version range from workspace:* to workspace:^ to avoid forced major bumps in fixed-group releases. workspace:* expands to an exact version on publish; any minor bump of the peer then falls out of range and triggers a semver-major bump on the dependent. workspace:^ expands to ^x.y.z which correctly accepts minor bumps.

    Affects:

    • service-ai peer on @objectstack/embedder-openai
    • runtime peer on @objectstack/driver-turso

  • Updated dependencies [6e88f77]

  • Updated dependencies [c8b9f57]

    • @objectstack/spec@6.8.0
    • @objectstack/core@6.8.0

6.7.1

Patch Changes

  • @objectstack/spec@6.7.1
  • @objectstack/core@6.7.1

6.7.0

Minor Changes

  • 4f9e9d4: Settings → runtime bridge: embedder_* settings now build a real IEmbedder and register it as a kernel-level DI service.

    @objectstack/spec

    • Exports EMBEDDER_SERVICE = 'embedder' from contracts/embedder.ts as the canonical DI token for the kernel-registered embedder.

    @objectstack/service-ai

    • Adds @objectstack/embedder-openai as an optional peer dependency (matches the @ai-sdk/* provider plugins pattern).
    • AIServicePlugin.bindSettings() now also:
      • Reads embedder_provider / embedder_api_key / embedder_model / embedder_base_url / embedder_dimensions from the ai namespace.
      • Dynamically imports @objectstack/embedder-openai and constructs an OpenAIEmbedder via createOpenAIEmbedder({ preset, … }).
      • Registers / replaces the instance under EMBEDDER_SERVICE. When the operator sets embedder_provider = none, the service is left unset so adapters can fail fast with a clear message.
      • Subscribes to settings:changed for the ai namespace so embedder swaps go live without restart (mirrors the chat-adapter pattern).
      • Overrides the manifest's fallback ai/test_embedder action with a live one-shot embed(['ping']) round-trip against the form's (possibly unsaved) values. Reports vector dims + latency.

    @objectstack/knowledge-turso

    • KnowledgeTursoPlugin's embedding constructor option is now optional. When omitted, the plugin resolves EMBEDDER_SERVICE from the kernel at start() time — typically the embedder built by @objectstack/service-ai from the ai settings namespace.
    • Explicit embedding still wins when both are present (useful for tests and multi-embedder setups).
    • Logs (embedder=<id>, dims=<n>) on adapter registration so operators can confirm wiring at a glance.
    • When neither path resolves, the plugin warns with a one-line hint pointing to Settings → AI & Embedder and no-ops gracefully (the host kernel still boots).

    Tests

    • service-ai: +5 cases (now 85) covering ai/test_embedder action registration, provider=none warning, missing-api-key error, custom-provider-without-base-URL error, and the full happy path (mocked fetch → embedder registered under EMBEDDER_SERVICE → test_embedder action returns vector dims).
    • knowledge-turso: new plugin.test.ts (+5 cases) covering deferred construction, EMBEDDER_SERVICE fallback, explicit-wins precedence, missing-both warn-and-noop, and missing-knowledge-service warn.

    End-to-end now possible: operator opens Settings → AI & Embedder, picks 硅基流动 + paste API key + chooses BAAI/bge-m3, hits Save. Within the same process, EMBEDDER_SERVICE is registered/replaced, KnowledgeTursoPlugin (if started without an explicit embedder) picks it up, and subsequent knowledge.search() calls embed via the new provider — no restart, no env vars.

Patch Changes

  • Updated dependencies [430067b]
  • Updated dependencies [4f9e9d4]
    • @objectstack/spec@6.7.0
    • @objectstack/core@6.7.0

6.6.0

Patch Changes

  • Updated dependencies [a49cfc2]
    • @objectstack/spec@6.6.0
    • @objectstack/core@6.6.0

6.5.1

Patch Changes

  • @objectstack/spec@6.5.1
  • @objectstack/core@6.5.1

6.5.0

Patch Changes

  • @objectstack/spec@6.5.0
  • @objectstack/core@6.5.0

6.4.0

Minor Changes

  • f8651cc: Knowledge Protocol MVP — protocol-first RAG via adapter plugins.

    What's new:

    • @objectstack/spec — new KnowledgeSource / KnowledgeDocument / KnowledgeChunk / KnowledgeHit schemas (under @objectstack/spec/ai) and IKnowledgeService / IKnowledgeAdapter contracts (under @objectstack/spec/contracts).
    • @objectstack/service-knowledge — KnowledgeService orchestrator + KnowledgeServicePlugin. Routes search/index calls to the appropriate adapter, runs permission-aware retrieval by re-checking every hit's sourceRecordId against the caller's ExecutionContext via IDataEngine (same RLS that gates plain ObjectQL), and subscribes to IRealtimeService for inline record→adapter sync.
    • @objectstack/knowledge-memory — deterministic, dependency-free in-memory adapter for dev/tests/reference. Hash-token embedder + brute-force cosine + paragraph chunking.
    • @objectstack/knowledge-ragflow — production-grade adapter against the Apache-2.0 RAGFlow REST API. Plug in your dataset id; ObjectStack handles permission filtering after retrieval.
    • @objectstack/service-ai — new search_knowledge tool wired through the registry. Threads the LLM caller's actor into KnowledgeService.search so retrieval honours RLS automatically.

    Why this design: ObjectStack does NOT own chunking / embedding / vector storage / rerank — those are commodity capabilities best handled by mature OSS (RAGFlow, LlamaIndex, Dify, …). What ObjectStack uniquely owns is the protocol + permission-aware orchestration on top.

    See content/docs/protocol/knowledge.mdx for the full design.

  • f8651cc: AI tools now execute with the end-user's ExecutionContext, so the existing ObjectQL row-level-security rules automatically scope what an agent can read and mutate.

    What changed

    • New ToolExecutionContext (on @objectstack/spec/contracts's ChatWithToolsOptions) carries the authenticated actor, conversation id, and environment id through to tool handlers.
    • The built-in data tools (query_records, get_record, aggregate_data, legacy query_data) and the auto-generated action_* tools now pass options.context to IDataEngine calls, mapping the actor to { userId, roles, permissions, isSystem: false }.
    • Assistant + agent REST routes forward req.user into the new context automatically — no caller changes required.
    • When no actor is provided (cron jobs, internal callers, existing tests) the helpers fall back to { isSystem: true }, preserving today's behaviour. Fully backward compatible.

    Why this matters

    Before this change, an AI tool call ran with system privileges and saw every row in the tenant. Now the agent sees exactly what the human operator would see — same RLS, same field-level masking, same audit trail. This is the foundation for trustworthy autonomous agents.

    For custom call sites

    If you invoke aiService.chatWithTools(...) from your own route, pass toolExecutionContext: { actor: { id, roles, permissions } } to inherit the user's permissions. Omit it to keep the legacy system-level behaviour.

Patch Changes

  • a981d57: Auto-persist chat messages when conversationId is supplied. AIService.chatWithTools and streamChatWithTools now write the inbound user turn, every intermediate assistant/tool round, and the final assistant turn to ai_messages via the configured conversation service. Persistence is best-effort: failures are logged at warn level and never fail the chat request.
  • b486666: Add GET /api/v1/ai/conversations/:id route to fetch a single conversation with its full message history. Enforces ownership via the authenticated user: returns 404 when the conversation does not exist and 403 when it belongs to another user. Enables clients to hydrate chat UIs from server-persisted history instead of relying on local storage.
  • Updated dependencies [f8651cc]
  • Updated dependencies [f8651cc]
  • Updated dependencies [0bf6f9a]
    • @objectstack/spec@6.4.0
    • @objectstack/core@6.4.0

6.3.0

Patch Changes

  • @objectstack/spec@6.3.0
  • @objectstack/core@6.3.0

6.2.0

Minor Changes

  • b4c74a9: Actions-as-tools Phase 3 — Human-In-The-Loop approval queue.

    Dangerous declarative actions (confirmText, mode:'delete', variant:'danger') can now be exposed to the LLM safely. Instead of being skipped outright, they are registered as tools whose handler enqueues a pending request and returns { status: 'pending_approval', pendingActionId } to the model. A human approves (or rejects) from Studio's pending-actions inbox; the service then re-runs the exact same dispatcher.

    New surface

    • New system object ai_pending_actions (id, conversation_id?, message_id?, object_name, action_name, tool_name, tool_input, status [pending|approved|executed|failed|rejected], result?, error?, rejection_reason?, proposed_by, decided_by?, proposed_at, decided_at?).
    • New built-in Studio view AiPendingActionView with pending / executed / rejected / failed sub-views and per-row Approve / Reject API actions.
    • New methods on IAIService (all optional, gated on a wired IDataEngine):
      • proposePendingAction(input) → { id }
      • approvePendingAction(id, actorId) → { status, result?, error? }
      • rejectPendingAction(id, actorId, reason?)
      • listPendingActions(filter?) → PendingActionRow[]
    • New exported types: PendingActionStatus, ProposePendingActionInput, PendingActionRow.
    • New REST routes (auth required):
      • GET /api/v1/ai/pending-actions (ai:read)
      • GET /api/v1/ai/pending-actions/:id (ai:read)
      • POST /api/v1/ai/pending-actions/:id/approve (ai:approve)
      • POST /api/v1/ai/pending-actions/:id/reject (ai:approve)
    • New exported predicate actionRequiresApproval(action) for Studio's exposure surface.

    Wiring

    AIServicePluginOptions gains enableActionApproval?: boolean (default false). When true and an IDataEngine is available, dangerous actions are registered and routed through the queue.

    kernel.use(
  new AIServicePlugin({
    enableActionApproval: true, // opt in
    apiActionBaseUrl: "http://localhost:3000",
  })
);

    Internals

    • actionSkipReason() accepts enableActionApproval + aiService in its ctx and stops returning "requires confirmation" / "mode='delete'" / "variant='danger'" when HITL is wired.
    • registerActionsAsTools() pre-registers a bypass-approval dispatcher per dangerous tool via aiService.registerPendingActionDispatcher(toolName, fn); approval calls back into the same code path with enableActionApproval flipped off, so a single handler implementation serves both proposal and execution.
    • createActionToolHandler() short-circuits to proposePendingAction() when enableActionApproval && actionRequiresApproval(action) && ctx.aiService?.proposePendingAction.

    Out of scope (deferred)

    Slack/email notifications, approver routing (any signed-in user can approve in v1), auto-expiry of pending requests, resuming the same LLM turn after approval (operators get a fresh assistant message instead).

  • bce47a0: Polish Studio HITL pending-action inbox UI

    The AiPendingActionView shipped by service-ai is now an actual operator console rather than a flat grid:

    • Drawer detail panel — clicking any row opens a side drawer (navigation: { mode: 'drawer', view: 'detail' }) with four sections: Proposal · Tool input · Conversation context · Decision.
    • JSON widget on tool_input, result, and error so structured tool arguments and responses are readable without copy-pasting into a formatter.
    • Relative timestamps (type: 'datetime-relative') on proposed_at / decided_at columns and form fields.
    • Conversation/message linkbacks — the existing Field.lookup references to ai_conversations / ai_messages are surfaced in a collapsed "Conversation context" section, giving operators one-click access from a pending action back to the chat that proposed it.
    • Status-conditional fields via visibleOn predicates — rejection_reason only appears for rejected rows, error only for failed rows, etc.
    • Per-row approve/reject buttons on the Pending tab via rowActions pointing at the existing approve_pending_action / reject_pending_action object actions; the same actions also render in the drawer header.
    • Status-coloured rows to make pending vs failed vs executed scannable.

    Snapshot-style tests in __tests__/ai-pending-action.view.test.ts lock the shape so future Studio contract changes (widget renames, navigation modes) fail loudly in one place.

    This is a metadata-only change — Studio (@object-ui/studio) interprets the new view automatically. No backend, REST, or HITL semantics changed; the end-to-end demos in examples/app-todo/test/ai-hitl*.test.ts continue to pass unmodified.

Patch Changes

  • 13a4f38: Actions-as-tools Phase 2: the AI tool runtime can now dispatch type:'api' and type:'flow' actions in addition to type:'script'.

    • New exported ApiActionClient interface and createFetchApiClient({ baseUrl, headers, fetch }) factory — default fetch-based dispatch resolves relative target paths against baseUrl, throws on non-2xx with ${method} ${url} → ${status}: ${body}, and JSON-parses the response.
    • New exported buildApiRequestBody(action, args, record, recordId) helper — honours bodyShape.wrap, recordIdParam + recordIdField (defaults to 'id'), and merges bodyExtra last so constants win.
    • ActionToolsContext extended (additive): automation, apiClient, apiBaseUrl, apiHeaders.
    • actionSkipReason() gains an optional second ctx parameter that returns precise wiring-availability reasons ('no automation service available', 'no apiClient or apiBaseUrl configured'). Studio-only types (url / modal / form) and all dangerous variants (confirmText, mode:'delete', variant:'danger') remain skipped.
    • AIServicePlugin options accept apiActionBaseUrl (falls back to OS_AI_ACTION_API_BASE_URL) and apiActionHeaders; the plugin now resolves the automation service silently and threads everything into registerActionsAsTools.

    Net result: every non-destructive declarative action with a target — script, api, flow — is now LLM-callable end-to-end as soon as the corresponding wiring is in place.

  • bce47a0: HITL Phase 3 — end-to-end demos + bug fix in handler-engine adapter.

    Two runnable integration demos for the action-approval queue ship under examples/app-todo/test/:

    • ai-hitl.test.ts — drives the tool registry directly (no LLM). Asserts variant:'danger' actions register as tools, invocation returns pending_approval, row persists, approvePendingAction(id, actor) re-runs the handler, row flips to executed. Reject path covered too. Run with pnpm --filter @example/app-todo test:hitl.
    • ai-hitl-llm.test.ts — same scenario behind a real model on Vercel AI Gateway. The LLM autonomously picks action_delete_completed, the framework gates the call with pending_approval, the model summarises the wait without retrying, and the operator-side approve completes the deletion. Gated on AI_GATEWAY_API_KEY. Run with AI_GATEWAY_API_KEY=... pnpm --filter @example/app-todo test:hitl:llm.

    While wiring the demos, two bugs surfaced in the bypass-approval dispatcher and the handler-engine adapter:

    1. Bulk delete from declarative handlers was silently failing. The adapter built by buildHandlerEngineAdapter() wrapped multi-id deletes as engine.delete(obj, { where: { id: { $in: ids } } }), but ObjectQLEngine.delete() prefers the scalar id branch whenever where.id is set — so the { $in: [...] } object was forwarded to driver.delete(scalar) and rejected as "Wrong API use: tried to bind a value of an unknown type ([object Object])". The adapter now loops scalar deletes, which is correct and driver-agnostic.

    2. Approval pathway swallowed handler errors. createActionToolHandler returns a { ok: false, error } envelope on failure rather than throwing. The pre-registered bypass dispatcher just JSON-parsed and returned that envelope, so approvePendingAction thought the run succeeded and flipped the row to executed. The dispatcher now treats ok === false as a thrown error, so failed approvals are correctly persisted as status: 'failed' with the original message.

    Also: added delete/remove/purge/destroy/erase to MemoryLLMAdapter.ACTION_VERBS so the in-memory adapter can route delete-style intents during tests that don't have a real LLM.

    Docs: content/docs/guides/ai-capabilities.mdx now points at the two integration demos with copy-pasteable run commands.

  • 449e35d: Real-LLM smoke test for the data_chat agent loop, plus two query_data robustness fixes shaken out by running it against openai/gpt-4.1-mini via the Vercel AI Gateway.

    query_data tool fixes

    • Removed the LLM-controllable model parameter from the public tool schema. Frontier models were hallucinating text-davinci-003 and other long-dead model ids, breaking every plan generation.
    • Switched the structured-output filter shape from z.record(...) (which emits propertyNames in JSON Schema, rejected by OpenAI Structured Outputs) to a whereJson string field. The model emits a JSON-encoded ObjectQL filter; the tool parses & validates it before execution. This also fixes a parallel issue with OpenAI's strict mode requiring every property to appear in required.
    • Switched all optional fields to .nullable() so the planner Zod schema satisfies OpenAI Structured Outputs' "every property must be required" rule.
    • Beefed up the planner system prompt with explicit operator hints — most importantly: use $contains for partial string matches ("task named Foo" → {"subject":{"$contains":"Foo"}}), not equality. Without this hint the model defaulted to exact-match equality and never found anything.

    New smoke test

    examples/app-todo/test/ai-llm.test.ts (gated on AI_GATEWAY_API_KEY): boots the full ObjectStack, registers query_data + the six auto-generated action_* tools, sends "Please mark the 'Build' task as complete." to a real LLM, and asserts that

    1. the model picked the right tools in the right order (query_data → action_complete_task),
    2. a task row actually flipped to completed, and
    3. an ai_traces chat_with_tools row landed.

    Run with: pnpm --filter @example/app-todo test:llm.

    Verified end-to-end against openai/gpt-4.1-mini (~6.6 s, 2 tool calls, 1 task completed, trace persisted).

  • Updated dependencies [b4c74a9]

    • @objectstack/spec@6.2.0
    • @objectstack/core@6.2.0

6.1.1

Patch Changes

  • @objectstack/spec@6.1.1
  • @objectstack/core@6.1.1

6.1.0

Minor Changes

  • 93c0589: AI v1: Actions-as-Tools — every declarative UI Action of type: 'script' is now auto-exposed as an AI-callable tool named action_<name>. Agents can perform business operations ("complete the groceries task") via natural language, routed through the same dataEngine.executeAction() dispatcher Studio uses. This is the write-side counterpart to query_data.

    Highlights

    • registerActionsAsTools(toolRegistry, { metadata, dataEngine }) walks every object's actions[] and registers script-type ones, auto-injecting a recordId argument for row-context actions and inheriting JSON-Schema parameter types from the owning object's fields.
    • Safety filters skip destructive actions by default: confirmText, mode: 'delete', variant: 'danger', or explicit aiExposed: false.
    • New aiExposed?: boolean flag on ActionSchema for fine-grained opt-out.
    • New actions_executor skill bundle subscribes to action_* (wildcard tool names now supported in SkillSchema.tools).
    • The built-in data_chat agent now references both data_explorer and actions_executor skills, so users get read + write capabilities out of the box.
    • MemoryLLMAdapter learned a small two-step heuristic — when it sees an action verb ("complete", "start", "clone", ...) it routes to the matching action_* tool, resolving recordId from any prior query_data result.
    • New examples/app-todo/test/ai-action.test.ts demo proves the loop: user says "please complete the groceries task" → agent finds the task → agent calls action_complete_task → task status flips → ai_traces records the run.

    Breaking changes

    None. aiExposed is additive; existing actions remain exposed unless they fail an existing safety filter.

    Phase-1 limitations (Phase-2 roadmap items)

    • Only type: 'script' actions; api/flow/url/modal/form skipped.
    • No human-in-the-loop approval flow for destructive actions yet.
    • No CEL evaluation of visible/disabled predicates against agent context.
    • No bulk action support (single-record only).

Patch Changes

  • Updated dependencies [93c0589]
    • @objectstack/spec@6.1.0
    • @objectstack/core@6.1.0

6.0.0

Minor Changes

  • dbc4f7d: feat(ai): v1 AI capabilities — ModelRegistry, structured output, tracing, schema retrieval, and query_data tool

    This release lights up the first concrete capabilities on the slimmed AI protocol. All additions are non-breaking — new contract methods are optional and existing callers keep working unchanged.

    What's new

    • ModelRegistry (@objectstack/service-ai): in-memory runtime registry for AI.ModelConfig. Wire models via AIServicePluginOptions.models / defaultModelId. Exposes get, getOrThrow, getDefault, list, and estimateCost(modelId, usage) for ex-post token cost computation.

    • ai_traces object + auto-tracing: every LLM call from AIService (chat, complete, stream_chat, chat_with_tools, generate_object, embed) is now instrumented with latency, token usage, status, and (when pricing is registered) cost. The default ObjectQLTraceRecorder is auto-wired when the runtime exposes an IDataEngine, persisting rows to the new ai_traces object. Drop in a custom TraceRecorder via AIServicePluginOptions.traceRecorder, or pass null to opt out.

    • Structured output (IAIService.generateObject): new optional method on IAIService and LLMAdapter that returns a parsed, schema-validated object instead of free-form text. Implemented end-to-end in VercelLLMAdapter (uses the AI SDK's generateObject — provider strict-mode is automatic when supported). MemoryLLMAdapter ships a deterministic heuristic implementation so tests and demos work without an API key.

    • SchemaRetriever: lightweight keyword-based retriever over IMetadataService.listObjects(). Scores by object name (×3), label/plural (×2), description (×1), field name (×2), and field label (×1) with English stop-word filtering. Tokenisation splits snake_case so todo_task in a query matches name: 'todo_task'. SchemaRetriever.renderSnippet() produces a Markdown block ready to inject into a system prompt — no embeddings, no extra infra.

    • query_data tool: auto-registered when AI + Metadata + Data engine are all present. Takes a natural-language request, retrieves relevant schemas, asks the model for a structured QueryPlan via generateObject, validates the plan targets a real object, and executes it through IDataEngine.find. Returns { plan, count, records }. The composed primitive that closes the loop from "ask in English" → "validated SQL-shaped result".

    • Working demo in examples/app-todo: pnpm --filter @example/app-todo test:ai boots the full Todo stack, invokes query_data against the seeded tasks, and verifies the call lands in ai_traces. Zero API keys, ~3 seconds end-to-end. Serves as the canonical reference for wiring AI into a real app.

    Hardening

    • Strict tool schemas: nested orderBy and aggregations items in data-tools now declare additionalProperties: false + required, matching the top-level contract and making them safe for provider strict mode.

    Breaking-ish

    • TraceOperation values are now snake_case (stream_chat, chat_with_tools, generate_object) to match the project's data-value convention and so the ai_traces.operation select validates. Custom TraceRecorder implementations that hard-code the old camelCase names need to be updated. The values are an internal observability artefact — no public protocol surface exposes them.

    Notes

    • zod is now a direct dependency of @objectstack/service-ai (previously transitive via ai) because contract signatures and the new tool definition use z.ZodType types directly.
    • All new methods on IAIService / LLMAdapter are optional — existing custom adapters and callers continue to work without changes.
    • 12 new unit tests cover ModelRegistry (cost math, defaults, throwing lookups) and SchemaRetriever (scoring, snake_case tokenisation, limits, snippet rendering). Full suite: 323/323 ✓.

Patch Changes

  • Updated dependencies [629a716]
  • Updated dependencies [dbc4f7d]
  • Updated dependencies [944f187]
    • @objectstack/spec@6.0.0
    • @objectstack/core@6.0.0

5.2.0

Patch Changes

  • Updated dependencies [bab2b20]
  • Updated dependencies [fa011d8]
  • Updated dependencies [b806f58]
    • @objectstack/spec@5.2.0
    • @objectstack/core@5.2.0

5.1.0

Patch Changes

  • Updated dependencies [75f4ee6]
  • Updated dependencies [823d559]
    • @objectstack/spec@5.1.0
    • @objectstack/core@5.1.0

5.0.0

Patch Changes

  • Updated dependencies [2f9073a]
    • @objectstack/spec@5.0.0
    • @objectstack/core@5.0.0

4.2.0

Patch Changes

  • Updated dependencies [2869891]
    • @objectstack/spec@4.2.0
    • @objectstack/core@4.2.0

4.1.1

Patch Changes

  • @objectstack/spec@4.1.1
  • @objectstack/core@4.1.1

4.1.0

Patch Changes

  • Updated dependencies [2108c30]
  • Updated dependencies [23db640]
    • @objectstack/spec@4.1.0
    • @objectstack/core@4.1.0

4.0.5

Patch Changes

  • 15e0df6: chore: unify all package versions to a single patch release
  • Updated dependencies [15e0df6]
    • @objectstack/spec@4.0.5
    • @objectstack/core@4.0.5

4.0.4

Patch Changes

  • Updated dependencies [326b66b]
    • @objectstack/spec@4.0.4
    • @objectstack/core@4.0.4

4.0.3

Patch Changes

  • ee39bff: fix ai.
    • @objectstack/spec@4.0.3
    • @objectstack/core@4.0.3

4.0.2

Patch Changes

  • 5f659e9: fix ai
  • Updated dependencies [5f659e9]
    • @objectstack/spec@4.0.2
    • @objectstack/core@4.0.2

4.1.0

Minor Changes

  • Route auth/permissions metadata: Every route definition (RouteDefinition) now declares auth and permissions fields, enabling HTTP server adapters to enforce authentication and authorization automatically.
  • User context on RouteRequest: RouteRequest now carries an optional user: RouteUserContext object populated by the auth middleware, providing userId, displayName, roles, and permissions.
  • Conversation ownership enforcement: Conversation routes (create, list, add message, delete) are scoped to the authenticated user when a user context is present and the conversation has a userId. For backward compatibility, requests without user context and conversations created without a userId remain accessible under the existing behavior.
  • Enhanced tool-call loop error handling: chatWithTools now tracks tool execution errors across iterations and supports an onToolError callback ('continue' | 'abort') for fine-grained error control.
  • streamChatWithTools: New streaming tool-call loop that yields SSE events while automatically resolving intermediate tool calls.
  • New RouteUserContext type: Exported from the package for use by HTTP adapters and middleware.

4.0.0

Major Changes

  • ad4e04b: service ai

Patch Changes

  • Updated dependencies [f08ffc3]
  • Updated dependencies [e0b0a78]
    • @objectstack/spec@4.0.0
    • @objectstack/core@4.0.0

3.3.1

Patch Changes

  • Initial release of AI Service plugin
    • LLM adapter layer with provider abstraction (memory adapter included)
    • Conversation management service with in-memory persistence
    • Tool registry for metadata/business tool registration
    • REST/SSE route self-registration (/api/v1/ai/*)
    • Kernel plugin registering as 'ai' service conforming to IAIService contract