# Changelog

## 0.3.8 - 2026-08-30

### Fixes

- Builds missing `dist/` entries during git-hosted installs through `prepare`,
  so DSH can install `git+https://github.com/sandbaseai/sandbase-harness.git`
  without a prior local `npm run build`. Published packages that already ship
  `dist/` skip the rebuild. MCP image builds install dependencies with
  `--ignore-scripts` because the Dockerfile copies `package.json` before
  `scripts/`.
- Counts model usage once per model request in runtime metrics and persists
  aggregate input/output token usage on each session.
- Validates confirmed tool calls against raw AI SDK v4 stream lifecycle data,
  rejects malformed or incomplete calls, and grants one-shot execution
  authority only after validation.
- Documents the maintainer and organic project-promotion workflow in
  `AGENTS.md` and keeps the DeepSeek Harness Handbook discovery link current.

## Unreleased

## 0.3.7 - 2026-08-20

### Fixes

- Makes the DeepSeek Harness bundle boot in a clean Web profile when
  `MANAGED_AGENTS_API_KEY` is unset by keeping all MCP `env` values string-typed.
- Resolves the MCP entry from the installed profile instead of requiring a
  globally linked `managed-agents-mcp` executable on `PATH`.
- Replaces the ambiguous npm-name installation command with a pinned local
  source install, preventing resolution of the unrelated unscoped package.

## 0.3.6 - 2026-08-20

### Fixes

- Restores the documented `/v1` operations API for webhooks, scheduled
  deployments, and outcomes while retaining `/v1/x` compatibility aliases.
- Exposes all three operations screens in the primary Console navigation and
  routes `#scheduled-deployments` to its intended page.
- Wires the complete Memory Stores implementation to prevent list and detail
  views from crashing on missing runtime bindings.

## 0.3.5 - 2026-08-20

### Highlights

- **Agent Skills compatibility metadata**: validates and preserves the standard
  optional `compatibility` field (1–500 characters), exposes it through the API,
  and renders it safely in Console search and detail views.
- **Instant Codespaces evaluation**: adds a development-container configuration
  and documented one-click path to build and open the local Console without
  preparing a host Node.js environment.
- **Verified DSH onboarding**: adds runtime/auth preflight checks and links the
  bilingual DeepSeek Harness developer walkthrough and operator tools.
- **Release-aligned Agent Plugin**: pins the portable MCP plugin to the v0.3.5
  multi-architecture bridge image produced from this tag.

## 0.3.4 - 2026-08-19

### Highlights

- **Published MCP container**: release tags now produce public multi-platform
  `linux/amd64` and `linux/arm64` images in GitHub Container Registry.
- **Verifiable supply chain**: images carry OCI and MCP ownership labels and a
  GitHub build-provenance attestation tied to the release digest.
- **Registry-ready metadata**: adds a version-aligned `server.json` for the
  official MCP Registry OCI distribution format.

## 0.3.3 - 2026-08-19

### Highlights

- **First-class MiniMax provider**: adds MiniMax to Settings V2 and the
  runtime provider boundary, with global and mainland China endpoints.
- **Current MiniMax models**: supports `MiniMax-M3` as the default and
  `MiniMax-M2.7` as an explicit alternative.
- **Verified setup guide**: documents environment-variable credentials,
  regional endpoint selection, agent model IDs, and configuration checks.

## 0.3.2 - 2026-08-15

### Highlights

- **Containerized MCP bridge**: adds `Dockerfile.mcp` and `.dockerignore` for
  the six-tool stdio bridge, with OCI and MCP server metadata.
- **Verified container build**: CI now builds the production MCP image from a
  clean checkout so broken Docker packaging blocks future changes.
- **Release-aligned quick start**: source, Docker, installation, deployment,
  and DeepSeek Harness examples now pin the immutable v0.3.2 tag that contains
  every referenced file.
- **Project trust and discovery**: adds security and contribution policies,
  private vulnerability reporting, and a direct SandBase Agent Skills link.

## 0.3.0 - 2026-08-14

### Highlights

- **DeepSeek Harness integration**: ships an installable Cordis bundle and the
  `managed-agents-mcp` stdio server so DSH can create and inspect agents,
  manage persistent sessions, stream turns, retrieve artifacts, and cancel
  runs through MCP.
- **Verified integration guide**: adds a reproducible DSH configuration,
  permission boundaries, troubleshooting guidance, and compatibility evidence
  against DeepSeek Harness commit `47f9438`.
- **DeepSeek V4 reasoning controls**: forwards Settings V2 `reasoning_effort`
  values to OpenAI-compatible model requests and documents a verified
  DeepSeek V4 setup.
- **Distribution metadata**: corrects package repository links to
  `sandbaseai/sandbase-harness` and adds Glama maintainer metadata for MCP
  directory verification.

## 0.2.0 - 2026-08-01

### Breaking changes

- **Workspace state directory moved**: runtime state (database, config, logs)
  now lives at `<workspace>/.managed-agents/` instead of
  `~/.managed-agents/<name>-<hash>/`. Existing workspaces need their state
  moved manually or will re-initialize on next start.
- **`composeRuntimeFromSettings`**: the `memorySeedEnabled` parameter was
  removed; use `settingsSeed: { memoryEnabled: true }` instead.
- **Legacy provider endpoints permanently removed**: `/v1/x/model-providers`,
  `/memory-providers`, `/storage-providers` return 404. Use
  `/v1/x/settings` to configure providers.

### Highlights

- **Runtime decomposition**: the monolithic route and session files are split
  into focused modules (environments, files, credential-vaults, memory-stores,
  runtime, settings, templates, session-normalizers, session-stream,
  session-records, session-lifecycle, session-recovery, secrets,
  skill-packages, skill-resources, resource-utils). `resources.ts` and
  `extended.ts` are now pure composition roots.
- **Sandbox backends**: local process, Docker (with per-session container
  labels and path confinement), and Kubernetes (kubectl exec/cp transport)
  are shipped; Environment write-side validation rejects unknown providers.
- **Docker environment mode**: `hosting_type: docker` with image/resources
  fields in the Environment config.
- **Console improvements**: YAML/JSON agent config editor, fix modal crashes,
  session send failure handling, environment editor polish, settings form
  splitting.
- **Release gate**: `npm run release:check` covers typecheck (src + tests),
  full test suite, production build, package dry-run, CLI init smoke, and
  example workspace startup.
- **Model registry**: `resolveModelConfig()` supports `openai/gpt-5.5`-style
  qualified model references.
- **Workspace registry**: `managed-agents init` and
  `createRegisteredWorkspace` write config to `.managed-agents/config.yaml`,
  consistent with the runtime's path resolution.

### Stats

- 75 test files, 569 tests (up from 67 / 544 in 0.1.0).
- 24 new source modules, 7 new test files.
- `release:check` passes on a clean checkout.

## 0.1.0 - 2026-07-18

First public release of `managed-agents`.

### Highlights

- Local-first managed agent runtime with a Claude Managed Agents-style `/v1`
  API surface.
- React Dashboard for agents, sessions, environments, credential vaults,
  memory stores, files, skills, settings, logs, monitoring, and API reference.
- SQLite-backed runtime state stored outside source-controlled workspaces by
  default.
- Settings V2 for one workspace model vendor, loop engine, storage backends,
  context-memory backend, sandbox provider, API keys, validation, and restart
  flows.
- Session lifecycle, event replay, resumable SSE streams, memory resources,
  file resources, credential vaults, snapshots, local/Docker/self-hosted
  sandbox registration, MCP tools, and skill packages.
- TypeScript SDK and CLI commands for init, start, list, reload, chat, deploy
  guidance, and templates.
- Release gate covering typecheck, tests, production build, package dry-run,
  CLI init smoke, and example workspace startup smoke.

### Known first-release boundaries

- One active model vendor, one built-in loop engine, SQLite metadata storage,
  local artifact storage, SQLite memory, and runtime-registered sandbox
  providers are supported in 0.1.0.
- Planned adapters such as S3, Postgres/MySQL, mem0, MemU, Harness, Codex, and
  Claude loop engines remain unavailable until their runtime implementations
  are added.
- Live remote model credential checks and production deployment examples are
  tracked as follow-up work.