Skip to content

v0.2.0

Latest
Latest

Choose a tag to compare

View all tags
@WardenAdmin WardenAdmin released this 06 Sep 21:05
· 3 commits to main since this release
v0.2.0
72bdc79
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

[0.2.0] - 2026-09-06

Breaking

  • Saga authoring no longer inlines reason/commit capability blobs. Deploy kind: step manifests first, then compose with use: / version / with / when (workers → steps → sagas). Migration 011_step_definitions.sql adds step_definitions and pins step_definition_name / step_definition_version on step instances.
  • Saga definitions store the authoring AST only. Deploy link-checks catalog refs (ports, tighten, artifacts) but does not persist an expanded reason/commit photocopy. At saga start (and child spawn), the engine hydrates use: refs into instance frozen_steps. Redeploy sagas after upgrading — expanded definition bodies are not accepted.
  • Step and worker (namespace, name, version) pins are append-only. Redeploying the same version fails with an immutable-version error; bump the version to change capability.
  • Worker tool_sources no longer support legacy transport: sse. Use streamable_http (the default when transport is omitted) with a Streamable HTTP endpoint URL (commonly /mcp). Deploy rejects sse with a clear validation error.
  • Compensation YAML must declare exactly one tool in tools.allow (same as commit). Reason and commit steps share that single-tool undo path. Worker-manifest / DB compensation_prompt is removed (migration 008_drop_compensation_prompt.sql).
  • Removed WARDEN_REACT_TOOL_MESSAGE_LIMIT and all client-side MCP tool-output clipping. Tool payloads flow unchanged to the LLM transcript and tool_results / facts extraction. Tier-1 memory redaction (coupled to the clip limit) is removed; golden-ratio compression still digest/drops historical turns.
  • Worker MCP client upgraded to MCP Python SDK 2.x (mcp>=2.0,<3 on the worker extra). Rebuild worker images / re-run uv sync --extra worker after pulling. Client code uses v2 pagination (PaginatedRequestParams), streamable HTTP 2-tuple transports, and SDK snake_case protocol fields; the stdio mock fixture uses the v2 lowlevel Server handler API.
  • Soft-disable catalog definitions moved from PATCH /v1/definitions/{kind}/{uuid} to PATCH /v1/definitions/{kind}?id= or ?namespace=&name=&version= (CLI flags match).

Added

  • First-class step manifests (kind: step): catalog capabilities with input ports, step_kind reason|commit, and capability fields; sagas compose via use: + version + with + when (tighten-only overrides). Saga deploy link-checks refs; saga start freezes hydrated steps onto the instance. Input port schema fragments are enforced at step deploy (valid Draft-7), saga deploy (value: literals), and schedule (resolved bindings). List with warden list definitions --type step / GET /v1/definitions/steps. Docs: Step manifests.

  • Worker definitions store the full validated manifest in body JSONB (same pattern as sagas/steps). Migration 012_worker_definition_body.sql backfills and drops denormalized model_provider / model_name / system_prompt / tool_sources / adapter columns. GET .../workers/{id}?include_body=true returns row.body. Worker temperature is persisted and applied at runtime.

  • Catalog definition list/get items are uniform: id, namespace, name, version, is_active, timestamps (+ optional body). Workers gained is_active (013_worker_is_active.sql); top-level list adapter was removed (still in body). Soft-disable via PATCH /v1/definitions/{workers|steps|sagas} with query identity (id or triple) or warden definitions set-active. Inactive workers/steps/sagas are rejected at deploy link-check and runtime load with structured HTTP codes.

  • Catalog vocabulary: executable saga graph type is HydratedSagaBlueprint (authoring remains SagaAuthoringBlueprint). Worker timing bucket hydration_ms renamed to worker_init_ms. Reserve hydrate for catalog→runtime graph resolution; Jinja uses render; worker command load uses prepare/worker_init_ms.

  • Registry deploy stores definition bodies via one shared _definition_body_payload helper. Removed unused get_saga_definition_by_id, unreachable worker version-mismatch branch, and dead CLI list params; tests share worker_definition_body / step_definition_body factories.

  • MCP and tools → Tool payload hygiene — pagination, field projection, and summary-tool patterns for bounded MCP returns.

  • Reason-step tools.bind (⊆ with): pin saga-resolved values onto ReAct MCP tool args (saga wins, ∩ tool inputSchema), strip those keys from the LLM-facing schema, and persist tools_bind on step rows (migration 009_tools_bind.sql). Rejected on commit, compensation, and agent-adapter: simple.

  • provider: azure — Azure OpenAI / Microsoft Foundry via LangChain ChatOpenAI and the OpenAI-compatible /openai/v1/ path (AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT; worker model_name is the deployment name). Defaults to Chat Completions for prompt-cache friendliness; Responses API is opt-in via WARDEN_AZURE_USE_RESPONSES_API.

  • Submit-mode ReAct soft-feeds recoverable tool mismatches (e.g. search_replace old_text not found / non-unique match, missing path, patch-apply text failures) into the transcript with a one-line recovery hint instead of failing the step with TOOL_OUTPUT_ERROR. apply_patch_sandbox JSON rejects get the same hint. Transport/MCP error / invalid-argument failures remain hard. Override via ToolLifecycleHooks.tool_output_is_recoverable.

  • Submit-mode ReAct softens recoverable tool invoke exceptions (e.g. IsADirectoryError, Pydantic arg validation) by normalizing them to Error: tool payloads and reusing the same recoverability classifier; MCP CallToolResult.isError is honored at the tool wrapper. Infrastructure failures (MCPError, connection/timeout, mixed ExceptionGroup with transport leaves) still raise TOOL_INVOKE_FAILED.

Changed

  • Definition list/get/patch routes share a generic CRUD helper; PATCH identity moved from path UUID to query id or namespace+name+version on PATCH /v1/definitions/{workers|steps|sagas}.
  • Catalog inactive/missing errors use a shared exception hierarchy (InactiveCatalogDefinitionError / CatalogDefinitionNotFoundError) with structured HTTP detail payloads.

Fixed

  • Parent saga deploy rejects inactive child saga definitions referenced by spawn_sagas (must exist and is_active); inactive catalog refs map to HTTP 409 with structured INACTIVE_CATALOG_DEFINITION (missing → 404 CATALOG_DEFINITION_NOT_FOUND).

  • Soft-disable catalog definitions via PATCH /v1/definitions/{workers|steps|sagas}?id= or ?namespace=&name=&version= (CLI: --id XOR --namespace/--name/--version).

  • Schedule-time input validation logs a warning when resolved arguments are present but input_ports is empty (no hard-fail for engine-native steps).

  • Child spawn maps missing/inactive child definitions to clean STEP_FAILED codes (SPAWN_CHILD_DEFINITION_NOT_FOUND / SPAWN_CHILD_DEFINITION_INACTIVE) instead of unhandled exceptions.

  • Policy CEL is frozen into frozen_steps / policy_definition at saga start (parity with compensation); runtime gates evaluate the embed strictly — no POLICIES_ROOT fallback (migration 014_policy_definition.sql).

  • output_schema JSON is frozen onto frozen_steps as output_schema_definition at start/spawn hydrate; materialize and loop mint read the embed only (no runtime disk reload).

  • Reason-step Jinja prompts and skill payloads are frozen at saga start as prompt_definition / skills_definition (static bare {% include %} inlined via Jinja lexer/AST; migration 015_prompt_skills_definition.sql). Workers render and load_skill from those embeds — no live PROMPTS_ROOT / SKILLS_ROOT reads during task execution. Incomplete skill embeds fail at start freeze; schedule also requires full skills.allow coverage.

  • Child spawn maps hydrate/asset freeze failures to SPAWN_CHILD_HYDRATE_FAILED. Child creates use a nested transaction (savepoint) so a mid-loop failure rolls back partial children instead of committing orphans beside a failed spawn.

  • Worker command setup catches missing/inactive worker catalog errors (CATALOG_DEFINITION_NOT_FOUND / INACTIVE_CATALOG_DEFINITION) after the idempotency claim is taken: emit STEP_FAILED/COMPENSATION_FAILED and release the claim instead of leaving the step stuck IN_PROGRESS.

  • Policy refs must be exact paths under POLICIES_ROOT (e.g. gate.yaml); the stem-only .yaml fallback and deploy-time legacy warning are removed.

  • Startup schema sentinels now require worker_definitions, definition body/is_active columns, and tools_bind / policy_definition / prompt_definition / skills_definition on step instances.

  • Prompt Jinja rendering uses SandboxedEnvironment (file + inline paths), strips introspection helpers (cycler / joiner / namespace / lipsum), and fails closed on unsafe attribute access.

  • ReAct _submit is accepted only as a singleton tool batch. Mixed batches run the other tools, feed a _submit must be alone tool error, and continue the loop (no silent short-circuit that drops sibling calls).

  • Reason / simple human messages no longer json.dumps string prompts (Jinja and inline templates stay plain text). Structured dict/list prompt inputs are still JSON-encoded.

  • Worker-scoped skills (SKILLS_ROOT/<worker>/<id>.md): step skills.allow drives a static allowed_skills index + virtual load_skill; skill frontmatter allowed_tools unions with tools.allow extras for react reason steps (007_skills_allow.sql).

  • Coerce sloppy LLM tool arguments against MCP inputSchema before ReAct validation (common Ollama/vLLM stringified array/object fields)

  • Admit LLM JSON against reason-step output_schema before validation (_submit and simple structured output)

  • LLM JSON admission no longer crashes on nullable union type arrays; coerce string "null" / "none" to JSON null when the schema allows it

  • no_submit_call no longer lists successful plain-text MCP tool output in last_tool_errors; expose final assistant text as last_assistant_content on model_text_exit

Assets 2
Loading