# Changelog

[完整中文记录](./CHANGELOG.zh.md)

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

## [0.2.2] — 2026-08-14

Security patch and repository-maintenance follow-up.

### Fixed

- Caught server exceptions are no longer copied into HTTP responses. Detailed
  messages stay in local logs; clients receive stable codes and fixed bilingual
  messages, preventing stack traces, local paths, and internal details from
  leaking through the Settings or artifact endpoints.
- Dependabot no longer opens automatic major-version upgrades. Patch and minor
  updates remain automated; major upgrades require explicit compatibility
  review.

## [0.2.1] — 2026-08-14

Security, lifecycle, and first-time-maintainer community release.

### Fixed

- Media provider endpoints now require HTTPS except for explicit loopback
  development hosts; embedded credentials, query strings, and fragments are
  rejected before any request is sent.
- Empty media credential references now perform genuinely unauthenticated
  image-generation and connection-test requests with no Authorization header.
- Subagent model selection is installed only after a validated rule hit and all
  installed listeners are removed when the plugin is disposed.
- Browser source maps now resolve to the real `src/client/index.tsx` source.

### Added

- A real repository and Git-history secret scanner, executable with
  `pnpm check:secrets` and required by CI.
- A CodeQL workflow for public-repository analysis; private-repository
  Advanced Security is never enabled automatically.
- Complete Chinese counterparts for public English documentation, bilingual
  community templates, a model-purchase strategy, an open roadmap, and a code
  of conduct.
- A first-time-maintainer note that invites technical review and small,
  educational contributions without presenting the project as mature.

## [0.2.0] — 2026-08-14

Breaking architecture change: the plugin is now a **thin routing policy layer
over the DSH native model registry** (option 1 of the V2 plan).

### Changed

- Removed the duplicated agent model stack: the `openai-chat` adapter,
  provider registry, wire transport, and `mr:<id>` route registration are
  gone. DSH 模型页 is the only provider/credential/catalog/modality/retry
  configuration entry.
- Rules target NATIVE provider ids only; every hit passes
  `resolveModelInfo` / `resolveCallConfig` before the session header is
  written; failures use stable error codes and never touch the session.
- No-match no longer force-routes (V1 `fallback` removed); the harness
  `agent-default-model` governs.
- `model_route` is default-off and gated by an explicit allowlist; the tool
  exposes only the allowlist and still validates live at execution.
- `reasoningEffort` is an opaque adapter-owned string (no fixed enum);
  adapter defaults are never written back into rules or session headers.
- Media moved to the isolated `mediaProviders` namespace; V1 channels with
  `image_gen` migrate to `openai-images` in memory only (a user-confirmed save
  persists `schemaVersion: 2`). Vision chat is no longer claimed by the
  plugin.
- Hardened media paths: https-only downloads with allowed hosts, ≤3 redirect
  revalidation, private-address/DNS-pinning rejection, 20 MiB streaming cap,
  timeouts, magic-byte image typing, and a realpath-fenced artifact preview.
- 模型中枢 page rewritten: reads the official `llm.providers`/`llm.models`
  RPCs, refreshes on `llm/adapters-updated`/`settings/document-updated`, and
  shows active/dormant provider states, migration warnings, and Beta media
  channels (video/audio/3D marked not implemented).
- Acceptance now separates direct and routing defaults, restores the exact
  deployment default after direct calls, and reports assistant block kinds
  without logging their content.
- `defaultSize` now survives settings snapshot, edit, save, and projection
  roundtrips. Kimi K3 receives route-specific visible-output discipline at
  the official system-prompt assembly boundary; responses are never hidden
  or relabelled by heuristics.
- Subagent children now install the official model-selection assembly, so a
  keyword rule reaches the child's first request instead of falling back to
  the parent's creation-time options.

### Added

- `src/catalog.ts` (official directory + exact target resolution),
  `src/migration.ts` (pure V1→V2 migration + `toPersistedV2`),
  `src/media/types.ts` + `src/media/download.ts` (restricted downloads).
- 161 tests: 154 runtime/regression tests plus 7 public-release contract tests,
  including entry lifecycle contracts (no `registerAdapter` /
  `registerModelDiscovery` / `mr:` in the V2 runtime).
- Node 24-compatible pinned DNS lookups: when Node requests
  `lookup({ all: true })`, the downloader returns the complete set of already
  validated public addresses without weakening DNS-rebinding protection.

## [0.1.0] — 2026-08-14

Historical pre-V2 design record for `@dsh-external/dsh-model-router`. The public
Git history does not contain a recoverable 0.1.0 release tag; use this section
only to understand migration inputs, not as an installable release.

### Added

- **模型中枢 Settings page** (`settings.section` id `model-router`): master
  switch, case-sensitivity switch, multi-category channel management (seven
  tags, adapter, base URL, credential reference, model list, default
  reasoning effort, price tier), per-channel connection test, keyword
  routing rules, fallback selection, current-effective display, and
  "apply to this session".
- **`openai-chat` adapter**: dynamic registration/unregistration of channel
  provider routes (`mr:<id>`) on the harness `llm` seam; OpenAI-compatible
  SSE translation, stable `LlmError` codes, per-operation connection facts
  (live settings changes without restart).
- **Purpose isolation**: non-agent channels (`vision` / `image_gen` /
  `video_gen` / `audio` / `model_3d`) never enter agent routing candidates
  unless they also carry an `agent_*` tag.
- **`model_route` tool**: manual session model switch through the durable
  `request/header` write channel (the `/model` channel), candidates limited
  to agent-capable channels plus host providers (e.g. `deepseek-official`).
- **Keyword rule engine**: ordered first-match-wins, case-sensitivity
  switch, configured fallback; mounted on `agent/inbox/inserted`.
- **`image_gen` tool**: single synchronous `images/generations` call on an
  `image_gen` channel with Artifact delivery into the session workspace,
  plus a read-only Web preview route.
- **`async-job` adapter interface** (v2 reserved): submit/poll/cancel
  contract for video / 3D / audio generation.
- Web routes: `/_dsh/model-router/settings` (snapshot / save / test /
  apply), `/_dsh/model-router/artifacts` (read-only image preview).
- Red-light-first test suite covering config validation, adapter wire
  mapping, dynamic registration, isolation, rule engine, session write
  channel, tools, Web backend, and the release layout.

### Historical verification

- DeepSeek Harness `0.1.0-rc.6` (web profile), Node 24.
- Provider-specific live details were intentionally removed from the public
  changelog. Current verification belongs to the active release and CI.