Upgrade @earendil-works/pi-ai to ^0.87.1 for the new model catalog — and fix model discovery to ask live endpoints (incl. defaults)

[GnaneshKunal]

Summary

Proposing a bump of @earendil-works/pi-ai from ^0.85.1 to ^0.87.1 in packages/llm/llm-pi-ai, so the new model catalog lands in the built-in catalog — pnpm-workspace.yaml itself notes that "fresh pi-ai releases carry the model catalog updates that are the whole point of bumping it". 0.87.1 pulls in the new DeepSeek model IDs (deepseek-flash, V4.1-era defaults in place of deepseek-v4-flash), the mistral-conversations API, and the latest compat metadata.

Since MRs/PRs aren't accepted here, I've put the complete upgrade together locally and am attaching the working patch so it can be picked up and landed directly.

What the patch contains (21 files, +376 / −445)

1. Dependency bump — @earendil-works/pi-ai ^0.85.1 → ^0.87.1 in packages/llm/llm-pi-ai/package.json, with pnpm-lock.yaml and pnpm-workspace.yaml (minimumReleaseAgeExclude now names @earendil-works/pi-ai@0.87.1 / @earendil-works/pi-telemetry@0.87.1) refreshed accordingly.
2. New models pulled in — pi-ai's updated catalog renames deepseek-v4-flash → deepseek-flash; tests, fixtures, and inline snapshots follow (the real-API e2e now targets "V4.1 Flash defaults"). This is the main reason for the bump.
3. Upstream API drift absorbed in packages/llm/llm-pi-ai/src/catalog.ts:
   • new MistralConversationsCompat type with its own drift gate (MISTRAL_COMPAT_GATE) and a mistral-conversations entry in COMPAT_GATES;
   • compat gates updated for 0.87's surface: deferredToolsMode removed from OpenAICompletionsCompat, new supportsMidConvoSystemMessages / supportsMidConvoToolAdditions / supportsMidConvoToolChanges fields gated as withhold, and supportsToolReferences → sessionAffinityFormat on the Anthropic gate.
4. Typing — packages/llm/llm-pi-ai/src/replay.ts now types tool-call arguments as pi-ai's JsonObject.
5. The local pnpm patch is re-based and renamed — patches/@earendil-works__pi-ai@0.85.1.patch → patches/@earendil-works__pi-ai@0.87.1.patch. It still removes the per-delta full re-parse of the accumulated tool-call argument JSON in every stream adapter (upstream earendil-works/pi#9265); unpatched, a multi-megabyte argument stream costs O(n²) CPU on the event loop and stalls every session in the process. Please keep re-applying it on each pi-ai bump (or retire it once the fix lands upstream).
6. Docs — THIRD_PARTY_NOTICES.md and packages/llm/llm-pi-ai/README.md / README.zh.md (plus the i18n pairing hashes) point at the renamed patch file.

Request

Please land the ^0.87.1 bump so the new models (new DeepSeek IDs, Mistral conversations support) reach the built-in catalog. The attached patch is the complete diff against base commit 00102833dfaee1da9f48a3a8eae9d34005a75218 (00102833df, "Merge pull request #4978 from deepseek-harness/worktree/release-dsh-0.1.7-alpha.2") — apply it there with git apply staged-changes.patch.

Attachment

📎 staged-changes.patch — the full working diff (raw)

[GnaneshKunal]

Patch provenance, for clarity: staged-changes.patch is a diff against base commit 00102833dfaee1da9f48a3a8eae9d34005a75218 — 00102833df ("Merge pull request #4978 from deepseek-harness/worktree/release-dsh-0.1.7-alpha.2"), the tip of master at the time the diff was captured (it is git diff --cached from a clean checkout at that exact commit).

So rather than applying it against whatever master happens to be, pin the base commit:

git checkout 00102833dfaee1da9f48a3a8eae9d34005a75218
git apply staged-changes.patch

If master has moved by the time this lands, the diff should be re-captured/rebased onto the then-current tip.

[randomix777]

There is a second issue in the same area that should be fixed alongside the pi-ai catalog bump.

Reproduction

1. Configure llm-pi-ai.providers.opencode-go with apiKeyEnv: OPENCODE_GO_API_KEY and a valid official key.
2. Open Settings -> Models and use Fetch available models.
3. https://opencode.ai/zen/go/v1/models returns the account-enabled catalog, including deepseek-v4.1-flash, mimo-v2.6-flash, and space-bunny-free (the exact count is account-dependent).
4. DSH does not issue that request and shows only the bundled pi-ai catalog instead.

Root cause

packages/llm/llm-pi-ai/src/discovery.ts returns catalogModels(request.provider) whenever request.provider is present, before it examines request.baseURL or resolves the stored credential. That makes catalog-backed OpenCode routes short-circuit the endpoint path, so the official key is never used for discovery.

Suggested fix and regression test

For an OpenCode route with a non-empty baseURL, probe /models first; retain the bundled catalog as the no-endpoint fallback. Add a test with a local listing server proving that a catalog-backed route with baseURL performs GET /v1/models, forwards the stored key, and returns the endpoint entries. Keep the existing no-network behavior for requests without a base URL.

This is separate from refreshing the generated catalog itself: after the catalog update, the discovery path still needs to use the live endpoint.

[randomix777]

Follow-up for OpenRouter: the same discovery short-circuit affects this route. https://openrouter.ai/api/v1/models currently returns 459 models, including stealth/space-bunny-alpha; the bundled pi-ai OpenRouter catalog in this checkout has 333 models and does not contain it. Also, an explicit llm-pi-ai.providers.openrouter.models list replaces the inherited catalog, so a newly listed model must either be adopted from live discovery or added to that list. The endpoint-backed discovery path should therefore cover OpenRouter as well as OpenCode, with a regression test for the exact model id.

[GnaneshKunal]

Thanks for the feedback. Lemme check this. Thanks

…

________________________________ From: randomix777 ***@***.***> Sent: Thursday, 24 September 2026 07:53:19 To: deepseek-ai/deepseek-harness ***@***.***> Cc: Gnanesh ***@***.***>; Author ***@***.***> Subject: Re: [deepseek-ai/deepseek-harness] Proposal: upgrade @earendil-works/pi-ai to ^0.87.1 so the new model catalog lands (Discussion #7660) Follow-up for OpenRouter: the same discovery short-circuit affects this route. https://openrouter.ai/api/v1/models currently returns 459 models, including stealth/space-bunny-alpha; the bundled pi-ai OpenRouter catalog in this checkout has 333 models and does not contain it. Also, an explicit llm-pi-ai.providers.openrouter.models list replaces the inherited catalog, so a newly listed model must either be adopted from live discovery or added to that list. The endpoint-backed discovery path should therefore cover OpenRouter as well as OpenCode, with a regression test for the exact model id. — Reply to this email directly, view it on GitHub<#7660?email_source=notifications&email_token=AE2GLX5QGUZ62GOJIOZHQIL5QSARPA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOBVG42DQNZWUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVRTG633UMVZF6Y3MNFRWW#discussioncomment-18574876>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AE2GLX6QNTQOKVHUHU3LBLL5QSARPAVCNFSNUABJKJSXA33TNF2G64TZHMYTGMZTGA3DKMBZGE5UI2LTMN2XG43JN5XDWMJQHA3TINBTGSQXMAQ>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[GnaneshKunal]

Follow-up: the discovery short-circuit is fixed — including the no-baseURL case. The patch below applies on top of staged-changes.patch (same base commit 00102833df…).

What changed since the original post — it covers both reports above, plus the case you described as "no base url, it'll take the default one":

draft	before	now
route with baseURL	bundled snapshot; endpoint never asked	probes the named endpoint with the stored key → the live listing
route, no baseURL	bundled snapshot	probes the route's default endpoint — OpenRouter https://openrouter.ai/api/v1, OpenCode Go https://opencode.ai/zen/go/v1, resolved from pi-ai's provider record and its catalog models
default endpoint unreachable	—	bundled catalog with a labeled fallback ({ reason: 'endpoint-unreachable', endpoint }), rendered as a warning in the picker
endpoint answers 401 / broken reply, or a named baseURL fails	error	unchanged: loud error — a stale list never reads as a live one
a form's own field-defaults read	shared the fetch call	explicit answer: 'catalog' — offline, no network

The regression tests pin exactly what you asked for: a catalog-backed route with a baseURL performs GET /v1/models forwarding the stored key and returns the endpoint entries (deepseek-v4.1-flash, mimo-v2.6-flash, space-bunny-free); stealth/space-bunny-alpha is discoverable through live discovery even when an explicit providers.<route>.models list replaces the inherited catalog; and the OpenCode Go / OpenRouter default-endpoint probes resolve to https://opencode.ai/zen/go/v1/models and https://openrouter.ai/api/v1/models.

Also verified in the web UI (Models → openrouter → Fetch available models): with Base URL empty the picker now returns the live ~459-entry OpenRouter list including stealth/space-bunny-alpha (the bundled snapshot has 386 and not it), and with the network blocked it shows the labeled bundled-catalog fallback naming the attempted endpoint.

Alongside this, the discovery reply becomes LlmModelDiscoveryResult (models plus the optional fallback) and the request gains answer; the LlmModelDiscoveryRequest contract, the llm-streaming type pastes (EN/中文), the llm-pi-ai READMEs, and the typed Remote catalog all follow in the same patch.

📎 discovery-fix.patch (raw) — 29 files, +591/−166. Apply order from the base commit:

git checkout 00102833dfaee1da9f48a3a8eae9d34005a75218
git apply staged-changes.patch   # the pi-ai ^0.87.1 upgrade (attached above)
git apply discovery-fix.patch    # this follow-up