feat(providers): add alibaba, kimi, zhipu with dual OpenAI/Anthropic mounts - #6054

No actionable comments were generated in the recent review. 🎉

Merge Risk: 🔵 Low · up to 744a3

The PR adds three providers, dual-protocol routing, and provider-specific reasoning controls. A few bounded issues remain: Zhipu General API credentials can be used with an incompatible Anthropic mode, equivalent GLM inputs can be serialized differently, and Alibaba’s effort documentation is inconsistent. These warrant owner follow-up but do not indicate a release-blocking failure.

_{Comment @coderabbitai help to get the list of available commands.}

Thanks for the review @coderabbitai. Triaged the four comments below — one fixed, two skipped with rationale, one deferred.

1. config.schema.json — use_anthropic_endpoints descriptions for kimi_key / zhipu_key ✅ fixed
Valid catch. Both said "Routes chat completions requests…" but this PR implements Responses routing for kimi/zhipu too. Updated both to "Routes chat completions and responses requests through Anthropic-compatible endpoints." to match deepseek_key and reflect actual behavior (8/8 descriptions now consistent).

2. llmtests/chat_completion_stream.go + responses_stream.go — make the tool-stream cap fatal 🚫 skipped
The graceful goto at the 1500-chunk bound is intentional, and its semantics differ from the main-stream fatal cap (2500):

• The main-stream cap is fatal because that path's only job is to stream content — hitting it means something is wrong.
• The tool-stream path's job is to detect a tool call; once one is seen, subsequent chunks are noise. GLM models stream reasoning token-by-token before emitting the tool call, so a legitimate trace can run 1000+ chunks. Truncating after detection and passing is correct for verbose models.
• A true runaway (no natural completion) is still caught by the per-request streamCtx 30s timeout → t.Fatal("Timeout…"), and by the main-path fatal cap when the no-tool-call validation fails ("received N chunks but no tool calls").

Making the tool-stream cap fatal would turn legitimate verbose GLM streams red, which is the exact regression the raise was fixing.

3. core/providers/alibaba/alibaba.go — defensively clone NetworkConfig maps 🚫 skipped
The alibaba constructor is structurally identical to the merged NewDeepSeekProvider — same config.CheckAndSetDefaults() call, same networkConfig: config.NetworkConfig storage. CheckAndSetDefaults already deep-copies ExtraHeaders (codebase gotcha #4 in AGENTS.md). Applying the extra clone only to alibaba would diverge from the established provider pattern; if the defensive copy is worth adding, it belongs in CheckAndSetDefaults (shared infra) as a separate change covering all providers.

4. apiKeysFormFragment.tsx — disable the toggle for Zhipu General API URLs ⏳ deferred
Agreed it's a real UX gap, but the current behavior already surfaces a clean upstream error (the derived General-API Anthropic mount 401s), and the docs page calls out that the Anthropic mount is Coding-Plan-only. URL-coupled conditional disabling in the key form couples the shared form to provider-specific URL rules — worth a dedicated follow-up rather than bundling into this provider PR. Happy to open a separate issue/PR for it post-merge.

The schema-description fix is pushed in ab45ab9.

#!/bin/bash
set -euo pipefail

printf '%s\n' '=== schema descriptions ==='
rg -n -A10 -B3 '"kimi_key"|"zhipu_key"|"use_anthropic_endpoints"' transports/config.schema.json

printf '%s\n' '=== stream-bound control flow ==='
ast-grep outline core/internal/llmtests/chat_completion_stream.go --items all
ast-grep outline core/internal/llmtests/responses_stream.go --items all
rg -n -A12 -B12 '1500|2500|streamCtx|goto|tool call|tool_calls' \
  core/internal/llmtests/chat_completion_stream.go \
  core/internal/llmtests/responses_stream.go

printf '%s\n' '=== provider defaults and Alibaba constructor ==='
ast-grep outline core/schemas --items all --match 'CheckAndSetDefaults'
ast-grep outline core/providers/alibaba/alibaba.go --items all --match 'NewAlibabaProvider'
rg -n -A30 -B8 'func \(.*\) CheckAndSetDefaults|CheckAndSetDefaults\s*\(|networkConfig: config\.NetworkConfig|ExtraHeaders' \
  core/schemas core/providers/alibaba/alibaba.go

@is911 Verified.

1. kimi_key and zhipu_key now state that use_anthropic_endpoints routes both chat completions and responses requests. The schema finding is fixed.

2. I accept the tool-stream rationale. The 1,500-chunk limit ends collection after tool-call detection. Validation still fails when no tool call is detected. Normal streams retain the fatal 2,500-chunk limit. Note that the Responses tool stream uses a 200-second timeout, while the chat tool stream uses a 30-second timeout.

3. I withdraw the NetworkConfig cloning concern. ProviderConfig.CheckAndSetDefaults() clones both ExtraHeaders and BetaHeaderOverrides before AlibabaProvider retains config.NetworkConfig. A provider-local copy would duplicate shared behavior.

4. I agree that the Zhipu General API toggle is a valid deferred UX improvement. Please confirm if you want a follow-up issue for this item after merge.

🐇 ✅

Learnt from: is911
URL: https://github.com/maximhq/bifrost/pull/6054

Timestamp: 2026-08-11T09:03:18.658Z
Learning: In `core/internal/llmtests/chat_completion_stream.go` and `core/internal/llmtests/responses_stream.go`, the 1,500-chunk limit for tool-call streams intentionally ends collection without failing after enough chunks have been inspected. Tool-stream validation still fails when no tool call or function-call arguments are detected. The normal streaming paths use a separate fatal 2,500-chunk safety limit.

Learnt from: is911
URL: https://github.com/maximhq/bifrost/pull/6054

Timestamp: 2026-08-11T09:03:18.658Z
Learning: In Go providers, `schemas.ProviderConfig.CheckAndSetDefaults()` defensively copies `NetworkConfig.ExtraHeaders` and `NetworkConfig.BetaHeaderOverrides`. Provider constructors, including `core/providers/alibaba/alibaba.go` `NewAlibabaProvider`, can retain `config.NetworkConfig` after calling this shared method without adding provider-local copies of these maps.

_{You are interacting with an AI system.}

Confirmed — I'll open a follow-up issue for the Zhipu General-API toggle validation once this PR merges and link it back here. Thanks for running the verification on items 1–3.

All committers have signed the CLA.

@akshaydeo @TejasGhatte @Pratham-Mishra04 — friendly nudge for a review pass on this one when you have a window. It's been open for a few days and the DeepSeek-precedent diff makes it a fast review.

Heads-up on one gap I just closed (commit 61cf0e4): the per-vendor reasoning_effort shaping originally hard-pinned glm-5.2, so a future revision like glm-5.5 would have repeated the exact normalization gap this PR set out to fix — max silently downgraded to high on OpenAI-dialect mounts, and reasoning_effort stripped entirely on the zhipu mount. Verified by probe before the fix:

The fix is a shared isGLM52OrLater version-floor matcher (GLM-5.x where minor ≥ 2, Coding-Plan [1m] aliases included, 5.0/5.1 and GLM-4.x still excluded since they ignore or 400 the field) now used by supportsMaxReasoningEffort, isZhipuReasoningEffortModel, and isAlibabaReasoningEffortModel (alibaba matches the hosted GLM-5 series generally). Covered by TestIsGLM52OrLater + new glm-5.3/glm-5.5 routing cases; full core/providers/openai suite green.

For existing custom/OpenAI-dialect users on current dev, the short-term glm-5.3 entry is up as #6162 (follow-up to #4467) — once this merges, its explicit prefixes are subsumed by the version floor.

Happy to split per-provider or address anything else that's blocking — just say the word.

Addressed both remaining review findings in 7d67d93:

• Custom-provider gating: DeploymentsTable now receives the base provider type, and provider-specific deployment sections (including the Zhipu Anthropic gating) key off the effective provider — so a custom provider on the zhipu base type gets the correct alias-level behavior. providerName stays the real name for model lookup and display.
• Edit-flow timing + saveability: the persisted-toggle clear now watches key.use_anthropic_endpoints (so it runs after the async form.reset populates edit values) and clears with shouldDirty: true, making the correction saveable via the enabled Save button. The per-deployment override stays selectable under the General-API restriction — On is blocked, but a previously persisted On can be remediated via Use key setting / Off.

tsc --noEmit (only pre-existing branding route-tree error), oxlint, and oxfmt clean on the touched files.

Snyk remediation — commit 83f7188 resolves 10 of the 11 findings from the security/snyk check:

Remaining (unfixable): GO-2026-5932 flags golang.org/x/crypto/openpgp as unmaintained — there is no fixed version (Fixed in: N/A), and nothing in this repo imports the openpgp package (verified by grep across all modules), so it is a module-level advisory only. If the Snyk gate still fails on it, that finding needs a policy/severity call on the Snyk side rather than a code change.

Verified after the bumps: all workspace modules + the two standalone modules (cmd/e2eseed, the token-exchange example) build, and the provider test suites (openai, alibaba, kimi, zhipu) pass. govulncheck -scan module now reports only the openpgp advisory.

Branch rewritten onto current dev (a7944c3) to clear the merge conflicts that had accumulated behind the force-pushed dev.

What changed:

• The branch previously carried ~90 files of unrelated drift from an older dev head (MCP credstore/token-exchange, e2eseed, branding, bedrock/gemini/xai, SSRF, migration scripts) that dev has since re-landed independently — that drift was the source of both the 38-file conflicts and the Snyk manifest findings. It is gone; the PR is now a single clean commit with only the provider work (40 files).
• All provider feature content is preserved: dual mounts, per-vendor reasoning shaping, the GLM-5.2+ version floor (future GLM revisions work on day one), the Zhipu Coding-Plan UI gating, NetworkConfig cloning, and the llmtests stream-cap failure semantics. Conflicts against new dev content (xAI end-to-end test, Bedrock VPC-endpoint form fields) were resolved by keeping both sides.
• No dependency manifests change anymore, so the Snyk gate now passes ("No manifest changes detected"). The repo-wide bumps from the earlier attempt (go 1.26.6, klauspost/compress 1.18.7, otel 1.44.0) address findings that exist on dev itself regardless of this PR — happy to send them as a separate maintenance PR if wanted.
• Verified locally: workspace + core build, provider/anthropic/schemas test suites, go vet, gofmt, tsc (pre-existing branding route-tree artifact aside), oxlint/oxfmt, config.schema.json + docs.json validity.

@coderabbitai review

@akshaydeo — state-change note rather than a third bump: CodeRabbit's stale changes-requested cleared today (fresh pass is APPROVED), all threads resolved, checks green, no conflicts with dev. The only remaining blocker is one human approval.

You'd have the most context right now having just verified the adjacent Anthropic-mount logic in #6392 — the core/providers/anthropic/* files here are the trickiest slice. One logistics note: the diff touches Makefile (test targets), which is CODEOWNERS'd to @maximhq/bifrost-admin — if code-owner review is enforced that slice needs an admin pass, so feel free to route.