Repository navigation
Releases
Releases · auggie246/dsh-llm-openai-codex
v0.7.0
Support DeepSeek Harness 0.2.0-rc.2; the whole 0.1.x harness series is no longer supported.
- All
@deepseek-ai/*peer dependencies retargeted from 0.1.7-rc.2 to 0.2.0-rc.2;@earendil-works/pi-aiwidened to ^0.87.1 (the version 0.2.0-rc.2 bundles). - No API changes in the packages this plugin uses — no runtime code changes.
- Tests no longer hardcode catalog model slugs (pi-ai 0.87 retired gpt-5.4); 80/80 pass.
v0.6.1
v0.6.0
v0.5.0
Added
- Live Codex model discovery: when credentials allow, the model picker adopts the backend's model manifest and refreshes it on a timer. A persisted cache keeps the discovered list across restarts, and the installed catalog always serves as the fallback.
Compatibility
- DeepSeek Harness
0.1.1-rc.2,0.1.2-rc.1, and0.1.5-rc.1or newer0.1.5-rc.x. dsh-llm-pi-ai 0.1.5 reads a new requiredmodelErrorsmap from every resolved provider profile and calls it unguarded on every turn; the route profile this plugin builds now carries it, so the first turn no longer dies on a 0.1.5 host. The 0.1.1 and 0.1.2 adapters never read the member. - Peer ranges widen to
^0.1.1-rc.2 || ^0.1.2-rc.1 || ^0.1.5-rc.1and@earendil-works/pi-aito^0.82.1 || ^0.84.2 || ^0.85.1. - The suite passes 76/76 against the 0.1.1-rc.2 baseline, the published 0.1.2-rc.1 train, and the published 0.1.5-rc.2 packages in both pi-ai layouts.
v0.4.1
Highlights
- The standard-readme README now ships to npm. v0.4.0 restructured the repository README to the standard-readme specification, but npm could not take it: the registry refuses a same-version republish, and the unpublished
0.4.0version number is blocked from republishing.0.4.1is the same code with the version bumped, so the npm package page now carries the restructured README. - Nothing else changed. The code is identical to
v0.4.0— harness0.1.1-rc.2/0.1.2-rc.1support with the legacy fallbacks from that release.
Install
From GitHub:
dsh plugin --profile web add github:auggie246/dsh-llm-openai-codexOr from npm:
dsh plugin --profile web add dsh-llm-openai-codexFull Changelog: v0.4.0...v0.4.1
v0.4.0
Highlights
- Works with harness 0.1.2-rc.1. The 0.1.2 release removed the
settingsNamespacehelper from@deepseek-ai/dsh-settings, so this plugin's named ESM import failed at module link time and took the whole host row down. The import is now a namespace import resolved throughsettingsNamespaceOf: on 0.1.1 the real validator still runs, on 0.1.2+ the validated lowercase-hyphenated name passes through untouched. - Mixed-version adapter fallback. Harness 0.1.2 calls
adapter.imageRequestPricing()unguarded from the token meter, and the method only ships with the 0.1.2 adapter base class.applynow attaches the same no-pricing default when the installed adapter predates it, so a tree that still resolves 0.1.1 packages keeps serving turns instead of dying withadapter.imageRequestPricing is not a functionon every priced request. - Nothing else changed on 0.1.1. The verified-compatible surface stayed identical across both lines:
LlmError/RetryPolicySchema/resolveRetryPolicy, the six-method adapter contract, thePiAiAdapteroptions and profile shape, the Typert manifests andremote.codexAuthmount, thesettings.plugin.itemslot, and the lazy-CJS client wrapper. pi-ai 0.84.2 (the version 0.1.2 pins) keeps theopenai-codexcatalog provider, the{credential}resolve contract, and the request-levelapiKeyoverride.
Changes
- Peer ranges widened to
^0.1.1-rc.2 || ^0.1.2-rc.1for the@deepseek-ai/dsh-*packages, and to^0.82.1 || ^0.84.2for@earendil-works/pi-ai. On a 0.1.2 profile, pnpm may print a peer warning fordsh-client-runtime; it is expected and harmless — the package no longer exists upstream and the client module loader skips missing inject rows. - Two regression tests cover the fallbacks. The suite passes against the 0.1.1-rc.2 baseline and against the published 0.1.2-rc.1 packages.
Install
From GitHub:
dsh plugin --profile web add github:auggie246/dsh-llm-openai-codexOr from npm, once published there:
dsh plugin --profile web add dsh-llm-openai-codexFull Changelog: v0.3.0...v0.4.0
Note
This release was re-tagged after a README-only restructure (now following the standard-readme specification). npm keeps the previously published 0.4.0 tarball: the registry refuses a same-version republish, so the publish step skips when the version already exists. The README here and on the tag is current; the npm package page updates with the next version.
v0.3.0
Highlights
- Cancel a login in progress. A browser login whose tab closed no longer dead-ends the card. Cancel this login releases the callback server (or the device poll loop) at once, so a new login can start without waiting out the timeout. Cancelling while idle is a safe no-op.
- Reopen the authorization page. Browser pending status now carries the authorization URL. The card's "Open the authorization page again" link survives a page reload.
- Login countdown. Both login kinds show their remaining time. A browser attempt waits 10 minutes at most.
- Default retry policy. An unconfigured route now retries transient provider errors twice (~0.5s, then ~1s). The Codex backend answers some transient failures with an empty body, and over HTTP/2 there is no status text either, so pi-ai raises the unclassified
PI_AI_ERROR. The harness default retry set excludes that code, so one such response killed the turn. Deliberate non-retryables (AUTH,INVALID_REQUEST) stay out. An explicitconfig.retryPolicyreplaces the default verbatim.
Changes
- Added
codexAuth/cancelLoginto the host gateway, the client facade, and the codecs. The auth Remote now exposes six direct endpoints. - Browser pending status carries
url. Device pending keepsuserCodeandverificationUri, and never carries a URL. - New troubleshooting rows in the README: the stuck "Waiting for approval" state, and
Request failedon a Codex turn.
Install
From GitHub:
dsh plugin --profile web add github:auggie246/dsh-llm-openai-codexOr from npm, once published there:
dsh plugin --profile web add dsh-llm-openai-codexFull Changelog: 89d9ddd...v0.3.0