# Changelog

## 1.2.11

### Fixed
- Free models can call tools again. The free-tier gate authenticates the client
  through headers alone (`User-Agent`, `x-opencode-client`,
  `x-opencode-session` + `x-opencode-session-id`); once the duplicate session
  header from 1.2.10 is present it accepts a normal pi request. The body-shaping
  layer (title-generator marker, flattened history, stripped `tools`) and the
  text-tool protocol that replaced native tools are therefore unnecessary, and
  they were what broke tool calls: `big-pickle` (zero-cost but without a `-free`
  suffix) was never converted back into a tool call, `mimo-v2.6-flash-free`
  emitted several calls inside one fenced block and lost all of them, and
  weaker models invented their own fences. Both are removed; free models now
  send and receive native tools, verified end-to-end (read + bash + follow-up
  turn) on `big-pickle`, `muse-spark-1.3-contributor-free`,
  `mimo-v2.6-flash-free`, `nemotron-3-ultra-free` and
  `nemotron-3.5-lightning-free`.
- `GATE_UA` tracks opencode 2.0.32. The gate also re-checked its content rules:
  free requests now need `stream: true` and at least one tool definition, and
  the versionless `opencode/cli` user agent is rejected; pi's own request
  satisfies all of it, so the body still passes through untouched.
- Cached catalogs cannot pin a stale protocol any more: `healZenRoute`
  re-derives the free-tier route from the model id on every read, in
  `getModels()`/`getAllModels()` — the accessors every request path uses.
  A `models-store.json` written by an older version (which survives upgrades
  because pi's post-login refresh is offline) no longer causes
  `400 ModelProtocolUnsupported`.

## 1.2.10

### Fixed
- Free models were unusable again: every request returned `403 FreeTierError`
  because the gate now validates a second header, `x-opencode-session-id`, that
  `freeTierHeaders` did not send (verified against opencode 2.0.29).
- Free models outside the `gpt` / `grok` / `muse` families (`cat-jam-free`,
  `deepseek-v4.1-flash-free`, `mimo-v2.6-flash-free`, ...) returned
  `400 ModelProtocolUnsupported`. `modelFromZen` forced every free model onto
  OpenAI Responses, but the gateway serves each one on a fixed protocol:
  Responses for those three families, Chat Completions for the rest.
- Chat Completions free models rejected pi's own request shape with
  `403 FreeTierError`. The gate only accepts a title-generator-shaped
  conversation, so a run of three or more `user` messages (what pi sends for a
  system-reminder-annotated prompt) and native `assistant.tool_calls` / `tool`
  turns (what pi sends after a tool run) were both refused. The Completions path
  now flattens the history into one transcript, like the Responses path did, and
  the marker prompt text was resynced with the live one.

## 1.2.9

### Added
- Bounded request retries (parity #6): transient network/server failures
  retry up to twice with backoff, and 400s naming a rejected option
  (e.g. `temperature`) drop the option and retry — only before any response
  event reaches the consumer, so partial streams are never re-sent.
- `/opencode-zen` command (`status`, `switch-org`): switch workspace without
  re-login. Session-scoped (the extension API has no credential-write path),
  catalog refresh, request org header, and model filtering follow the switch.

### Fixed
- Console `test*` smoke-test ids and `jev*` decision-protocol ids no longer
  leak into the model picker.
- Failed catalog discovery and cache-only startup keep serving the stored
  catalog instead of wiping it (locked in by regression tests).

## 1.2.8

### Fixed
- Claude models that only accept adaptive thinking (`claude-haiku-5-5`,
  `claude-sonnet-5/5-5`, `claude-opus-5/5-5`, late 4.x sonnet/opus,
  `claude-fable-5*`) no longer fail on Zen with
  `"thinking.type.enabled" is not supported for this model` (Issue #10):
  `modelFromZen` now sets `compat.forceAdaptiveThinking` for them, so pi-ai
  sends `thinking.type: adaptive` with `output_config.effort`.

### Changed
- Dev dependencies track pi 1.2.0 (`@earendil-works/pi-ai`,
  `@earendil-works/pi-coding-agent` `^0.99.1` → `^1.2.0`).

## 1.2.7

### Fixed
- Text-protocol tool calls no longer fail validation with empty arguments:
  the prompt example is a filled command and the rules require every
  required parameter.

### Changed
- Free-tier identity UA tracks the current CLI release (2.0.28).

## 1.2.6

### Changed
- Free-tier (zero-cost) models work through the v2 content gate again. The
  gateway now checks the request body: a leading developer message must carry
  OpenCode's own title-generator prompt verbatim, native `tools` declarations
  are rejected, and identity headers must name the official client
  (`opencode/latest/2.0.27/cli`, `x-opencode-client: cli`, one
  `x-opencode-session` header). The extension injects the marker, collapses
  history to `[marker, system, user]`, and strips tool declarations on free
  models only; paid requests are never modified.
- Tool calling on free models runs as a text protocol: the system prompt
  teaches fenced ```tool JSON blocks, `message_end` converts them back into
  real pi tool calls (parallel blocks supported), and results return through
  the transcript. Verified live: chat, read, bash on muse-spark free.

### Fixed
- Free-model detection no longer depends on the catalog refresh having run:
  zero-cost cost metadata or a `-free` id suffix both qualify.

## 1.2.5

### Added
- Free-tier (zero-cost) models are usable again. The Zen gateway rejects them
  with `FreeTierError` ("OpenCode's free tier can only be used from within
  OpenCode") unless the request looks like an OpenCode-client turn. Probed
  against the live gateway, the check is exactly three conditions — token,
  TLS fingerprint, HTTP version and source IP are *not* among them:
  1. `User-Agent` shaped `opencode/(latest/)?<version>/…` — `opencode/cli` is
     rejected. Any semver-ish version passes.
  2. an `x-opencode-session` header (`x-session-affinity`/`x-session-id` also
     satisfy it, `x-opencode-session-id` alone does not) whose value matches
     the official `ses_<12 hex><14 alnum>` session-id shape.
  3. a streamed body (`stream: true`) whose `tools` array carries a read tool
     plus a command-execution tool — verified pairs `{read, shell}` and
     `{read, bash}`; single tools and other pairs are rejected.
  Pi's agent already advertises `read` + `bash`, so agent turns only need the
  header pair; tool-less completions (`streamSimple`, title generation, …)
  additionally get both stubs injected via the adapter's `onPayload` hook.
  Stubs follow the route's own tool format — `openai-completions`,
  `anthropic-messages` (flat `{name, description, input_schema}`) and
  `openai-responses` (flat function tools) are all covered. On completions
  the stubs are muted with `tool_choice: "none"`; anthropic and responses
  have no such value so the stubs stay callable in theory (in practice
  tool-less completions rarely emit calls). Paid models are never modified.
  See `src/free-tier.ts`.
- Zero-cost models are no longer filtered out of the catalog during
  `/api/config` — they are gated per-request instead of hidden.

### Fixed
- The console now attaches `options.apiKey: "{env:OPENCODE_CONSOLE_TOKEN}"`
  and an org header to the managed `opencode` entry. The BYOK check treated
  any `options.apiKey` as "bring your own key" and dropped the whole entry —
  the catalog came back empty ("logged in, no models"). An entry is now only
  treated as BYOK when its endpoint is off the Zen gateway.
- "Logged in but no models": a `/api/config` result that yielded no usable
  entries stored `allowedModels: []`, and `filterModels` then hid the entire
  catalog — including the offline baseline. An empty or missing list now means
  "no workspace restrictions"; the catalog is never wiped.
- Models with a `provider` override are no longer dropped (Issue #2). They
  route to the matching Zen gateway wire protocol and the provider registers
  the matching API implementations (`anthropic-messages`,
  `google-generative-ai`, `openai-responses` alongside `openai-completions`):
  - `@ai-sdk/anthropic` → `anthropic-messages`, with the console's trailing
    `/v1` stripped — the anthropic SDK appends `/v1/messages` itself
  - `@ai-sdk/google` → `google-generative-ai`
  - `@ai-sdk/openai` → `openai-responses` for the gpt/grok/muse families,
    `openai-completions` otherwise
  - unknown npm packages (vertex, bedrock, openrouter, …) stay excluded —
    they cannot route through the Zen gateway
- Console switches are respected: `blacklist` entries and models with
  `status: "deprecated"` are excluded (opencode itself treats
  `status !== "deprecated"` as enabled).
- `/api/config` entries beyond `provider.opencode` are read too: entries with
  a Zen-routable `npm`/`api` contribute their models; BYOK entries (own
  `options.baseURL`/`apiKey`) and entries with no routable protocol are
  skipped.
- A failing `/api/config` fetch during login no longer aborts the login —
  `allowedModels` stays unset and the next catalog refresh retries.

## 1.2.4

### Fixed
- Exclude zero-cost (free-tier) models from the catalog. Zen answers them
  with `FreeTierError` outside the opencode client even when the workspace
  whitelist includes them — listing them only produced dead ends.

## 1.2.3

### Changed
- Verified against pi 0.99.1 and opencode v2.0.20; devDependencies bumped.
  No code changes.

### Changed
- Rewrote README (hook-first structure) and added Vietnamese translation
  (`README.vi.md`) with a language switcher. No code changes.

### Fixed
- Console API calls use the `x-org-id` header. The previous `x-opencode-org-id`
  header is only valid on the inference endpoint — on `/api/config` the console
  answers `400 OrgRequired`. Shared workspace tokens (with a server-side active
  org) masked this; fresh device-flow tokens exposed it.
- `fetchZenModels` failure now hints at re-login (`/login opencode-zen`) on
  token rejection instead of a bare HTTP status.

### Added
- Security notes in README (no token/workspace-ID logging, leak recovery).

## 1.1.1

### Changed
- Versionless `User-Agent` (`opencode/cli`). The npm `latest` dist-tag tracks the
  v1 line which the console rejects, and pinning a local version drifts — the
  console accepts the versionless form, verified live.

## 1.1.0

### Added
- Dynamic model catalog via `fetchModels`: the workspace-allowed list
  (not disabled, on the org whitelist, no custom provider override) is pulled
  from `/api/config` on every pi catalog refresh, with context limits and
  prices mapped by `modelFromZen`. The static `MODELS` remains as the offline
  baseline.

## 1.0.1

### Added
- `repository` / `homepage` metadata so the pi package gallery links GitHub.

## 1.0.0

### Added
- `opencode-zen` provider for pi: console device-flow OAuth login, no API key
  management. Workspace members approve once in the browser.
- Workspace selection in the pi TUI after approval; requests carry the
  matching org header.
- Workspace-allowed model check after login and on every token refresh;
  `filterModels` narrows `/model` to what the workspace permits.
- Dynamic opencode `User-Agent` (later replaced, see 1.1.1).
- Paid-tier prices from the console workspace config
  (glm-5.3-flash, deepseek-v4.1-flash).