Model catalog lagging behind pi-ai releases: glm-5.3 unavailable on zai-coding-cn

[explosivee]

Environment

dsh 0.1.1-rc.2 (latest as of today), installed globally via npm
Provider route: zai-coding-cn (Zhipu coding plan)
Problem GLM-5.3 (released by Zhipu in mid-August) cannot be selected in the model picker for the zai-coding-cn route. The installed catalog tops out at glm-5.2.

Root cause @deepseek-ai/dsh-llm-pi-ai declares "@earendil-works/pi-ai": "^0.82.1". Under 0.x semver rules this range resolves to 0.82.x only, so even the newest dsh ships pi-ai 0.82.1 — whose zai-coding-cn catalog predates GLM-5.3. pi-ai 0.84.2 (published Aug 14) already includes glm-5.3 and glm-5.2-highspeed for this route in its builtin catalog.

Request Bump the dependency to ^0.84.2 (or track pi-ai releases more closely), so new models reach dsh users without manual config.

Workaround (for anyone hitting this): add the model explicitly under llm-pi-ai.providers.zai-coding-cn.models in ~/.dsh/settings.yaml, e.g. glm-5.3 with contextWindow: 1000000, maxTokens: 131072, compat: { thinkingFormat: zai, supportsReasoningEffort: false } — values taken from the pi-ai 0.84.2 catalog entry.

[weijiafu14]

Same root cause as #2330 (the model catalog ships inside the @earendil-works/pi-ai package; the rc pin holds an older minor, so mid-August models like GLM-5.3 simply aren't in the installed data files). Two additions:

A validated way to get GLM-5.3 today: declare the model explicitly instead of waiting for the catalog — a declared route in settings.yaml's llm-pi-ai: section (model id, contextWindow, maxTokens, plus thinkingLevelMap/compat if needed) bypasses the built-in catalog entirely. Working declaration recipes: examples/custom-gateways — our validation harness drives real endpoints through exactly this shape, with recorded-wire assertions that declared values actually go out.

The systemic fix: this is the most user-visible payoff of the pi-ai version-boundary migration proposed in #3752 — we posted cross-version behavioral A/B data there (0.82.1 / 0.84.1 / 0.84.4, accumulator/transform/retry) supporting that upgrade path being made regular and regression-gated.

Scope note: an explicit declaration gives you the route and capacity semantics; whether your Zhipu coding-plan credential unlocks GLM-5.3 on their side is between your subscription and their gateway.

[fan56]

This is exactly what dsh-model-sync fixes: the route stays in step with pi.dev without waiting for a pi-ai release or a dsh upgrade.

Measured today: pi.dev's zai-coding-cn listing already carries glm-5.3, glm-5.3-flash and glm-5.3-highspeed. Install, and the first sync (5s after startup, or /model-sync on demand) lands them in providers.zai-coding-cn.models through the official settings.mutate seam — zero patches — then re-syncs every 4 hours:

npm i @aiwayds/dsh-model-sync
dsh plugin add @aiwayds/dsh-model-sync

0.1.6 notes: listing capacity values are sanity-gated (a maxTokens that merely echoes the contextWindow is stripped with a DEGRADED report line instead of tripping the "output token limit" family, see #1166), and your own modelOverrides (think levels, narrowed context windows) survive every sync round. Subscription/entitlement for glm-5.3 still follows your zai credentials — the sync only fixes "the catalog carries it and you can select it".