v0.2.0

Note

如果你是从 v0.1.x 升上来的：那两个版本完全不可用（每个请求 403 / 每轮 PI_AI_ERROR），
没有需要迁移的配置，直接换到这一版即可。

Fixed

• Every chat turn failed with Cannot read properties of undefined (reading 'apiKey') (PI_AI_ERROR), while the availability probe kept
  reporting models as reachable. The provider was built without an auth
  plane, so pi-ai dereferenced provider.auth.apiKey on the first request and
  threw before anything reached the network — which is exactly why the probe,
  which speaks HTTP directly rather than through pi-ai, still looked healthy.

  The provider now carries a placeholder api-key method that satisfies pi-ai's
  resolution path, and the adapter's credential seam returns undefined rather
  than '' so the request-level override is omitted entirely. The lane remains
  anonymous: no credential is sent, and Authorization is suppressed explicitly
  rather than filled with a placeholder value.

• The catalog re-downloaded 5.69 MB on every catalogue build. models.dev
  answers with a single document covering every provider it knows, of which this
  plugin reads one section. It is now cached on disk with its own TTL, and the
  previously declared-but-unused METADATA_TTL_MS (24 hours) is what actually
  governs it — the on-sale list keeps the shorter 5-minute TTL, because a new
  model should appear in the picker promptly.

  Startup no longer waits for that download, an offline start still gets a full
  list, and the refresh button still forces a real round trip instead of being
  answered from cache.

Changed

• The upstream endpoint is https://opencode.ai/inference/v1; the previously
  used /zen/v1 path is retired. The request shape that the lane accepts is
  now centralised in lib/upstream.js.

Notes

• Verified against the live lane with the plugin's own code path
  (ZenCatalogService.probeAll()): 8 of 12 advertised free models answer,
  including multi-turn recall. The remaining failures are upstream-side and are
  expected — cat-jam-free and muse-spark-1.3-contributor-free are
  region-blocked (403 RegionError), jev-1.13-free is
  400 ModelProtocolUnsupported.
• Cache location is ~/.cache/dsh-zen-adapter/catalog.db, following the same
  convention as dsh-free-search. Delete that directory to clear it; nothing
  else is left behind. ZEN_CACHE_DIR relocates it.
• The free lane is undocumented upstream behaviour and may change without
  notice. If every model starts returning 403, the gate has moved — the
  availability button is the early warning for that.

v0.1.1

Warning

此版本同样不可用。 403 已经修好，但对话每一轮都会失败
（PI_AI_ERROR：Cannot read properties of undefined (reading 'apiKey')）——
provider 缺了 auth plane，pi-ai 在发请求前就抛错了。
请直接安装 v0.2.0。

Fixed

• Free lane rejected every request with 403 FreeTierError. The session id
  sent upstream was well-shaped but derived wrongly: its 12 hex characters came
  from a SHA-256 digest, while the lane decodes them as a bit-inverted wall
  clock and refuses anything that is not a plausible timestamp. Every request
  failed, including ones that previously worked, while the catalog stayed
  readable — so the plugin looked healthy and simply could not talk.

  The clock part is now produced with upstream's own arithmetic
  (BigInt(timestamp) * 0x1000n + counter, then ~, taking 6 bytes), in a new
  lib/session.js shared by the adapter and the availability probe.

  Two consequences worth recording, because both mislead:

  • A shape match proves nothing. /^ses_[0-9a-f]{12}[0-9A-Za-z]{14}$/
    accepts the broken ids, so no shape assertion could have caught this.
  • ffe is not a token. Ids that end in ffe happen to work when
    counter is 1, because ~1 ends in ffe; ~2 ends in ffd and ~300
    in ed3. Hard-coding the suffix still yields a digest, not a clock.

• The availability probe generated its own session id — a second copy of the
  same broken derivation. The probe button would have reported 403 for every
  model while chat worked, which reads as "upstream is down" rather than "the
  probe is malformed". Both call sites now share one implementation.

Notes

• Verified against the live lane with the plugin's own code path
  (ZenCatalogService.probeAll()): 5 of 10 advertised free models answer.
  The remaining failures are upstream-side and are expected —
  ling-3.1-flash-free and step-5-preview-free throttle (429),
  jev-1.13-free is SystemOne-only (400 ModelProtocolUnsupported),
  muse-spark-1.3-contributor-free is region-blocked (403 RegionError).
• A wrong recipe for this outage circulated publicly (switch to
  /inference/openai, drop Authorization, add a shell tool). It does not
  work: the gate is the session id, not the endpoint, the auth header, or the
  tool names.

v0.1.0

Warning

此版本完全不可用。 每一个请求都会被上游拒绝（403 FreeTierError）——
会话 id 的时钟推导方式不对，上游只接受看起来像时间戳的值。
请直接安装 v0.2.0。
Initial release.

Added

• Native DSH provider zen-free — OpenCode Zen's free models registered
  through llm.registerAdapter(), so they appear in the model picker as the
  OpenCode Zen (free lane) group. No signup, no API key, no extra process.
• Live model catalog — the free roster is discovered at runtime from
  opencode.ai/zen/v1/models intersected with models.dev/api.json, rather
  than hard-coded. Refreshed in the background every 30 minutes, and a refresh
  is announced to the client so the picker updates without a restart.
• Free determination from pricing — a model is offered only when
  models.dev reports input and output cost of 0. Deprecated or paid models
  are refused even when their id ends in -free, because upstream keeps
  delisted ids on sale for a while. The -free suffix is a fallback used only
  when metadata is missing altogether.
• Real context limits — contextWindow and maxTokens come from
  models.dev per model instead of a fixed constant, so a 1M-token model is not
  understated and a 200K model is not overstated.
• Plugin detail page — a panel under Settings → Plugins → dsh-zen-adapter
  with 获取模型 (re-pull the catalog) and 检测可用性 (send a real request
  per model to see which ones answer right now), plus a table of model id,
  context size, output limit, availability status and the reason behind it.
• Zero dependencies — @deepseek-ai/* and @earendil-works/pi-ai are
  declared as optional peerDependencies and resolved through the host's
  module tree, so exactly one pi-ai generation exists on the machine and the
  plugin can be installed alongside other provider plugins.

Notes

• Upstream availability is per-IP quota'd, and some models carry regional
  restrictions. A model being listed does not mean it will answer: 429,
  403 RegionError and 5xx are upstream-side behaviour this plugin cannot
  fix. Use 检测可用性 to see current status.
• PROVIDER_ID is deliberately zen-free rather than zen, because
  llm.registerAdapter() is all-or-nothing and a second claimant of an
  existing provider id throws DUPLICATE_ADAPTER.