fix: route Kimi coding fallback via Anthropic messages

[ksukys]

Summary

• Detect api.kimi.com/coding endpoints as Anthropic Messages-compatible during agent initialization
• Apply the same detection when activating fallback providers
• Add regression coverage for primary and fallback Kimi coding routing

Why

When a primary provider such as OpenAI Codex is rate-limited, fallback to the Kimi coding provider could incorrectly remain in chat_completions mode. That caused Hermes to create an OpenAI-style client against https://api.kimi.com/coding and hit a 404 instead of using the Anthropic Messages adapter.

Test Plan

• python -m pytest tests/run_agent/test_run_agent.py::TestInit::test_kimi_coding_base_url_uses_anthropic_messages_mode tests/run_agent/test_run_agent.py::TestFallbackAnthropicProvider::test_fallback_to_kimi_coding_sets_anthropic_messages_mode tests/run_agent/test_run_agent.py::TestFallbackAnthropicProvider::test_fallback_to_openrouter_uses_openai_client -q -o 'addopts='
• python -m pytest tests/run_agent/test_run_agent.py::TestInit tests/run_agent/test_run_agent.py::TestFallbackAnthropicProvider -q -o 'addopts='

[alt-glitch]

Note: This PR treats api.kimi.com/coding as Anthropic Messages-compatible, which conflicts with #17107 (open) that argues it should NOT be treated as Anthropic Messages. Related: #17027 (merged, added the detection), #13782 (merged, earlier Kimi fix). Needs clarification on which wire protocol Kimi coding actually speaks.

[ksukys]

Follow-up after retesting against current origin/main:

I verified the latest upstream behavior in a clean worktree before reducing this patch. Upstream now covers the normal provider-resolution path for Kimi, but two paths still keep Kimi Coding in chat_completions mode:

1. Direct agent initialization with base_url=https://api.kimi.com/coding and provider=kimi-coding
2. Fallback activation when switching from a rate-limited primary provider to Kimi Coding

Both should route through anthropic_messages, otherwise Hermes constructs an OpenAI-style client for https://api.kimi.com/coding and hits the wrong endpoint.

I also reduced the patch scope after that verification:

• keep only URL/API-mode detection in the agent init path and fallback activation path
• add regression tests for direct init and fallback activation
• drop the previous AnthropicAuxiliaryClient.default_headers / auxiliary-client change, since upstream already fixed the normal provider-resolution/auxiliary-client side

On the #17107 concern: this PR should apply only to the bare Kimi Coding route (https://api.kimi.com/coding), which speaks Anthropic Messages. The OpenAI-compatible /coding/v1 path should remain chat_completions and should not be matched by this fix.

Local validation on the reduced patch:

• venv/bin/python -m compileall -q agent run_agent.py
• targeted regression tests for direct init + fallback activation: passed
• Kimi fallback verifier: passed, including the broader Kimi/provider-related selected tests
• hermes chat -q 'Reply with exactly: OK' --quiet --toolsets safe: OK

[teknium1]

Thanks for isolating the direct-init and runtime-fallback paths; both remain unhandled on current main (agent/agent_init.py:441-454 and agent/chat_completion_helpers.py:1363-1380).

Problems

• The new "/coding" in ... predicates and both new tests include /coding/v1, but the follow-up comment says only bare /coding should use Anthropic Messages and /coding/v1 must remain chat_completions. This needs resolution before the patch can express the intended contract.
• agent/agent_init.py:989-1028 also activates a configured fallback without recomputing api_mode; it is a sibling fallback path not covered by this PR.
• Current main already has a competing /coding/v1 Anthropic classification in hermes_cli/runtime_provider.py:131 and tests/agent/test_auxiliary_transport_autodetect.py:37, matching the open #17107 discussion conflict.

Suggested changes

• Establish and test the bare /coding versus /coding/v1 wire contract, then make all agent fallback/init paths use that same exact rule.
• Include init-time fallback coverage if its desired transport behavior is the same.

Automated hermes-sweeper review.

[teknium1]

This regression encodes /coding/v1 as Anthropic, but the follow-up comment says /coding/v1 must remain chat_completions and only bare /coding should change. Please split the endpoint cases after the transport contract is confirmed.

[GottZ]

This was generated by AI during triage.

Summary

Three open PRs address the Kimi fallback transport bug: #25706 covers direct initialization and runtime fallback, #33276 reuses the shared URL detector only in runtime fallback, and #41020 adds provider-wide runtime-fallback classification. All three currently over-classify at least one OpenAI-compatible Kimi or Moonshot route instead of distinguishing the bare api.kimi.com/coding endpoint from /coding/v1 and Moonshot /v1.

Related pull requests

• #25706 related — (+56/-4) — preferred consolidation target, but not merge-ready: it is the only diff covering both direct initialization and runtime fallback with regression tests, yet its substring predicate and tests classify /coding/v1 as Anthropic Messages and it omits the sibling init-time fallback path. This aligns with the keep_open review on #25706: establish the exact bare /coding contract, preserve /coding/v1 as OpenAI-compatible, and apply the shared rule to every init/fallback path.
• #33276 related — (+10/-0) — duplicate fallback-only approach: it delegates to the shared URL detector, but that detector currently also classifies /coding/v1 as Anthropic Messages and the PR adds no regression tests. Despite the keep_open review on #33276, its diff is a narrower subset of #25706 and does not independently resolve the disputed wire contract, so its useful detector-reuse approach should be incorporated into #25706 rather than maintained separately.
• #41020 related — (+4/-0) — duplicate with an over-broad fix: provider-identity routing sends all kimi-coding and kimi-coding-cn fallbacks through Anthropic Messages, incorrectly including Moonshot /v1 endpoints documented by existing tests as non-Anthropic-wire. Despite the keep_open review on #41020, the diff duplicates only the fallback portion and requires replacement with URL-specific detection plus tests, making #25706 the stronger consolidation target.

Duplicates

#33276 and #41020 substantially duplicate the runtime-fallback portion of #25706; #33276 proposes shared URL-detector reuse, while #41020 uses a broader and incorrect provider-wide branch.

Suggested consolidation

Merge #25706 only after narrowing the contract to bare api.kimi.com/coding, keeping /coding/v1 and Moonshot /v1 OpenAI-compatible, centralizing that URL rule, and covering direct init, runtime fallback, and init-time fallback with positive and negative tests. Then close #33276 and #41020 as superseded duplicates; their keep_open reviews identify real fallback gaps, but their current diffs neither resolve the protocol distinction nor provide broader correct coverage than the amended #25706.

Cross-PR triage: Reviewed 3 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 7 kB of PR diffs, 4 kB of issue/PR text, 6 kB of discussion (6 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.