Skip to main content

Supervaizer Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]​

Added​

  • supervaizer login — Logs a developer in to Studio from the terminal with a device authorization: the command prints an approval page and a code, the developer signs in or signs up in the browser (Google included) and approves, and the command writes the three SUPERVAIZE_* variables and the three SUPERVAIZER_WORKSPACE_AUTH_* variables to an environment file (.env by default). It prints variable names only, never a value. supervaizer start --env-file <file> loads such a file before the control script. After a registration, the start log shows the Studio page of each agent when Studio sends it (studio_url in the handshake agent entries). Needs a Studio with the /api/v1/cli/device/ endpoints; an older Studio gets a clear error.
  • Integration skill in the SDK (PR #165) — The Supervaizer integration skill for coding agents moves from runwaize/runwaize_skills to skills/supervaizer-integration/ and is released with the SDK. It interviews the developer about the mission, job, case and step mapping in rounds of questions, then builds the integration topic by topic (parameters, fields, shared resources, cost, human in the loop, guardrail controls, Configure screen, analytics). It holds the process only and points to the guides at docs.runwaize.com, supervaizer scaffold, docs/ and the Hello World example for the contract, so it cannot drift from them. .claude-plugin/marketplace.json lets Claude Code install it with /plugin marketplace add supervaize/supervaizer. The skill is not part of the wheel. docs/2026_05_SUPERVAIZER_v2.md now documents the job.configurator and job.analytics surfaces, the Dashboard document and the job snapshot keys Studio reads. The skill also covers the Studio side (account, Developer access, API key, registration, workspace authorization, refused registrations), hosting and deployment, artifacts, scheduled start and the billing catalog. Its helper prints no environment value, fails on a JSON-RPC error, forwards the job, case, step and draft-session ids, sends no made-up job id, keeps an existing spec file unless --force is given, and lists the cost task only when the spec has a cost.
  • Agent billable consumption (PR #155) — Agent(billing_catalog=[...]) and AgentRegistrationContract.billing_catalog (BillingCatalogItem: sku of 1 to 64 characters matching ^[a-z0-9_]+$ and unique in the list, unit of 1 to 32 characters, description of at most 255) are sent only when set, so an agent without a catalog registers with the same keys as before. Account.report_consumption(records, *, agent_slug, workspace_id=None) and report_consumption_sync post a ConsumptionReport (agent_slug, 1 to 200 BillableConsumption records) to /w/<workspace>/api/v1/billing/consumption/ with the workspace header set to the same workspace, and return one ConsumptionReceipt (accepted, duplicate, or rejected with a reason) per record, in order, parsed through ConsumptionReportResponse. quantity and agent_cost_usd are Decimal values sent as strings; occurred_at must be timezone-aware. Transport errors and whole-batch 400/403 are re-raised like send_event; local mode accepts every record without HTTP. HELLO_WORLD_BILLING_CATALOG in the Hello World example shows a one-SKU catalog. See docs/2026_10_BILLABLE_CONSUMPTION.md. Requires a minor release.

Changed​

  • Production Python refactors (PR #163): Agent lifecycle, CLI startup, contracts, authentication, REST routes, deployment, and example jobs use smaller functions. Shared helpers replace repeated decisions. Characterization tests preserve existing payloads, errors, logs, and execution order. All scored production Python functions have CRAP below 10. The documentation generator remains exempt.
  • just ship-reconcile takes no arguments (PR #156) — The recipe only forwarded its own defaults, so the script's repo argument and --timeout-seconds/--poll-seconds parser are gone. Override with environment variables: TIMEOUT_SECONDS (default 1800), POLL_SECONDS (default 15), e.g. TIMEOUT_SECONDS=7200 just ship-reconcile. The script refuses any argument (exit 2), so a stale command cannot run against a repo it did not name; the repo is fixed to supervaize/supervaizer. Tooling only, no package change.
  • Admin and workbench route factories split (PR #163) — create_admin_routes (CRAP 1046) and create_workbench_routes (CRAP 1104) held every handler as a nested closure. Handlers are now AdminHandlers methods and module-level workbench functions, registered from route tables; the OpenAPI schema is unchanged. Shared helpers replace copied filter/sort/paginate, status-update, step-validation and error-to-500 code. set_server_start_time, SERVER_START_TIME and the EntityFilter model were unused and are removed. Internal module only, no public API change.

Fixed​

  • Workbench job monitor returned 500 (PR #163) — workbench_monitor.html used x is not none if x is defined else false, which Jinja 3.1 parses as a test argument, so GET /manage/agents/<slug>/workbench/jobs/<job_id> raised TemplateSyntaxError for every existing job. The condition is now x is defined and x is not none; a test compiles every admin template.

Tests​

uv run just test-cov -p no:sugar -q

StatusCount
✅ Passed1559
🤔 Skipped0
🔴 Failed0
⏱️ in127s
MetricBefore (5ad84c6)After (97ea54f)
Tests1,0031,559
src + tools coverage74.4% (9,415 statements, 2,413 missed)85.4% (9,475 statements, 1,380 missed)
src only coverage82.0% (1,541 missed)94.1% (508 missed)
tools only coverage0% (872 statements)0% (872 statements)

[1.9.0] - 2026-10-29​

Added​

  • Agent full description (PR #151) — Agent(full_description=...) and AgentRegistrationContract.full_description (V2AgentFullDescription: format is markdown (default) or html, content is 1 to 100,000 characters and not whitespace only) are sent only when set, so an agent without one registers with the same keys as before. The SDK types and bounds the content; Studio sanitizes it before rendering it on the agent detail page. Requires a minor release.

  • Guided setup contract (freeze of 2026-10-28) (PR #149) — Typed, optional additions for chat-guided job setup. An unmodified v2 agent registers byte-for-byte as before; tests/test_registration_compat.py holds the golden payloads. Agent(introduction=...) and AgentRegistrationContract.introduction (V2AgentIntroduction: tagline, summary, example_goals, avatar_url) are sent only when set. V2SetupFieldDefinition with V2FieldOption, V2FieldDependsOn (wire key in), V2FieldValidation and list fields with item_fields; V2AwaitingFieldDefinition and V2ResourceFieldDefinition gain the same optional keys. Validators reject bare-string options, rules that do not fit the field type, min_rows without multiline, and nested list items. V2JobSetupPolicy.plan and V2ActionResult.setup_plan accept a V2SetupPlan (confirmation_required, steps[] of V2SetupStep, cost_estimate) or a dict that passes through untouched; a step's help_actions must be declared in capabilities.actions with kind="help". V2ActionDefinition.kind (command or help; help actions must be mutating=False), target_field and dry_run; V2ActionResult.draft (V2FieldDraft, V2Suggestion) and V2ActionResult.dry_run (V2DryRunOutcome). V2CostEstimate is an estimated range or unknown. Confirm idempotency rules are documented in docs/2026_05_SUPERVAIZER_v2.md, with the conformance helper assert_confirm_idempotent(handler, request) in supervaizer.testing. Requires a minor release.

Changed​

  • A2UI file fields and Studio-stored artifacts (PR #139) — docs/2026_05_SUPERVAIZER_v2.md documents type: "file" awaiting fields (accept, max_bytes, storage), the HITL answer shapes for storage: "studio" and storage: "inline", the inline request-body limit, and artifact refs that point at Studio-held bytes ("storage": "studio" + file_id). File fields are case awaiting steps only for now.
  • just ship-reconcile opens a back-merge PR (PR #137) — The script pushed develop directly, which passed only through an admin bypass of the develop PR rule. It now merges origin/main into a chore/sync-main-<tag> branch cut from origin/develop, pushes that branch, and opens a PR to develop; merge it with a merge commit. The --mode option (and the recipe's mode argument) is removed: rebase needed a non-fast-forward push to develop, which the ruleset blocks. .claude/ship.md records the release flow for agents.
  • Dependencies — uv.lock: opentelemetry-api 1.45.1 (PR #150), virtualenv 21.14.13 (PR #146), uv 0.13.4 (PR #144), certifi 2026.10.19 (PR #145), filelock 4.2.1 (PR #147). GitHub Actions: astral-sh/setup-uv 10.3.0 (PR #142), trufflesecurity/trufflehog 3.99.3 (PR #138, #148), google/osv-scanner-action reusable workflows 2.7.0 (PR #140, #143). No runtime change.

Tests​

just test

StatusCount
✅ Passed930
🤔 Skipped0
🔴 Failed0
⏱️ in44s

[1.8.0] - 2026-10-11​

Added​

  • Scheduled start capability (PR #132) — v2 agents opt in with capabilities.scheduled_start: true; the default and explicit False omit the field. Registration rejects the opt-in unless the final actions declare job.start. Studio owns the clock and invokes the existing action with a stable occurrence idempotency_key. See docs/2026_10_SCHEDULED_START.md.
  • Product feature map (PR #131) — docs/product/feature-map.md indexes user workflows with routes, source, and pytest evidence. AGENTS.md and the README point at it. Update the map when a workflow, route, or access rule changes.
  • Mission context on v2 requests and DataResource calls (PR #130) — New V2MissionContext (id, name, metadata) carries Studio Mission.metadata to the agent, uninterpreted. V2SurfaceRequest, V2ActionRequest, DataResourceContext, and DataResourceContextContract gain an optional mission; a model validator requires a non-blank mission.id equal to mission_id, and metadata accepts a JSON object only. New header X-Supervaize-Mission-Metadata (base64url UTF-8 JSON, 4096 bytes max) on DataResource proxy calls: build_data_resource_context_headers(..., mission_metadata=...) writes it, encode_mission_metadata_header / decode_mission_metadata_header are exported, and the data routes set context.mission or return HTTP 400 on a bad header or a header without a non-blank X-Supervaize-Mission-Id. Optional and additive: requests without mission behave as before. Requires a minor release.

Changed​

  • Ruff 0.16 default lint rules (PR #133) — pyproject.toml drops the select = ["E4", "E7", "E9", "F"] pin from PR #123 and lints with the Ruff 0.16 defaults. No-op findings are fixed: import order, quoted annotations, explicit subprocess.run(check=False), string os.getenv defaults, and similar rewrites. Rules whose fix would change runtime behavior are ignored with a reason in pyproject.toml (BLE001, S110, DTZ005, TRY004, RUF012, UP040/UP046/UP047, plus per-file ignores); B008 accepts FastAPI Body, Depends, and Query defaults. No runtime change. A temporary B023 ignore in routes.py marked a known bug in v1 custom-method routes; PR #134 fixes it and removes the ignore.
  • Dependencies — jaraco-context 6.1.2 in uv.lock (dev extra, via hatch and keyring; fixes PYSEC-2026-1469 / GHSA-58pv-8j8x-9vj2) (PR #135); astral-sh/setup-uv 10.2.0 (PR #128); softprops/action-gh-release 3.0.3 (PR #126); trufflesecurity/trufflehog 3.97.9 (PR #129); google/osv-scanner-action reusable workflows 2.6.0 (PR #125, #127).

Fixed​

  • v1 custom-method routes ran the last registered method (PR #134) — POST /api/supervaizer/agents/{slug}/custom/{method} now calls service_job_custom with its own method name, and the response message names that method. Before, every custom route ran the last method in agent.methods.custom. Route paths, operation names, and OpenAPI output do not change. The B023 ignore from PR #133 is removed.
  • README deploy section — deploy up builds the image, pushes it to the platform registry, and stores generated SUPERVAIZER_API_KEY and SUPERVAIZER_PRIVATE_KEY as platform secrets. The previous "experimental, does not push" note described an older command.
  • Workbench is v1-only — POST /manage/agents/{slug}/workbench/start returns HTTP 400 for an agent with no v1 methods and does not create a job. v2-only agents run through POST /a2a.
  • Data-resource routes were undocumented — CRUD lives at /api/agents/{slug}/data/{name}/, requires X-API-Key plus a workspace token, and is not covered by the local-mode /a2a bypass.
  • Local-mode host and workspace-auth docs — Wildcard hosts (0.0.0.0, ::, empty) bind 127.0.0.1; an explicit --host is kept. The /a2a bypass ignores leeway and SUPERVAIZER_WORKSPACE_AUTH_REQUIRED=false. POST /a2a is mounted by default; only /api waits for a Studio account or local mode. JSON-RPC -32030 uses error.data.code workspace_authorization_missing or workspace_authorization_not_configured.

Tests​

just test

StatusCount
✅ Passed851
🤔 Skipped0
🔴 Failed0
⏱️ in48s

[1.7.0] - 2026-10-04​

Added​

  • Local mode can invoke v2 actions and surfaces (PR #117) — With SUPERVAIZER_LOCAL_MODE=true, no supervisor_account, and no SUPERVAIZER_WORKSPACE_AUTH_* trust setting, POST /a2a supervaizer/action.invoke and supervaizer/surface.load no longer answer -32030 workspace_authorization_not_configured. Handlers get an unverified workspace context with grant_id="local-mode" (LOCAL_MODE_WORKSPACE_GRANT_ID) and only the scopes the call needs, and startup logs a bypass warning. Any trust setting (even without SUPERVAIZER_WORKSPACE_AUTH_REQUIRED=true) or an attached Studio account keeps the fail-closed check, and data-resource routes never bypass.
  • Rule checkpoint gates (PR #113) — V2 controllers can explicitly declare before and after rule-checkpoint support, including declared business-secret paths. RuleCheckpointGate and Case.run_guarded_step() request Studio authorization around a cooperative effect, persist the outcome for safe recovery, and receive approved pauses through the declared rule.checkpoint.resume action.
  • Optional label and description on surfaces and actions (PR #103) — Additive v2 registration contract: V2SurfaceDefinition(id, label, description), V2AgentCapabilities.surface_definitions, and optional label / description on V2ActionDefinition. Display text is trimmed and capped at 300 characters; blank counts as absent. Existing agents that omit the fields keep byte-identical cards (capabilities.surfaces stays a list of ids).

Changed​

  • Shorter v2 on-ramp (PR #121) — build_v2_agent_registration needs only agent_slug, display_name, and a2ui_catalog_version. agent_id defaults to the slug, controller_url to /a2a, and agent_card_url to /.well-known/agents/<slug>_agent.json (the unversioned card route); explicit values keep their behavior. supervaizer scaffold generates a minimal Supervaizer v2 controller (one registration, one job.start surface, one job.start action) instead of the v1 AgentMethod template. docs/api/openapi.json is regenerated from it, so the v1 job routes of the old example leave the published reference; the SDK still serves them for v1 agents. The README quick start is a single short v2 agent.
  • v1 methods are optional for v2 agents (PR #114) — Server(agents=[agent]) no longer raises Agent <name> has no methods defined when the /api router is mounted (Studio account or SUPERVAIZER_LOCAL_MODE=true) and the agent declares only supervaizer_v2_registration. Such agents keep the agent-level routes (info, instructions, parameter validation, /parameters) but get no v1 job routes (/jobs, /stop, /status); Studio runs their jobs through /a2a. The registration payload is unchanged (methods: {}). An agent with neither methods nor a v2 registration still raises. The README v2 sample drops its placeholder job_start method.
  • Documentation audit (PR #104) — README, AGENTS.md, CONTRIBUTING.md, SECURITY.md, and docs/ verified against the 1.6.0 code. Fixed: admin UI path and access model (/manage, Tailscale-gated, ADMIN_ALLOWED_IPS removed), the README v2 sample (mission_id, explicit V2ActionDefinition), Studio connection (Account wiring and the SUPERVAIZER_WORKSPACE_AUTH_* launch requirement), CLI reference (actual deploy options, non-functional --reload/--debug), persistence (the singleton ignores db_path; no CachingMiddleware), workspace authorization (WorkspaceAuthorizationClaims, exact bootstrap ids, env vars), CONTRIBUTING (mypy path, 88 columns, gitmoji commits, PRs target develop), and SECURITY (supported versions, actual branch rulesets). docs/2025_08_ADMIN_README.md and docs/2025_08_REST_API.md rewritten; docs/2025_10_LOCAL_TESTING.md marked legacy; RFC statuses updated; GEMINI.md and .cursorrules now point to AGENTS.md.
  • Deploy credentials stay out of the image (PR #111) — The generated Dockerfile no longer bakes SUPERVAIZE_API_KEY and the other host variables in as ARG/ENV; up passes them at runtime and stores SUPERVAIZE_API_KEY in the platform secret manager. docker-compose.yml and the DigitalOcean app spec are written with mode 600, and deploy local publishes its port on 127.0.0.1 only.
  • deploy up --image takes name[:tag] (PR #111) — up adds the platform registry and rejects a value that contains /.
  • DigitalOcean uses the account container registry (PR #111) — up no longer creates a {name}-{env} registry and down no longer deletes it; DigitalOcean allows one registry per account.
  • Deployment drivers declare their registry (PR #111) — BaseDriver gains prepare_registry() and an overridable registry_auth(). prepare_registry() is not abstract, so existing third-party driver subclasses still instantiate; deploy up raises NotImplementedError for a driver that does not implement it.
  • Agent Card points at real routes (PR #115) — api_endpoints[].url and its examples now use /api/supervaizer/agents/{slug}, and human_url points at the agent's instructions page, which needs X-API-Key like every /api/* route. authentication declares the X-API-Key header instead of none. Behavior change for card consumers: authentication.type is now apiKey; version_info.latest (always equal to current), the dead /changelog/{slug} URL, and logo_url (no route served it) are removed; version_info.changelog_url appears only when the agent sets release_notes_url. Studio reads only the card's supervaizer.v2 extension, so it is not affected.
  • Dependencies — anyio 4.14.2 (PR #102); httpcore2 2.10.0 (PR #100); httpx2 2.12.0 (PR #101); urllib3 2.8.0 (PR #105); virtualenv 21.7.13 (PR #106); actions/checkout 7.0.1 (PR #97); astral-sh/setup-uv 10.0.1 (PR #96); trufflesecurity/trufflehog 3.97.1 (PR #98); google/osv-scanner-action reusable workflows 2.5.1 (PR #95, #99). Minimum versions raised (PR #122): deploy extra boto3 1.43.108, docker 7.2.0, google-cloud-artifact-registry 1.23.0, google-cloud-run 0.16.2, google-cloud-secret-manager 2.31.0; dev extra hatch 1.18.1, mypy 2.4.0, pre-commit 4.6.2, pytest-mock 3.16.0, ruff 0.16.10.

Deprecated​

  • supervaizer start --debug and --reload (PR #111) — Both options still run but are hidden from --help and print a deprecation warning. --debug still sets SUPERVAIZER_DEBUG and turns on FastAPI debug mode in the no-script --local fallback; Server itself never reads the variable. --reload still sets SUPERVAIZER_RELOAD, but the fallback no longer passes it to uvicorn, where it crashed (You must pass the application as an import string). For debug mode, set debug=True on the Server in a control script. Restart after code changes.

Removed​

  • Stale docs and TODO block (PR #104) — docs/2025_04_API_REFERENCE.md (v1 examples that no longer run), docs/2025_08_PROTOCOLS.md (subset of 2026_05_PROTOCOLS.md), docs/superpowers/ plans (shipped or removed features), and the CHANGELOG TODO block. docs/2025_08_PARAMETER_VALIDATION_IMPLEMENTATION.md trimmed to docs/2025_08_PARAMETER_VALIDATION.md.
  • Dead admin allowlist and Server.verify_api_key (PR #112) — Deletes admin/ip_allowlist.py (ADMIN_ALLOWED_IPS, unused since 0.15.0), Server.verify_api_key / Server.api_key_header (API-key checks live on the router), and tools/test_admin.py.
  • Root-level smoke scripts (PR #111) — test_local.py (it ran python -m supervaizer, which has no __main__; use supervaizer deploy local) and test_dockerfile_generation.py (license header only).
  • No-op deploy flags (PR #111) — deploy up --yes, deploy up --no-rollback, and --verbose on deploy plan, up, down, and status did nothing. They now fail as unknown options. deploy down --yes and --verbose on deploy local and clean remain.

Fixed​

  • /manage pages no longer expose the API key (PR #116) — Every admin page put the write-scope SUPERVAIZER_API_KEY in a data-admin-key attribute on <body>, and the workbench page also inlined it in its script. Anyone who could open /manage could read the key and call the write-scope /api/* routes with it. Since 0.15.0, require_tailscale gates /manage and its endpoints need no key, so the templates and workbench-form.js no longer receive or send it. The /manage/job-start-test page and job-start-form.js are removed: the page never loaded the script, and the script targeted an agent path that does not exist.
  • Job responses and storage no longer hold decrypted parameters (PR #118) — POST /api/supervaizer/agents/{slug}/jobs returned the decrypted agent_parameters list, secret values included, and Job.to_dict wrote it to the job store. AbstractJob.agent_parameters is now excluded from serialization and repr; the agent method still gets the values at run time. service_job_custom passes its decrypted list to that run only (new Agent.job_start agent_parameters argument), because a job reloaded from storage no longer has it, and the job_start and _execute debug logs no longer print it. Job records saved before this release keep their parameters, because the store merges updates: after you upgrade, remove agent_parameters from the Job records in $DATA_STORAGE_PATH/entities.json (default ./data/entities.json). Behavior change: Job.to_dict, the job start response, and the OpenAPI Job schema no longer include agent_parameters.
  • Valid empty registration parameter setup (PR #109) — Agents without parameter definitions now send an empty list instead of null.
  • Safer HTTP error diagnostics (PR #107, #108) — Diagnostic curl output masks authorization and API-key headers, serialized event details redact API keys and token fields, and HTTP errors log the remote status and request ID.
  • just dev starts the controller (PR #111) — The recipe ran uvicorn controller:app --reload, but there is no controller module. just dev is now an alias of just local (uv run supervaizer start --local).
  • Correct refresh hint (PR #111) — When supervaize_instructions.html already exists, the CLI now points to supervaizer scaffold refresh-instructions instead of the nonexistent supervaizer refresh-instructions.
  • Git hooks run the pre-commit checks (PR #111) — just install-hooks installed the .pre-commit-config.yaml hooks in .git/hooks and then set core.hooksPath=.githooks, so Git ignored them, and .githooks/pre-commit only called a pre-commit-user script that did not exist. .githooks/pre-commit now runs the pre-commit stage (Ruff, mypy, license header) on the staged files and still refuses a plain git commit on gitbutler/workspace; the new .githooks/commit-msg runs the gitmoji and commitizen checks. .githooks/post-commit reads src/supervaizer/__version__.py, so a version-bump commit is tagged vX.Y.Z instead of v. just install-hooks only sets core.hooksPath. Removed the GitButler-managed .githooks/post-checkout, which deleted itself from the tree on each checkout away from gitbutler/workspace. just ship passes --no-verify to its release merge, because the commit-msg hook rejects the [MINOR] title. GitButler's but commit does not run Git hooks.
  • Post-commit hook tags release versions only (PR #119) — A plain git commit of just version-dev on created a local vX.Y.Z.dev0 tag, and just version-dev off failed with tag 'vX.Y.Z' already exists but still printed Created tag. The hook now calls just tag-version, which tags only X.Y.Z versions, skips an existing tag, and prints Created tag only when git tag succeeds.
  • deploy up pushes the image it deploys (PR #111) — up now pushes the built image to the platform registry (Artifact Registry, ECR, or the account DigitalOcean registry, with short-lived credentials) and deploys that reference. Before, Cloud Run and App Runner pointed at an image that was never pushed, and DigitalOcean used a placeholder GitHub source. Cloud Run now deploys with allow_missing and waits for the operation; App Runner gets a string map in RuntimeEnvironmentVariables and secret ARNs in RuntimeEnvironmentSecrets; DigitalOcean SECRET envs carry their value.
  • Deployed containers get the variables the server reads (PR #111) — Generated secrets arrive as SUPERVAIZER_API_KEY and SUPERVAIZER_PRIVATE_KEY (was {name}-{env}-api-key, {name}-{env}-rsa-key, and SV_RSA_PRIVATE_KEY), and the log level as SUPERVAIZER_LOG_LEVEL (was SV_LOG_LEVEL). SUPERVAIZER_SERVER_ID and the SUPERVAIZER_WORKSPACE_AUTH_* variables are forwarded, so a Studio-registered v2 controller starts in the container. --port now sets SUPERVAIZER_PORT as well as the container port.
  • deploy local health check and Compose (PR #111) — The health check probes /.well-known/health (the /agents/health route no longer exists), commands use Docker Compose v2 (docker compose), the host SUPERVAIZE_API_KEY is no longer replaced by the controller test key, and multi-line values such as PEM keys are quoted in docker-compose.yml.
  • Launch stops when the Studio agent lookup fails (PR #120) — Agent.update_agent_from_server used to log a failed GET /agents/{id} or /agents/by-slug/{slug} and return None, so Server.launch kept going without the agent's Studio status or parameters. It now raises RuntimeError naming the slug, server_agent_id, URL, and HTTP status. A 404 from Studio hid an Agent ID mismatch this way. Behavior change: a Studio-registered controller no longer starts while Studio cannot return its agent. Local mode (no Studio account) still returns None.
  • Parameter validation accepts the job-start payload (PR #115) — POST /api/supervaizer/agents/{slug}/validate-agent-parameters turned the decrypted list of {name, value} objects that Studio sends into param_0, param_1, ... keys, which then failed as unknown. Job start (Job.new) only accepts that list, so an agent with required parameters had no payload that passed both. The route now reads the list by name, so one encrypted payload passes both. The {NAME: value} dict form stays accepted for validation only; job start still requires the list. A list item without a string name and a value, or a payload of another type, now fails with a clear error instead of being validated as empty.
  • Deploy home-page template links (PR #115) — deploy/templates/index.html linked to /admin and /console; it now links to /manage and /manage/console.
  • Rule checkpoint recovery (PR #113) — Early resume delivery wakes live waiters, resumed proposals retain step identity, and lost after responses retry only evaluation. Concurrent original execution and recovery cannot advance a step twice.
  • Ruff rule set is explicit (PR #123) — Ruff 0.16 enlarged its default rule set, and the pre-commit check failed with 266 findings. [tool.ruff.lint] select = ["E4", "E7", "E9", "F"] keeps the rules that ran before. Ruff 0.16 also formats Python code blocks in Markdown, so three docs files are reformatted.

Tests​

just test

StatusCount
✅ Passed808
🤔 Skipped0
🔴 Failed0
⏱️ in43s

[1.6.0] - 2026-08-30​

Added​

  • Declared action scope and mutability (#92) — V2AgentCapabilities.actions is now list[V2ActionDefinition] instead of an opaque list[str]. Each action declares id, mutating (does invoking it change agent-side state?), and scope (workspace / mission / job), so a consumer can authorize and group actions without pattern-matching the identifier. id is caller-supplied at invocation time, so deriving an authorization decision from it puts that decision in the caller's hands.

    Plain strings are still accepted everywhere a list of actions is taken — in build_v2_agent_registration(actions=...) and when validating a registration payload — and coerce to mutating=True, scope="job", the fail-closed reading that matches today's behaviour. Only the serialized form changes: capabilities.actions now emits objects. Consumers must accept both, because agents pinned to older Supervaizer releases keep sending bare strings. No protocol version discriminates the two shapes: versions.a2ui_version and versions.a2a_version cover the surface catalog and the A2A protocol, not the registration payload.

    The builder derives the metadata from the definition each action comes from: resource actions take V2ResourceDefinition.scope, dataset queries are mutating=False, V2JobSetupPolicy.preview_action is mutating=False, and workspace binding actions are workspace-scoped. Resource operation ids are freeform, so operations stay mutating=True unless the agent declares otherwise.

    Metadata precedence is explicit > derived > bare: an explicit V2ActionDefinition in actions= overrides the derived metadata — that is how a job-scoped resource.invoice.reconcile on a workspace-scoped resource is declared — while a bare id string declares nothing and defers to the definition it was derived from. Precedence is about metadata, not position: an id declared both ways within actions= keeps the explicit metadata regardless of which form comes first, and the earlier mention still fixes the order.

  • V2DatasetDefinition.scope — Mirrors V2ResourceDefinition.scope (workspace / mission / job, defaulting to workspace) so dataset query actions carry a declared scope rather than an assumed one.

  • V2AwaitingState.reopenable — Declares whether an already-answered awaiting step may be reopened and resubmitted with edited values. Defaults to False; previously consumers inferred this from a substring of surface.

Tests​

just test

StatusCount
✅ Passed698
🤔 Skipped0
🔴 Failed0
⏱️ in67s

[1.5.0] - 2026-08-28​

Added​

  • Generic job setup contract — New V2JobSetupPolicy on V2JobPolicy.setup lets an agent declare the actions Studio may use to preview, start, and submit a job setup (preview_action, start_action, submit_action, defaulting to job.start.preview, job.start, and step.awaiting.submit), the contexts those actions apply to (action_scopes), and an opaque plan payload. Declared actions are folded into the registration's capability action list. V2ActionResult.setup_plan carries the agent's plan back to Studio. V2JobSetupPolicy is exported at package level. Additive only; agents that omit setup are unaffected.

    action_scopes defaults to an empty list: an agent that opts in with setup: {} declares no scopes rather than claiming all of workspace, job, case, and step. This matches V2JobSyncPolicy.supported_statuses and avoids advertising context support the agent never configured. The three action ids are validated as non-blank, so a blank id is rejected at registration instead of being silently dropped from capabilities.actions while still appearing in the serialized job_policy.setup. The same validation was applied to V2JobSyncPolicy.action, which had the identical gap.

Changed​

  • Regenerated model reference and OpenAPI — docs/model_reference/ and docs/api/openapi.json were rebuilt from the current models, picking up the job setup contract along with accumulated drift since 0.20.1. V2JobSetupPolicy is now declared before V2JobPolicy so the public reference renders the real type instead of a ForwardRef.

Fixed​

  • PyPI publish (Metadata 2.5) — Pin pypa/gh-action-pypi-publish to v1.14.2 so hatchling ≥1.32 wheels upload. Skip version bump unless the triggering commit has [MAJOR]/[MINOR]/[PATCH].

Tests​

just test

StatusCount
✅ Passed691
🤔 Skipped0
🔴 Failed0
⏱️ in65s

[1.4.0] - 2026-07-07​

Added​

  • Security & performance review summary — Added docs/2026_07_SECURITY_REVIEW.md, a non-actionable high-level summary of a full-source security and performance/scalability review (posture, verified-sound controls, severity counts, and remediation themes). Per SECURITY.md, detailed findings (locations, attack scenarios, remediation specifics) are handled through the private vulnerability channel and are intentionally omitted from the public repository.

Fixed​

  • Hardened API-key checks — API keys are compared in constant time.
  • Safer local test mode — supervaizer start --local binds to loopback (127.0.0.1) by default instead of all interfaces; pass an explicit --host to override.
  • Baseline security response headers — Responses now set X-Content-Type-Options, X-Frame-Options, Referrer-Policy, and Strict-Transport-Security (streaming/SSE-safe).
  • Reduced sensitive data in logs — Agent parameter values are no longer written to logs during parameter validation.
  • Scheduled-step execution hardening — The scheduler only runs methods declared by the agent that owns the step's job.

Tests​

StatusCount
✅ Passed683
🤔 Skipped0
🔴 Failed0
⏱️ in83s

[1.3.1] - 2026-07-02​

Changed​

  • pyproject.toml (since v1.3.0) — Runtime lower bounds: FastAPI >=0.139.0, sse-starlette >=3.4.5, Typer >=0.26.8, Uvicorn >=0.49.0. deploy extra: boto3 >=1.43.39, google-cloud-run >=0.16.1, google-cloud-secret-manager >=2.23.0. dev extra: hatch >=1.17.0, pytest >=9.1.1, pytest-asyncio >=1.4.0, ruff >=0.15.20. Lock file refreshed via uv lock.

Fixed​

  • A2A event scope test — tests/test_a2a.py now walks FastAPI included-router wrappers when locating /a2a/events, preserving the read-scope assertion under FastAPI 0.139.0.

Tests​

StatusCount
✅ Passed683
🤔 Skipped0
🔴 Failed0
⏱️ in56s

[1.3.0] - 2026-07-02​

Added​

  • context.assign contract — New typed wire contract for assigning a frozen Studio context selection to a job: V2ContextAssignment (items, mission_id, job_id, assigned_at), V2ContextAssignmentItem (ref, version, scope, title), and the V2_ACTION_CONTEXT_ASSIGN action id in supervaizer.contracts, exported at package level. Additive only; dispatch uses the existing generic Server.v2_action machinery.

Tests​

  • tests/test_common.py — structured JSON log output for API access-denial records
  • just test
StatusCount
✅ Passed677
🤔 Skipped0
🔴 Failed0
⏱️ in65s

[1.2.0] - 2026-05-27​

Changed​

  • FastAPI lifespan cleanup — Controller shutdown now cancels the scheduled-step background loop and waits briefly for it to stop.
  • Server module refactor — Split server startup configuration, runtime server info persistence, Studio registration handshakes, and scheduled-step loop management out of server.py into focused modules while preserving the public Server behavior.
  • Supervaizer v2 agent methods — SDK agents can now declare optional standard actions such as agent.refresh plus custom agent actions through the same AgentMethods structure used for job methods, and the A2A runtime registers those handlers automatically.

Tests​

  • tests/test_server.py — scheduler task cancellation and bounded shutdown waiting during FastAPI lifespan shutdown.
  • tests/test_server_refactor_modules.py — parity coverage for the extracted server configuration, server info, registration, handshake, and scheduled-step helpers.
  • tests/test_a2a.py — standard and custom agent method dispatch through the v2 A2A controller.
  • tests/test_agent.py — agent-level v2 method registration and contract validation.
  • tests/test_contracts.py — typed agent method contract serialization.

Tests​

  • tests/test_common.py — structured JSON log output for API access-denial records
  • just test
StatusCount
✅ Passed672
🤔 Skipped0
🔴 Failed0
⏱️ in86s

[1.1.1] - 2026-05-20​

Changed​

  • Cloud Logging structured output — SUPERVAIZER_LOG_FORMAT=json now switches controller stderr logs to newline-delimited JSON with Cloud Logging-compatible severity plus bound fields such as access-denial path, reason, and truncated key_preview; local text logging remains the default.

Tests​

  • tests/test_common.py — structured JSON log output for API access-denial records
  • just test
StatusCount
✅ Passed658
🤔 Skipped0
🔴 Failed0
⏱️ in92s

[1.1.0] - 2026-05-19​

Added​

  • Workspace agent authorization — Studio-signed Ed25519 workspace authorization tokens on X-Supervaize-Workspace-Authorization; the SDK verifies JWKS-backed tokens and exposes V2VerifiedWorkspaceContext for handlers. Workspace and tenant slugs remain routing hints only.
  • Workspace binding rotocol — Agents can declare optional workspace_binding metadata with bootstrap workspace_binding.options, workspace_binding.create, and the workspace_binding.create surface so Studio can bind an agent-side record before a Workspace Agent Grant exists.
  • Workspace authorization docs — docs/2026_05_WORKSPACE_AGENT_GRANTS.md plus workspace authorization and binding bootstrap rules in docs/2026_05_PROTOCOLS.md and docs/2026_05_SUPERVAIZER_v2.md.

Changed​

  • Fail-closed Studio A2A — Workspace-scoped v2 JSON-RPC actions and surfaces reject requests without a valid workspace authorization token when workspace authorization is enabled.
  • Workspace-scoped data resources — Data resource routes require verified workspace context from the authorization token instead of trusting caller-supplied slugs alone.
  • Studio server audience handoff — server.register handshakes can now supply the Studio-persisted server audience for workspace authorization tokens, and Supervaizer adopts that audience before serving protected v2 calls so workspace grants survive agent process restarts.
  • Controller version registration — server.register now sends the Supervaizer controller package version directly as controller_version, so Studio no longer depends on OpenAPI scraping to refresh the server detail page version.

Fixed​

  • Workspace authorization validation — Malformed or incomplete workspace authorization tokens and settings now fail with explicit workspace_authorization_* errors instead of ambiguous handler failures.

Tests​

  • tests/test_a2a.py — workspace authorization accept/reject paths, JWKS verification, binding bootstrap exceptions, and protected action/surface enforcement.
  • tests/test_agent.py — v2 registration carries workspace binding and authorization settings.
  • tests/test_contracts.py — workspace binding and authorization contract models.
  • tests/test_routes.py — data resource routes require verified workspace context.
  • tests/test_server.py — registration handshake audience handoff, workspace authorization startup validation, and controller_version registration.

just test

StatusCount
✅ Passed657
🤔 Skipped0
🔴 Failed0
⏱️ in76s

[1.0.1] - 2026-05-17​

Supervaizer v2 2️⃣​

  • Supervaizer v2 case metadata — V2CaseSnapshot now carries optional public metadata so agents can expose case-level business context such as contact interview links without adding agent-specific fields to the protocol.
  • Studio registration handshake validation — After server.register, startup validates supervaizer_handshake.controller_api_key_match and fails when Studio persisted a different controller API key than the one this process is using (configured SUPERVAIZER_API_KEY or auto-generated key).
  • Controller API key reload stability — Auto-generated controller keys are written to SUPERVAIZER_API_KEY before registration so Uvicorn reload keeps the same key; startup logs a short SHA-256 fingerprint instead of printing the full secret.

Tests​

  • tests/test_contracts.py — V2CaseSnapshot accepts public metadata.
  • tests/test_server.py — registration handshake accept/reject paths and generated API key export for reload.

just test

StatusCount
✅ Passed645
🤔 Skipped0
🔴 Failed0
⏱️ in77s

[1.0.0] - 2026-05-17​

Supervaizer v2 2️⃣​

  • Supervaizer v2 documentation — Added docs/2026_05_SUPERVAIZER_v2.md and updated the README/protocol docs to explain the v2 A2A/A2UI layering, registration model, resources, datasets, surfaces, actions, HITL, artifacts, and job.sync convergence semantics.
  • Supervaizer v2 README — Reworked the repository README as a v2-first onboarding guide, linking to the protocol docs and Hello World example while removing the old v1 scaffold/job-field quick start path.
  • Supervaizer v2 contract primitives — Added typed SDK models for the v2 registration and action contract, including pinned A2UI/A2A versions, resources, datasets, case lanes, artifact declarations, job snapshots, sync metadata, and replay-safety metadata.
  • A2A JSON-RPC action runtime — Added the /a2a supervaizer/action.invoke dispatcher, v2 Agent Card extension payloads, and public SDK helpers for registering typed v2 actions through Server.register_v2_action() and @server.v2_action(...).
  • A2A SSE event stream — Added /a2a/events and an in-process v2 event bus so action effects returned through supervaizer/action.invoke can also be observed over Server-Sent Events.
  • Supervaizer v2 transport honesty — V2A2ATransport.push_notifications now defaults to false; the MVP advertises JSON-RPC and SSE support only until A2A push notifications are implemented.
  • A2A JSON-RPC surface runtime — Added supervaizer/surface.load, typed V2SurfaceRequest/V2SurfaceResult models, and public SDK helpers for registering A2UI surface handlers through Server.register_v2_surface() and @server.v2_surface(...).
  • Supervaizer v2 agent identity guard — Agent now rejects v2 registration payloads whose declared agent.slug differs from the runtime SDK slug, preventing A2A action handlers from registering under one slug while Studio invokes another.
  • Supervaizer v2 job sync state — V2JobSyncResult now carries an optional job_state snapshot so agents can return convergent Job/Case/Step/Artifact state through job.sync.
  • Supervaizer v2 job source target metadata — V2JobSource now includes an optional target_type so external sources can declare the business object Studio should use for dedupe and catch-up.
  • Supervaizer v2 resource form fields — V2ResourceDefinition now carries typed fields metadata so Studio can render simple agent-owned resource create/edit forms without callback-style dynamic field logic.
  • Supervaizer v2 ResourceImport documents — Added a typed A2UI ResourceImport document contract so agents can declare import formats, contextual fields, row columns, and submit actions without adding agent-specific protocol fields.
  • Supervaizer v2 resource option sources — Resource fields can now declare typed resource-backed options_source metadata so Studio can render relationship selectors without callback-style dynamic choices.
  • Supervaizer v2 workspace-scoped resources — Resource declarations now advertise scope and requires_context metadata, and legacy DataResource registration info carries the same workspace context requirement for Studio-side access control.
  • Supervaizer v2 dataset display metadata — V2DatasetDefinition now carries typed display columns so Studio can render generic dataset surfaces from the registration contract.
  • Supervaizer v2 dashboard widgets — Added generic dashboard and widget registration contracts, including dataset/action/inline data refs and Vega-Lite widget specs under visualization: { type: "vega-lite", spec: ... } without AnalyticsResource REST routes.
  • Supervaizer v2 mounted HITL review docs — Documented DocumentReview as a generic A2UI document payload for mounted awaiting-step surfaces without adding agent-specific protocol fields.
  • Supervaizer v2 registration builder — Added build_v2_agent_registration() to derive validated v2 registration payloads from public SDK primitives, including generated resource/dataset surfaces and actions.
  • Supervaizer v2 public SDK ergonomics — Exported the remaining v2 registration helper models from supervaizer and neutralized agent-specific examples in v2 public documentation and contract descriptions.
  • Legacy dynamic choices removed — Removed the v1 dynamic_choices field metadata, dynamic_choices_callback, /start/dynamic_choices route, and related contract exports; dynamic options now belong to v2 resource options_source metadata or typed A2A actions.
  • Legacy job poll removed — Removed job_poll from the public v1 method contract and the local workbench poll route/button; v2 status convergence is represented by the typed job.sync action.
  • Supervaizer v2 awaiting form fields — Step awaiting state can now declare typed form fields so Studio can submit HITL actions through step.awaiting.submit.
  • Local Hello World v2 contract — The built-in local Hello World agent now declares a minimal Supervaizer v2 registration through the SDK builder and registers job.start, job.sync, case.step.awaiting, and a generated resource handler for local Studio and SDK smoke tests.
  • Controller correctness fixes — Scoped Jobs.get_job(..., agent_name=...) to the requested agent, gave EventType.AGENT_SEND_ANOMALY a distinct value, and fixed custom method routes to parse normal JSON request bodies.
  • Supervaizer v2 runtime hardening — Added typed common fields to v2 action effects, including resource import counts, rows, errors, gaps, summaries, and case snapshots; validated job_state on action results; authenticated the A2A JSON-RPC controller and SSE event stream while keeping A2A discovery/health public; sanitized A2A handler errors; and constrained legacy method execution to declared non-blocked method paths.
  • Supervaizer v2 replay and context contracts — Action results now validate and serialize replay-safety metadata, and DataResource context header generation includes the agent slug advertised by the contract model.

Tests​

  • uv run pytest tests/test_a2a.py tests/test_contracts.py -q

just test

StatusCount
✅ Passed601
🤔 Skipped0
🔴 Failed0
⏱️ in69s

[0.20.1] - 2026-05-13​

Security​

  • uv.lock — Refreshed transitive versions to address open Dependabot / GHSA advisories on the default branch graph: urllib3 (redirect and decompression-chain issues), requests (extract_zipped_paths temp reuse), protobuf (JSON recursion depth), pyasn1 (decoder / recursion DoS), pygments (ReDoS in GUID lexer), and uv (ZIP / tar / RECORD handling; dev dependency via hatch).

[0.20.0] - 2026-05-13​

Security​

  • Secret scanning + push protection — enabled on the repository; credentials pushed by mistake are now blocked at the source.
  • Dependabot alerts + automated security updates — enabled; known-CVE dependency updates are now proposed automatically.
  • Private vulnerability reporting — enabled; researchers can report issues privately via GitHub Security Advisories instead of public issues.
  • Branch ruleset on main — replaces classic branch protection; requires all CI checks (pre-commit, build 3.10–3.13) to pass before merge, enforced on admins. Classic protection had no required checks and did not enforce on admins.
  • Tag ruleset — v* release tags are now immutable; they cannot be moved or deleted.
  • pypi environment protection — required reviewer approval added; a push to main no longer publishes to PyPI without a human gate.
  • SHA-pinned GitHub Actions — all third-party actions pinned to commit SHAs (trufflehog, setup-uv, gh-action-pypi-publish, action-gh-release). Previously trufflesecurity/trufflehog@main was a floating branch reference — a critical supply-chain risk.
  • uv sync --frozen — enforced in CI; the lockfile can no longer silently change during a build.
  • OSV-Scanner (google/osv-scanner-action) — added as a daily scheduled scan and on every PR targeting main or develop, against the OSV.dev malicious package index.
  • Dependabot configuration (.github/dependabot.yml) — weekly updates for pip and github-actions ecosystems with a 5-day cooldown window to mitigate supply-chain worm attacks; all PRs routed to develop.
  • SECURITY.md — added; documents the responsible disclosure process (GitHub private advisories) and the repository's supply-chain posture.
  • AGENTS.md security rules — mandatory rules added for coding agents: no direct pushes to main, no manual lockfile edits, no workflow modifications without approval, no local hatch publish.

Tests​

just test

StatusCount
✅ Passed570
🤔 Skipped0
🔴 Failed0
⏱️ in69s

[0.19.0] - 2026-05-11​

Added​

  • Agent method timeout metadata — AgentMethod and the registration contract now expose is_async and timeout metadata for Studio. timeout defaults to 600 seconds and can be null for controller jobs that should run until Studio stops them manually.

Changed​

  • pyproject.toml (since v0.18.0) — Runtime lower bounds: FastAPI, orjson, packaging, Pydantic, python-slugify, Rich, sse-starlette, TinyDB, Typer, Uvicorn. deploy extra: boto3 and Google Cloud libraries (Artifact Registry, Cloud Run, Secret Manager). dev extra: add boto3, docker, and black so uv sync --extra dev matches deploy/docker-heavy tests and Black in CI; add yamllint; bump mypy, pre-commit, pytest, respx, ruff. [tool.black]: explicit Black config (line length 88, Python 3.12–3.13 targets, excludes for caches/venv/dist).

Tests​

  • uv run pytest tests/test_agent.py tests/test_contracts.py

just test

StatusCount
✅ Passed570
🤔 Skipped0
🔴 Failed0
⏱️ in69s

[0.18.0] - 2026-05-04​

[0.17.3] - 2026-05-04​

Added​

  • POST /api/supervaizer/registration/refresh — Adds an authenticated controller endpoint that lets Studio request asynchronous re-registration. The route requires write-scope API-key access, reuses the existing supervisor account registration flow, and is advertised through the controller contract as Controller.POST_CONTROLLER_REGISTRATION_REFRESH.

Tests​

just test

StatusCount
✅ Passed564
🤔 Skipped0
🔴 Failed0
⏱️ in69s

[0.17.2] - 2026-05-03​

Added​

  • POST /api/supervaizer/agents/{agent_slug}/status — Controller contract adds Controller.POST_AGENT_STATUS; HTTP handler returns the agent's job_status result as a JobResponse (job id, status, message, payload). OpenAPI lists explicit 200 (JobResponse) and error responses (ErrorResponse for 400/404/500). Returns 404 when job_status yields no result.

Changed​

  • routes.py — synchronous agent hooks off the event loop — job_stop, job_status, and the start flow's dynamic_choices_callback run inside asyncio.to_thread so blocking synchronous agent code does not stall the async server.
  • routes.py / OpenAPI — Agent POST /status success payload is a JobResponse with explicit HTTP 200 (replacing the previous AgentResponse-shaped merge).
  • case.py — logging — Routine case-method logs demoted from INFO to DEBUG to reduce noise under normal operation.
  • agent.py — docstrings — Agent._execute, job_stop, and job_status document synchronous hook behavior and that the controller invokes them from a worker thread.
  • AGENTS.md / repo hygiene — Dropped generated AGENTS.compiled.md; Supervaizer's in-repo agent guide is consolidated in AGENTS.md.
  • publish-pypi.yml — post-publish release automation — After PyPI publish, CI now runs the same GitHub release flow as just gh-release (tools/gh-release-latest-tag.sh) to create/update and mark the latest release from the newest origin/main tag.
  • publish-pypi.yml — branch reconciliation in CI — Added an automatic main -> develop merge-back step after publish/release so long-lived branch history stays synchronized without manual follow-up.

Fixed​

  • GitHub Actions Node 20 deprecation warnings — Upgraded workflow action majors across CI/release/publish pipelines: actions/checkout@v5, actions/setup-python@v6, and astral-sh/setup-uv@v7 to avoid Node 20 runtime deprecation warnings and align with Node 24 transition.

Tests​

just test

StatusCount
✅ Passed560
🤔 Skipped0
🔴 Failed0
⏱️ in69s

[0.17.1] - 2026-04-26​

Fixed​

  • server.py — startup lifespan migration — Replaced deprecated @app.on_event("startup") with a proper asynccontextmanager lifespan passed to FastAPI(lifespan=...). Eliminates DeprecationWarning from FastAPI 0.100+.
  • account_service.py — atexit robustness — close_httpx_client_sync now guards the aclose() call with getattr(_httpx_client, "aclose", None) to prevent AttributeError when the module-level client is a sync httpx.Client in older-installed environments.

Tests​

just test

StatusCount
✅ Passed559
🤔 Skipped0
🔴 Failed0
⏱️ in74s

[0.17.0] - 2026-04-26​

Changed​

  • Async control-event HTTP client — account_service.send_event, Account event helpers, and case event-reporting methods now use httpx.AsyncClient and must be awaited in async agents. Explicit _sync shims remain for CLI, startup, and synchronous controller methods. See RFC-002 for details.

Tests​

just test

StatusCount
✅ Passed559
🤔 Skipped0
🔴 Failed0
⏱️ in
⏱️ with cov102s

Coverage​

NameStmtsMissCover
src/supervaizer/init.py23387%
src/supervaizer/version.py30100%
src/supervaizer/access/init.py40100%
src/supervaizer/access/api_auth.py38197%
src/supervaizer/access/client_ip.py37781%
src/supervaizer/access/tailscale.py23291%
src/supervaizer/account.py1221588%
src/supervaizer/account_service.py80298%
src/supervaizer/admin/routes.py54919065%
src/supervaizer/admin/workbench_routes.py38025633%
src/supervaizer/agent.py3548077%
src/supervaizer/case.py2676476%
src/supervaizer/cli.py2107564%
src/supervaizer/common.py142497%
src/supervaizer/contracts.py168398%
src/supervaizer/data_resource.py700100%
src/supervaizer/data_routes.py982179%
src/supervaizer/deploy/init.py20100%
src/supervaizer/deploy/cli.py1044656%
src/supervaizer/deploy/commands/init.py20100%
src/supervaizer/deploy/commands/clean.py1581789%
src/supervaizer/deploy/commands/down.py611280%
src/supervaizer/deploy/commands/local.py2141593%
src/supervaizer/deploy/commands/plan.py75791%
src/supervaizer/deploy/commands/status.py1004159%
src/supervaizer/deploy/commands/up.py113893%
src/supervaizer/deploy/docker.py1862388%
src/supervaizer/deploy/driver_factory.py200100%
src/supervaizer/deploy/drivers/init.py13469%
src/supervaizer/deploy/drivers/aws_app_runner.py22115729%
src/supervaizer/deploy/drivers/base.py82890%
src/supervaizer/deploy/drivers/cloud_run.py20512539%
src/supervaizer/deploy/drivers/do_app_platform.py16412524%
src/supervaizer/deploy/health.py1611293%
src/supervaizer/deploy/state.py115992%
src/supervaizer/deploy/utils.py240100%
src/supervaizer/event.py40198%
src/supervaizer/examples/local_server.py17382%
src/supervaizer/instructions.py56198%
src/supervaizer/job.py1662883%
src/supervaizer/job_service.py340100%
src/supervaizer/lifecycle.py153398%
src/supervaizer/parameter.py79791%
src/supervaizer/protocol/init.py20100%
src/supervaizer/protocol/a2a/init.py30100%
src/supervaizer/protocol/a2a/model.py37586%
src/supervaizer/protocol/a2a/routes.py35391%
src/supervaizer/routers/init.py40100%
src/supervaizer/routers/api.py18194%
src/supervaizer/routers/private.py100100%
src/supervaizer/routers/public.py22577%
src/supervaizer/routes.py3336082%
src/supervaizer/server.py2828470%
src/supervaizer/server_utils.py250100%
src/supervaizer/storage.py1601292%
src/supervaizer/telemetry.py400100%
src/supervaizer/utils/init.py20100%
src/supervaizer/utils/version_check.py27678%
---------------------------------------------------------------
TOTAL6133155175%

[0.16.0] - 2026-04-25​

Added​

  • Controller contract metadata for Studio — server.register now advertises controller_contract_version, /api base path, and canonical endpoint templates so Studio can route without hardcoded Supervaizer paths. The same contract is available at GET /api/supervaizer/contract.
  • DataResource request context — DataResource callbacks can opt into a context keyword with workspace, mission, agent, and request identifiers while legacy callbacks continue to work unchanged.
  • CaseNodes update helpers — CaseNodes.node_index() and CaseNodes.make_update() derive step indexes from the declared node set instead of requiring consumers to maintain a separate index map.

Changed​

  • DataResource fields can be marked sensitive — DataResourceField.sensitive is included in registration payloads for Studio masking.

Tests​

  • New: tests/test_contracts.py for contract endpoint templates, schema export, import safety, resolver behavior, and DataResource context headers.
  • New: tests/test_case.py coverage for CaseNodes.node_index() and CaseNodes.make_update().
  • Updated: tests/test_routes.py to verify DataResource callbacks receive DataResourceContext from forwarded X-Supervaize-* headers.
  • Updated: tests/test_server.py to assert registration info and /api/supervaizer/contract expose matching /api contract metadata.
  • Deleted: none.

just test

StatusCount
✅ Passed514
🤔 Skipped0
🔴 Failed0
⏱️ in01:03

[0.15.1] - 2026-04-20​

Fixed​

  • TemplateResponse Starlette 0.50 compatibility — Updated routers/public.py and routes.py to use the new TemplateResponse(request, name, context) API. The old two-argument form TemplateResponse(name, context) was removed in Starlette 0.50.0, causing a TypeError: unhashable type: 'dict' crash on every GET / request.

[0.15.0] - 2026-04-20​

Added​

  • supervaizer.access module — New package with three focused sub-modules for centralized access control:

    • client_ip.py — _extract_client_ip(scope) extracts the real client IP from ASGI scope, honoring TRUSTED_PROXIES (env var, comma-separated CIDRs) to safely parse X-Forwarded-For; returns "" on any parse error (fail-closed).
    • tailscale.py — require_tailscale FastAPI dependency enforces that requests originate from the Tailscale CGNAT range 100.64.0.0/10; raises HTTP 403 or WebSocketException(1008) on denial; logs via log_access_denied_tailscale.
    • api_auth.py — require_api_key / require_scope FastAPI dependencies for machine-to-machine auth with a hierarchical scope model (write implies read). API_KEYS registry is empty by default; SUPERVAIZER_API_KEY env var is pre-loaded as a write-scope entry at import time.

  • supervaizer.routers module — Three router factories that replace scattered per-route Security(...) calls:

    • public_router — unauthenticated surface for home page (/) and A2A discovery (/.well-known/*).
    • private_router (prefix /manage) — admin UI and workbench WebSocket, gated by require_tailscale at router level.
    • api_router (prefix /api) — machine-to-machine surface (/api/supervaizer/…, /api/agents/{slug}/…), gated by require_api_key; write-mutating endpoints additionally enforce require_scope("write").

  • log_access_denied_tailscale and log_access_denied_api helpers in supervaizer.common — structured WARNING log entries for every denied request, including IP, path, reason, and a truncated key preview (never the raw key value).

Changed​

  • Admin UI moved from /admin to /manage — All admin routes, workbench, and HTML template links updated. private_router (prefix /manage) gates the surface with Tailscale-only access instead of the previous AdminIPAllowlistMiddleware + API-key combo.

  • API routes moved from /supervaizer/… to /api/supervaizer/… — All machine-to-machine endpoints now live under the /api prefix provided by api_router. Clients must update base paths accordingly.

  • AdminIPAllowlistMiddleware removed — Replaced by require_tailscale at router level. The admin/ip_allowlist.py module is deleted.

  • Per-route Security(server.verify_api_key) removed from routes.py and data_routes.py — Authentication is now enforced once at the api_router level.

  • Admin auth simplified to Tailscale-only — verify_admin_access, ?key= query-param handling, and console-token generation/validation are removed from admin/routes.py and admin/workbench_routes.py.

Security​

  • Removed hard-coded default API keys — API_KEYS is now empty at startup; no credentials ship with the package. Only SUPERVAIZER_API_KEY (operator-supplied env var) populates the registry.

Tests​

  • New: tests/test_access_client_ip.py, tests/test_access_tailscale.py, tests/test_access_api_auth.py covering the new access layer.
  • Updated: test_routes.py, test_routes_case_update.py, test_data_resource.py — paths prefixed with /api.
  • Updated: test_admin_routes.py, test_workbench_routes.py — prefix /admin → /manage; Tailscale gate bypassed via dependency_overrides.
  • Deleted: test_admin_ip_allowlist.py — coverage moved to new access tests.

just test

StatusCount
✅ Passed505
🤔 Skipped0
🔴 Failed0
⏱️ in01:07

[0.14.2] - 2026-04-16​

Added​

  • Automatic data resource route mounting at startup — Server now mounts generated CRUD routers for any agent that declares data_resources, so endpoints are available immediately after server boot without extra manual wiring.

Fixed​

  • DataResource.registration_info now includes computed display_label per field — Resource field payloads sent to Studio now expose resolved labels (display_label) instead of only raw model dumps, ensuring proper field naming in the UI.

  • JobStartConfirmationEvent is emitted after job_start execution metadata is available — The confirmation event is now sent after job.metadata is populated from the job_start response payload, so Studio receives complete metadata in the start confirmation event.

Unit Tests Results​

just test

StatusCount
✅ Passed500
🤔 Skipped0
🔴 Failed0
⏱️ in~70s

[0.14.1] - 2026-04-15​

[0.14.0] - 2026-04-14​

Added​

  • DataResource class — Declares agent-owned CRUD endpoints exposed to Studio with name, entity_type, description, operations (list of CRUD operations), importable (bulk import support), and deletable flags.

  • DataResourceField class — Describes field schema with name, type, description, editable, and visible attributes for validated rendering in Studio forms.

  • FieldType enum — Validated field types: STRING, INTEGER, BOOLEAN, DATE, DATETIME, TEXT, EMAIL, URL for consistent data handling across agents and Studio.

  • Editable enum — Controls Studio form behaviour per field: ALWAYS (edit in all forms), CREATE_ONLY (edit only on creation), NEVER (read-only display).

  • metadata: dict on AbstractJob and CaseAbstractModel — Arbitrary metadata flows through registration_info to Studio, enabling agents to attach custom context and tracking data to jobs and cases.

  • data_resources: list[DataResource] on Agent — Included in registration_info for Studio to discover and render CRUD interfaces for agent-managed data.

  • Auto-generated FastAPI CRUD routes — For each declared DataResource operation, Supervaizer auto-mounts routes (GET, POST, PATCH, DELETE) at /agents/{slug}/data/{resource}/....

  • Bulk import route — When importable=True on a DataResource, a POST /data/{resource}/import/ route accepts CSV or JSON for batch creation, enabling Studio to load data in bulk.

Unit Tests Results​

just test

StatusCount
✅ Passed492
🤔 Skipped0
🔴 Failed0
⏱️ in~70s

[0.13.3] 2026-04-14​

Added​

  • CaseNodeUpdate.upsert and Case.patch_step — Optional step update path for Studio: when upsert is true, the existing case step at the same index is updated instead of appending. Case.patch_step(index, update) sets index and upsert on the update, sends send_update_case, and replaces the matching entry in Case.updates. Serialized in CaseNodeUpdate.registration_info for the controller payload.

  • Human answer with casestep_index — POST /jobs/{job_id}/cases/{case_id}/update: if request.answer includes casestep_index, the controller calls case.patch_step(int(casestep_index), update) and runs PersistentEntityLifecycle.handle_event(..., INPUT_RECEIVED) instead of receive_human_input. Omit casestep_index for the previous append/receive-human-input behavior.

  • Tests — tests/test_routes_case_update.py covers job 404, workbench-style human_answer params (including casestep_index stripped from fields), single owning agent vs multiple agents, and skip when job.agent_name is not on the server.

Changed​

  • Dynamic choices request context — POST .../start/dynamic_choices now passes workspace_slug through to dynamic_choices_callback alongside workspace_id and mission_id (Supervaize Studio sends it in the JSON body).

  • POST /jobs/{job_id}/cases/{case_id}/update (human_answer) — Resolves the job with in-memory Jobs().get_job first, then persisted (include_persisted=True) if missing. Returns 404 when the job does not exist. Dispatches human_answer only for the job owner via server.get_agent_by_name(job.agent_name) and agent._execute(...), using the same parameter shape as the workbench HITL route (fields, context, payload, job_id, case_id, optional message). Strips casestep_index from fields for the hook. Runs the hook in a thread pool executor to avoid blocking the event loop.

Unit Tests Results​

just test

StatusCount
✅ Passed473
🤔 Skipped0
🔴 Failed0
⏱️ in~70s

[0.13.1]​

Added​

  • ADMIN_ALLOWED_IPS for admin UI — When set, only matching client IPs may access /admin (HTML, APIs, static files, WebSocket). Comma-separated IPs and optional CIDR notation; empty or unset allows all. Uses the first address in X-Forwarded-For when present.
  • Dynamic choices for AgentMethodField — Fields can now use dynamic_choices instead of static choices to resolve options at runtime via a callback. Add a dynamic_choices_callback callable (signature: (method_name: str, context: dict) -> dict[str, list[tuple[str, str]]]) to the Agent constructor and a dynamic_choices key to your AgentMethodField. Supervaize Studio fetches choices from the new POST /supervaizer/agents/{slug}/start/dynamic_choices endpoint (with workspace_id, workspace_slug, and mission_id in the request body) when rendering the job start form. Static choices and dynamic_choices are mutually exclusive on a field. See Dynamic Choices guide.

Unit Tests Results​

just test

StatusCount
✅ Passed464
🤔 Skipped0
🔴 Failed0
⏱️ in65s

[0.12.0]​

Added​

  • Custom routes — Agents can mount their own FastAPI APIRouter via the new custom_routes field on Agent. Supervaizer mounts them under the API router at /api/agents/{slug}/... (paths defined on the nested router append after that prefix) without inspecting or managing the routes. Enables agents to expose tool endpoints, webhooks, or any HTTP API alongside the workbench.

  • Scheduled steps — CaseNodeUpdate gains scheduled_at, scheduled_method, scheduled_params, scheduled_status fields. Steps with scheduled_at are deferred until the scheduled time. A background executor polls every 60 seconds and calls the agent method automatically. The workbench shows countdown, "Execute now", and "Cancel" controls on pending scheduled steps. Enables time-based orchestration (call scheduling, retries with backoff, follow-up actions).

    • Model: CaseNodeUpdate.scheduled_at / scheduled_method / scheduled_params / scheduled_status
    • Property: CaseNodeUpdate.is_scheduled returns True when scheduled_at is set
    • Executor: _run_scheduled_step_loop — background asyncio task, polls every 60s
    • Routes: POST .../execute, POST .../cancel, PATCH .../schedule for step management
    • UI: status badges (pending/executing/completed/failed/cancelled), execute now / cancel buttons
    • Case.cancel_scheduled_steps() — cancels all pending steps on job stop
    • Cases.get_due_scheduled_steps() — returns steps where scheduled_at <= now() and status is pending

  • AgentResponse export — AgentResponse is now exported from supervaizer.__init__ for use as a typed response model in custom routes and agent endpoints.

Fixed​

  • OpenAPI / JSON Schema (Pydantic 2.12+) — Building the full FastAPI schema (GET /openapi.json, Swagger UI) could raise PydanticInvalidForJsonSchema: … CallableSchema. Three root causes fixed:
    • AgentMethodAbstract.model_config["example_dict"] used "type": str (Python builtin) — changed to string "str"
    • AgentResponse nested schemas not rebuilt after dependent models — added AgentResponse.model_rebuild() after Case.model_rebuild() in supervaizer/__init__.py
    • CaseNode.factory typed as Callable which cannot appear in JSON Schema — changed to Any (behaviour unchanged)

[0.11.0]​

  • Job Poll mechanism — New optional job_poll method in AgentMethods for manual external service polling. When defined, the workbench shows a "Check for updates" button on active jobs. Clicking it calls the agent's poll handler, which checks external services (email inboxes, call status APIs, etc.) and updates cases accordingly. Enables local development without webhooks — production uses real-time webhooks, local mode uses the poll button.

    • AgentMethods: new job_poll: AgentMethod | None field
    • Route: POST /workbench/jobs/{job_id}/poll triggers the agent's poll handler
    • UI: "Check for updates" button with loading state, conditionally rendered via has_poll template variable
    • JS: pollJob() in workbench-form.js with disable-during-request pattern

  • HITL double-click guard — All HITL buttons are disabled during submission via _postAnswer(). Buttons get opacity-50 cursor-not-allowed while in flight and re-enable on error. Dialog HITL submitMessage/confirmDialog are now async, properly awaiting the form submission before resetting state.

  • Monitor reply display — Step payloads with a reply key render as a blue blockquote below the step row. Step payloads with approved_content render as a green confirmed card.

  • Monitor duration display — step_duration format now handles both numeric and string values, preventing Jinja2 TypeError when duration comes from HITL form input.

  • WebSocket terminal signal — Terminal job state sends a terminal WS signal that keeps the connection idle, preventing reconnect loops.

  • is_local_mode() helper — Centralized in supervaizer.common, replacing inline os.environ.get("SUPERVAIZER_LOCAL_MODE") checks across account_service.py, server.py.

  • deque for log buffer — _workbench_log_buffer changed from list to collections.deque(maxlen=500), removing manual trim logic.

  • Dialog HITL — New HITL type for interactive content review via chat interface. When a CaseNodeUpdate payload contains supervaizer_dialog, the workbench renders a conversation UI instead of a fixed form. Supports iterative refinement through LLM-powered feedback loops. Fields: content (JSON string or plain text), content_type (email/text/code), objective, instructions, messages (conversation history), confirm_label.

    • Template: dialog_renderer.html with render_hitl_dialog and render_dialog_confirmed macros
    • JS: submitDialogMessage(caseId, message) and confirmDialog(caseId) in workbench-form.js
    • Route: workbench_routes.py detects supervaizer_dialog in AWAITING case payloads

  • Local mode URL fixes — supervaizer start --local --port N now correctly shows localhost URLs everywhere: CLI output, admin interface logs, storage, and uvicorn. Fixed by resolving Server.__init__ defaults from env vars at call time (not class definition time) and ensuring CLI-provided values take precedence over .env file values.

  • Local mode event skipping — account_service.send_event() returns a no-op ApiSuccess when SUPERVAIZER_LOCAL_MODE=true, preventing HTTP errors against the SaaS API during local development.

  • Agent parameter env pre-fill — In local mode, the workbench auto-loads .env values into agent parameter fields with green .env badge indicators. Secret fields show a masked placeholder; non-secret fields display the value. Backend falls back to env values for empty fields on job submission.

  • HTMX polling guard — Monitor template suppresses hx-trigger polling when a HITL dialog or form is active, preventing DOM overwrites while users interact with forms.

v0.10.27​

  • Agent Workbench — Full-featured testing interface for agents directly from the admin panel. Four-zone layout with agent parameters, job control, execution monitor, and live console log. Supports starting/stopping jobs, real-time case and step tracking via HTMX polling, and Human-in-the-Loop (HITL) form rendering and submission. Job history panel lists all past executions with status badges.

    • Backend: workbench_routes.py with 8 FastAPI endpoints (page, start, stop, status, monitor, console, HITL answer, job history)
    • Frontend: workbench.html, workbench_monitor.html, workbench_console.html, workbench_jobs_list.html, workbench-form.js
    • Shared field renderer macros: field_renderer.html

  • Local test mode (--local) — Run the supervaizer server without Supervaize Studio credentials for local development and workbench testing. Includes built-in Hello World agent with configurable case count, optional HITL steps, and JobInstructions.check() support for graceful stop. Default API key local-dev, amber banner in admin UI.

    • CLI: supervaizer start --local / just local
    • Agent: examples/hello_world_agent.py with _LocalAccount (no-op Studio stub)
    • Server factory: examples/local_server.py

Changed​

  • Admin navigation — Workbench access moved from nav dropdown to per-agent cards with direct green "Workbench" button. Removed "New Agent" button from agents list.
  • Admin auth — Refactored API key resolution to support local server instance (_get_admin_api_key, _is_local_mode), with fallback chain: server state → env var → default.

Fixed​

  • CaseNode.factory now optional — CaseNode.factory field changed from required to Optional[Callable] = None. Previously, factory was exclude=True (correctly omitted from serialization), but still required by Pydantic validation. This caused AgentResponse(**agent.registration_info) to fail with Field required errors on the GET /supervaizer/agents route, because the round-tripped dict never contained factory. Agents with CaseNode-based workflows were unable to register. (#6)

Changed​

  • CI – Python package workflow: dependency install now uses uv sync --extra dev --extra deploy (adds deploy extras for CI).

Unit Tests Results​

StatusCount
✅ Passed433
🤔 Skipped0
🔴 Failed0
⏱️ in50s

[0.10.19]​

Added​

  • Server: optional local index.html – If a local index.html exists, the server serves it for the home page response.

Changed​

  • Dependencies – Bumped boto3/botocore to 1.42.41 (lock file). Updated ruff to 0.15.0 and pre-commit config.
  • Admin index – Simplified displayed information in the admin index template.
  • CI – PyPI publish workflow: added concurrency settings to cancel in-progress runs.

[0.10.18]​

Added​

  • SUPERVAIZER_PRIVATE_KEY – _get_or_create_private_key reads key from env or generates and sets it. Admin dashboard shows a development-only warning (not for production exposure).

Changed​

  • Refactor – Code formatting and trailing whitespace cleanup (admin routes, server, examples, tests, templates).

[0.10.17]​

Added​

  • SUPERVAIZER_PRIVATE_KEY from env if set; else create key and set env.

  • Admin: live server and data when no persistence – When there is no persistent storage (e.g. Vercel, serverless), the admin dashboard and agents page no longer fail. Server status, server config, and agents use the live server instance (app.state.server) when storage has no ServerInfo. Dashboard stats, jobs list, cases list, and recent activity use live Jobs/Cases registries when storage is in-memory and empty.

Changed​

  • Register with supervisor (no frontend API key) – The "Register with supervisor" button now calls the backend without any API key. The backend sends the SERVER_REGISTER event to SUPERVAIZE_API_URL; no authentication is required from the frontend for this action.
  • Admin server/agents fallback – Server info and agents are read from storage first, then from the live server when storage has no ServerInfo (e.g. persistence disabled). get_server_status and get_server_configuration accept an optional request for the live fallback.

Fixed​

  • Admin: API key "None" in URLs – When SUPERVAIZER_API_KEY is unset, the template no longer renders the literal string "None" (e.g. in data-admin-key). SessionStorage no longer stores "None" from ?key=None, and the frontend treats the string "None" as no key where applicable.

Unit Tests Results​

StatusCount
✅ Passed419
🤔 Skipped6
🔴 Failed0
⏱️ in53.16s

[0.10.11]​

Added​

  • Optional data persistence (default off) – Data is no longer persisted to file by default, so the server runs correctly on Vercel and other serverless platforms where the filesystem is ephemeral.

Changed​

  • Controller unreachable – When the Supervaize controller server is not available (connection refused or timeout), event send now fails gracefully with a short, clear error message instead of dumping the full payload and traceback. Logs: Supervaize controller is not available at {url}. Connection refused or timed out. Is the controller server running? and the event type/exception.
    • Set SUPERVAIZER_PERSISTENCE=true (or 1/yes) to enable file persistence.
    • CLI: supervaizer start --persist enables persistence for that run.
    • Explicit StorageManager(db_path=...) in code still uses file storage (e.g. tests).
    • See 2025_08_PERSISTENCE.md for configuration.

Unit Tests Results​

StatusCount
✅ Passed419
🤔 Skipped6
🔴 Failed0
⏱️ in55.28s

[0.10.4]​

  • Relax some dependencies requirements

[0.10.2]​

Changed​

  • 🐍 Python 3.13 Support - Added Python 3.13 compatibility
    • Updated package classifiers to include Python 3.13
    • Build process now uses Python 3.13 for wheel generation
    • CI test matrix includes Python 3.13
    • Wheels built with Python 3.13 are compatible with both Python 3.12 and 3.13

[0.10.0]​

Added​

  • 🚀 Cloud Deployment CLI - Complete automated deployment system for Supervaizer agents

    • Full implementation of RFC-001: Cloud Deployment CLI
    • Support for three major cloud platforms:
      • Google Cloud Run with Artifact Registry and Secret Manager
      • AWS App Runner with ECR and Secrets Manager
      • DigitalOcean App Platform
    • New deployment commands:
      • supervaizer deploy plan - Preview deployment actions before applying
      • supervaizer deploy up - Deploy to cloud platform with automated build, push, and verification
      • supervaizer deploy down - Tear down deployment and clean up resources
      • supervaizer deploy status - Check deployment status and health
      • supervaizer deploy local - Local Docker testing with docker-compose
      • supervaizer deploy clean - Clean up deployment artifacts and state
    • Automated Docker Workflow: Build → Push → Deploy → Verify
    • Secret Management: Secure handling of API keys and RSA keys via cloud provider secret stores
    • Health Verification: Automatic health checks at /.well-known/health endpoint
    • Idempotent Deployments: Safe create/update operations with rollback on failure
    • Local Testing: Full Docker Compose environment for pre-deployment testing
    • See Local Testing Documentation for details

  • Agent Instructions Template - New HTML page served by FastAPI for Supervaize integration instructions

    • Accessible at /admin/supervaize-instructions
    • Provides step-by-step setup guide for agents

  • Version Check Utility - Automatic check for latest Supervaizer version

    • Helps users stay up-to-date with latest features and fixes
    • Located in supervaizer.utils.version_check

  • Enhanced Admin Interface

    • New agents listing page with grid view
    • Improved agent detail views
    • Better navigation and UI consistency

Changed​

  • 🔄 Protocol Unification - Removed ACP protocol in favor of unified A2A protocol

    • Removed src/supervaizer/protocol/acp/ directory and all ACP-specific code
    • Removed acp_endpoints parameter from Server class
    • Removed ACP route registration and test files
    • Updated all documentation to reflect A2A-only support
    • The A2A protocol has evolved to incorporate features from multiple agent communication standards, including the former ACP
    • All A2A protocol links updated to https://a2a-protocol.org/
    • Breaking Change: acp_endpoints parameter no longer accepted in Server initialization

  • 📦 Dependency Optimization - Cloud SDKs moved to optional dependencies

    • Base package size significantly reduced
    • Cloud deployment dependencies now optional: pip install supervaizer[deploy]
    • Optional deploy group includes: boto3, docker, google-cloud-artifact-registry, google-cloud-run, google-cloud-secret-manager, psutil
    • Removed unused pymongo dependency
    • Updated dependency versions for better compatibility

  • Improved Error Handling - Enhanced API error responses with better context

  • Documentation Updates

    • Added comprehensive deployment documentation
    • Updated model reference documentation
    • Improved README with deployment examples
    • Updated 2025_08_PROTOCOLS.md to focus on unified A2A protocol
    • Added Protocol Evolution section explaining ACP merger

Fixed​

  • API documentation errors corrected
  • Improved type hints for agent_parameters and case_ids in job.py
  • Health logging optimized in A2A routes

Unit Tests Results​

StatusCount
✅ Passed415
🤔 Skipped6
🔴 Failed0
⏱️ in50.56s

Migration Notes​

  • ACP Protocol Removal: If your code uses acp_endpoints=True parameter, remove it from Server initialization. The A2A protocol now provides unified agent communication.
  • If you need deployment features, install with: pip install supervaizer[deploy]
  • For development, install with: pip install supervaizer[dev,deploy]
  • No other breaking changes to existing APIs or functionality

[0.9.8]​

Added​

  • Parameter Validation System: Added comprehensive parameter validation for job creation with clean error messages
    • New validate_parameters() method in ParametersSetup class for agent parameter validation
    • New validate_method_fields() method in AgentMethod class for job field validation
    • Two separate validation endpoints for different validation needs:
      • /validate-agent-parameters - Validate agent configuration parameters (secrets, API keys, etc.)
      • /validate-method-fields - Validate job input fields against method definitions
    • Support for validating both job fields and encrypted agent parameters
    • Clean error messages with specific details about invalid parameter types and missing required parameters

Fixed​

  • Execution of supervaizer start was not maintaining the main namespace so the fastapi server was never starting. Replaced execution by sub-process.
  • Type of agent.choice. #TODO: test and decide which to keep (list[str] or list [tuple[str,str]])
  • When supervisor_account is provided, A2A endpoints are automatically activated, because Supervaize needs to be able to trigger the healthchecks. -export_openapi.py tool to generate openapi.json (for docusaurus documentation) - automation in docusaurus to do.

Changed​

  • Parameter Validation System: Refactored to provide separate validation endpoints for different concerns

    • Agent Parameters: Now validated separately through /validate-agent-parameters endpoint
    • Method Fields: Now validated separately through /validate-method-fields endpoint
    • Clean Architecture: Removed legacy endpoint for cleaner, more focused API design
    • Code Deduplication: Eliminated redundant validation code in job start endpoints
    • Clearer Separation: Agent configuration validation vs. job input validation are now distinct operations

  • pytest does not run with coverage by default (change in pyproject.toml)

Unit tests results​

StatusCount
✅ Passed308
🤔 Skipped6

[0.9.6]​

  • Public release to Pypi
  • Fixed the github workflows
  • Improve README.md

[0.9.5]​

Fixed​

  • Setup : missing py.typed in pyproject
  • clarified public_url (replaced registration_host by public_url)
  • changed "supervaizer install" to "supervaizer scaffold"

Added​

  • gen_model_docs.py: tool for documentation generation - see disclaimer

Unit tests results​

StatusCount
✅ Passed277
🤔 Skipped6

[0.9.4]​

Added​

  • CICD : release, deploy
  • gen_model_docs.py : to generate the documentation of the models.

Changed​

  • Moved "example" to src/supervaizer
  • Improved and Moved some documentation to docs
  • Added python-package.yml github action, triggered on push / PR of "develop" branch

[0.9.3]​

Added​

  • Data persistence with tinyDB
  • Admin UI with fastAdmin
  • Dynamic content on:
    • Server page
    • Agent
    • Jobs
    • Cases
  • Add persisted data to job status check.

Changed​

  • Parameter.to_dict : override to avoid storing secrets.
  • Removed Case Nodes
  • Improved test coverage : accounts, admin/routes,

Unit tests results​

StatusCount
🤔 Skipped6
⚠️ Failed0
✅ Passed281

Test Coverage : Test Coverage

Emoji Legend
🌅 Template🏹 Service👔 Models
🐛 Bug🛣️ Infrastructure/CICD🔌 API
💼 Admin📖 Documentation📰 Events
🧪 Tests🧑‍🎨 UI/Style🎼 Controller

Uploaded on 2026-11-06 15:07:19