feat(poe): default to the openai responses api endpoint by kovsu · Pull Request #14144 · CherryHQ/cherry-studio

kovsu · 2026-04-09T09:29:43Z

Branch strategy

Active development targets main.

What this PR does

Before this PR:

Poe was a chat-completions-only registry entry (plus the Anthropic-compatible mirror for official Claude bots). On chat-completions, reasoning is fail-closed by design and only works for the model families with audited per-model extra_body contracts (GPT-5.4, Gemini 3.1 Pro, Claude bots); every other bot has no reasoning control, and built-in web search rides the Poe-specific extra_body.web_search parameter.

After this PR:

Poe's registry entry declares the OpenAI Responses API endpoint (https://api.poe.com/v1/, which Poe now serves natively) and makes it the default chat endpoint:

Reasoning: the standard openai-responses reasoning format applies to all bots — no per-model contract needed on this path.
Web search: the standard Responses web_search tool via the existing server-tools config.
Chat-completions stays as a fallback endpoint and keeps its current behavior unchanged (fail-closed wire + the per-model extra_body contracts); the anthropic-messages endpoint for official Claude bots is untouched.
Poe website metadata updated to the current URLs (creator.poe.com/docs, poe.com/api/models).
A registry contract test locks the routing: default endpoint is openai-responses with the standard reasoning format, chat-completions stays fail-closed (provider-reasoning-contracts.test.ts).

Fixes #

Why we need it and why it was done in this way

The following tradeoffs were made:

Additive endpoint switch, not a removal. Earlier iterations of this branch deleted the hand-rolled chat-side Poe special-casing in src/main/ai/; since then main rebuilt that logic as data-driven registry wire contracts (feat(ai-reasoning): data-driven reasoning controls from the provider registry #17034) and server-tools (feat(server-tools): add provider-native search capabilities #17322), so this PR keeps them as the chat-endpoint fallback and only adds the Responses route on top. Nothing regresses for users pinned to chat-completions, and the PR no longer touches src/main/ai/ at all.
The new default reaches existing installs too: provider rows are deliberately sparse ([Feature]: Reconcile provider-registry connection-config changes into already-persisted user_provider rows #17096) — seeded and v1-migrated rows keep a NULL default endpoint and delegate to the current registry preset, so the Responses default applies to fresh and existing installs alike. An explicit user endpoint choice always wins. Recorded as a notice in the breaking-changes log.
Claude bots keep anthropic-messages-first routing — Poe's Anthropic-compatible endpoint is already the best path for them.

The following alternatives were considered:

3-way model-ID routing (Claude → Anthropic SDK, GPT/O-series → Responses, rest → Chat Completions) — an earlier iteration; rejected as unnecessary once Poe shipped native Responses support.
Pinning existing rows to the legacy chat-completions endpoint (one-time data migration + migrator special-case) — rejected: it fights the sparse-row delegation design ([Feature]: Reconcile provider-registry connection-config changes into already-persisted user_provider rows #17096, which has a philosophy test asserting registry defaults flow to inherited rows), and the Responses path is the better default for every bot family.
Deleting the chat-completions extra_body reasoning/web-search paths (the original shape of this PR) — rejected after the data-driven reasoning refactor landed on main; they now serve as the fallback endpoint's audited contracts.

Links to places where the discussion took place:

Breaking changes

Poe's default chat endpoint moves to openai-responses for fresh and existing installs (sparse rows inherit the registry default; an explicit endpoint choice is preserved). Plain chat, reasoning and web search keep working — reasoning coverage widens from contract-listed families to all bots. Users who want the legacy endpoint can select openai-chat-completions explicitly in Settings → Model Providers → Poe. Recorded (severity: notice) in v2-refactor-temp/docs/breaking-changes/2026-07-13-poe-responses-endpoint-required.md.

Special notes for your reviewer

packages/provider-registry/data/*.json are regenerated via pnpm --filter @cherrystudio/provider-registry generate; the diff includes expected upstream (models.dev/OpenRouter) metadata drift per packages/provider-registry/CLAUDE.md. Two pieces of drift needed source fixes (each its own commit): the new ling-3-0-flash-fin model ships reasoning controls, so the bailing creator gains a reasoning-membership rule; and models.dev de-listed Fireworks' accounts/fireworks/routers/* kimi SKUs, so their overrides pin apiModelId/name to stay resolvable as standalone rows.
Verified: pnpm test:provider-registry (346/346 incl. catalog-source-sync and the new Poe contract test) and pnpm --filter @cherrystudio/provider-registry compat:check (compatible with frozen registry schema v1).

Checklist

This checklist is not enforcing, but it's a reminder of items that could be relevant to every PR.
Approvers are expected to review this list.

Branch: This PR targets main
PR: The PR description is expressive enough and will help future contributors
Code: Write code that humans can understand and Keep it simple
Refactor: You have left the code cleaner than you found it (Boy Scout Rule)
Upgrade: Impact of this change on upgrade flows was considered and addressed if required
Documentation: A user-guide update was considered and is present (link) or not required. Check this only when the PR introduces or changes a user-facing feature or behavior.
Self-review: I have reviewed my own code (e.g., via /gh-pr-review, gh pr diff, or GitHub UI) before requesting review from others

Release note

Poe now defaults to the OpenAI Responses API endpoint, enabling reasoning-effort control for all bots and built-in web search for supported bots.

Signed-off-by: Konv Suu <hi@kovsu.com>

cherry-ai-bot

Cherry Review · 阻塞

1 blocker
逐条见行内评论。

cherry-ai-bot · 2026-08-24T05:56:40Z

+
+## What the user should do
+
+Open Settings → Model Providers → Poe and switch the API endpoint to


必须修复：Existing Poe provider rows remain routed through openai-chat-completions, but this PR removes Poe-specific reasoning and web-search handling while documenting a Settings endpoint switch that the UI does not persist for existing providers.

inv_564ed6016e8f062a#c0

kangfenmao

发现一个必须在合并前处理的升级兼容性问题，详见行内评论。

kangfenmao · 2026-08-27T11:25:57Z

        "developerRole": false
      },
-      "defaultChatEndpoint": "openai-chat-completions",
+      "defaultChatEndpoint": "openai-responses",


[P1] 为现有 Poe provider 行保留/迁移旧端点。PresetProviderSeeder 的 batchUpsertTx 只插入不存在的 providerId，因此已存在的 Poe 行仍为 openai-chat-completions；本 PR 同时移除了 reasoning.ts/websearch.ts 的 Poe 分支，升级用户的 reasoning effort 和 web search 会静默失效。请在升级路径增加一次性迁移或保留兼容分支，并确保 endpoint config/default/model endpoint 一致后再移除旧协议支持。

…rovider Signed-off-by: Konv Suu <hi@kovsu.com> # Conflicts: # packages/provider-registry/data/models.json # packages/provider-registry/data/providers.json # packages/provider-registry/src/providers/poe.ts # src/main/ai/utils/reasoning.ts # src/main/ai/utils/websearch.ts # src/renderer/pages/settings/ProviderSettings/providerUrls.ts

models.dev now ships reasoning controls for ling-3-0-flash-fin; the membership invariant requires the owning creator to claim every catalog reasoning model. Signed-off-by: Konv Suu <hi@kovsu.com>

Signed-off-by: Konv Suu <hi@kovsu.com>

…breaking The chat-completions fallback keeps its reasoning contracts and web search, so existing installs lose nothing; only the default for fresh installs changes. Signed-off-by: Konv Suu <hi@kovsu.com>

models.dev dropped the accounts/fireworks/routers/* kimi SKUs, leaving the source overrides dangling; pin apiModelId and name so they stay resolvable as standalone rows. Signed-off-by: Konv Suu <hi@kovsu.com>

cherry-ai-bot

Cherry Review · 阻塞

1 blocker
逐条见行内评论。

已有更新的评审结论，撤回这条 REQUEST_CHANGES

Poe's Responses emulation for Claude (LiteLLM→Vertex) breaks streams, so the 4.5-line official bots join the anthropic-messages pin, using the exact lowercase bot names api.poe.com/v1/models serves; an explicit-set contract test locks the full pinned roster. Also pins the Fireworks wire ids that models.dev de-listed (gpt-oss-20b, minimax-m2p7) so requests keep valid model paths. Catalog regenerated from source. Signed-off-by: Konv Suu <hi@kovsu.com>

The manifest embeds the provider catalog content hash; regenerate it against the updated catalog so build:builtin-knowledge:check passes. Signed-off-by: Konv Suu <hi@kovsu.com>

Sparse provider rows delegate to the registry preset by design (CherryHQ#17096), so the endpoint switch applies to existing installs too; document the real semantics and the explicit-choice escape hatch instead of promising a fresh-install-only change. Signed-off-by: Konv Suu <hi@kovsu.com>

Signed-off-by: Konv Suu <hi@kovsu.com>

cherry-ai-bot

Cherry Review · 通过

1 warning
逐条见行内评论。

cherry-ai-bot · 2026-08-28T04:35:43Z

  { modelId: 'gpt-oss-120b', values: ['low', 'medium', 'high'] },
-  { modelId: 'gpt-oss-20b', values: ['low', 'medium', 'high'] },
-  { modelId: 'minimax-m2-7', values: ['low', 'medium', 'high'] },
+  { modelId: 'gpt-oss-20b', apiModelId: 'accounts/fireworks/models/gpt-oss-20b', values: ['low', 'medium', 'high'] },


建议修复：The manually pinned Fireworks model overrides omit provider-specific pricing, so regenerated catalog rows lose known Fireworks prices. GPT OSS 20B and MiniMax M2.7 then inherit zero input/output pricing from the global catalog, while the standalone Kimi router rows have no pricing at all.

inv_7806d61e74d9fe01#c0

已有更新的评审结论，撤回这条 REQUEST_CHANGES

This comment was marked as outdated.

Sign in to view

kovsu requested a review from DeJeune April 9, 2026 09:35

kovsu force-pushed the hotfix/poe-custom-provider branch from 9b7e423 to 0911df1 Compare July 13, 2026 08:15

kovsu requested a review from 0xfullex as a code owner July 13, 2026 08:15

kovsu added 2 commits July 13, 2026 16:21

update

7eab401

docs(breaking-changes): record poe responses endpoint requirement

8cfb7fc

Signed-off-by: Konv Suu <hi@kovsu.com>

kovsu force-pushed the hotfix/poe-custom-provider branch from 0911df1 to 8cfb7fc Compare July 13, 2026 08:22

cherry-ai-bot Bot previously requested changes Aug 24, 2026

View reviewed changes

kangfenmao requested changes Aug 27, 2026

View reviewed changes

kovsu changed the title ~~feat: poe support responses api~~ feat(poe): default to the openai responses api endpoint Aug 28, 2026

kovsu added 5 commits August 28, 2026 09:13

fix(provider-registry): claim ling-3-0-flash-fin reasoning membership

aa1bbe4

models.dev now ships reasoning controls for ling-3-0-flash-fin; the membership invariant requires the owning creator to claim every catalog reasoning model. Signed-off-by: Konv Suu <hi@kovsu.com>

test(provider-registry): lock poe responses endpoint contract

6ede9e3

Signed-off-by: Konv Suu <hi@kovsu.com>

docs(breaking-changes): poe endpoint default switch is a notice, not …

41566de

…breaking The chat-completions fallback keeps its reasoning contracts and web search, so existing installs lose nothing; only the default for fresh installs changes. Signed-off-by: Konv Suu <hi@kovsu.com>

fix(provider-registry): pin de-listed fireworks kimi router identities

c188555

models.dev dropped the accounts/fireworks/routers/* kimi SKUs, leaving the source overrides dangling; pin apiModelId and name so they stay resolvable as standalone rows. Signed-off-by: Konv Suu <hi@kovsu.com>

kovsu force-pushed the hotfix/poe-custom-provider branch from f97cf67 to c188555 Compare August 28, 2026 01:17

cherry-ai-bot Bot previously requested changes Aug 28, 2026

View reviewed changes

Comment thread packages/provider-registry/data/provider-models.json

kovsu added 3 commits August 28, 2026 11:32

chore(builtin-knowledge): regenerate product manifest

de62f51

The manifest embeds the provider catalog content hash; regenerate it against the updated catalog so build:builtin-knowledge:check passes. Signed-off-by: Konv Suu <hi@kovsu.com>

kovsu requested a review from kangfenmao August 28, 2026 03:37

DeJeune approved these changes Aug 28, 2026

View reviewed changes

fix(provider-registry): align poe catalog contracts

ae4517d

Signed-off-by: Konv Suu <hi@kovsu.com>

cherry-ai-bot Bot approved these changes Aug 28, 2026

View reviewed changes

Merge branch 'main' into hotfix/poe-custom-provider

8df8538

Navigation Menu

Uh oh!

feat(poe): default to the openai responses api endpoint - #14144

feat(poe): default to the openai responses api endpoint#14144
kovsu wants to merge 12 commits into
CherryHQ:mainfrom
kovsu:hotfix/poe-custom-provider

kovsu commented Apr 9, 2026 •

edited

Loading

Uh oh!

This comment was marked as outdated.

cherry-ai-bot Bot left a comment

Uh oh!

cherry-ai-bot Bot Aug 24, 2026

Uh oh!

kangfenmao left a comment

Uh oh!

kangfenmao Aug 27, 2026

Uh oh!

cherry-ai-bot Bot left a comment

Uh oh!

Uh oh!

cherry-ai-bot Bot left a comment

Uh oh!

cherry-ai-bot Bot Aug 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants


		## What the user should do

		Open Settings → Model Providers → Poe and switch the API endpoint to

Uh oh!

Conversation

kovsu commented Apr 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Branch strategy

What this PR does

Why we need it and why it was done in this way

Breaking changes

Special notes for your reviewer

Checklist

Release note

Uh oh!

This comment was marked as outdated.

cherry-ai-bot Bot left a comment

Choose a reason for hiding this comment

Uh oh!

cherry-ai-bot Bot Aug 24, 2026

Choose a reason for hiding this comment

Uh oh!

kangfenmao left a comment

Choose a reason for hiding this comment

Uh oh!

kangfenmao Aug 27, 2026

Choose a reason for hiding this comment

Uh oh!

cherry-ai-bot Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

cherry-ai-bot Bot left a comment

Choose a reason for hiding this comment

Uh oh!

cherry-ai-bot Bot Aug 28, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

kovsu commented Apr 9, 2026 •

edited

Loading