feat(agent): adaptive thinking for Kimi-family Anthropic endpoints

[chouqin]

Kimi's Anthropic-compatible endpoints (api.moonshot.cn/anthropic, api.kimi.com/coding) implement the adaptive thinking contract — they accept thinking.type=adaptive + output_config.effort (all of low, medium, high, xhigh, max verified live) and return thinking blocks, and the replay-validation 400s that originally motivated dropping the parameter (#13848) no longer occur.

_supports_adaptive_thinking() now returns True for Kimi-family models, so they get thinking={type: adaptive, display: summarized} + output_config.effort via ADAPTIVE_EFFORT_MAP instead of nothing, and the blanket drop of the thinking parameter for Kimi-family endpoints is removed. MiniMax and other non-adaptive third parties keep the manual budget_tokens path; Claude behavior is unchanged.

What does this PR do?

Kimi's Anthropic-compatible endpoints (api.moonshot.cn/anthropic, api.kimi.com/coding) implement the adaptive thinking contract. Stop dropping the thinking parameter for the Kimi family and send it in adaptive form:

Related Issue

Fixes #

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• _supports_adaptive_thinking() returns True for Kimi-family models → thinking={type: adaptive, display: summarized} + output_config.effort via ADAPTIVE_EFFORT_MAP (never manual budget_tokens)
• the blanket drop of the thinking parameter for Kimi-family endpoints ([Bug]: 400 error for kimi-for-coding #13848-era mitigation) is removed
• MiniMax and other non-adaptive third parties keep the manual budget_tokens path; Claude behavior unchanged

How to Test

Test it against the kimi for coding endpoint:

• thinking.type="adaptive" + output_config.effort in {low, medium, high, xhigh, max} → all HTTP 200 with thinking blocks; display="summarized" accepted
• multi-turn tool-call replays in every history shape — signed thinking blocks verbatim, unsigned placeholder blocks, and no thinking blocks at all (the exact [Bug]: 400 error for kimi-for-coding #13848 trigger) — all return HTTP 200; the replay-validation 400s that motivated the original drop no longer occur

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 15.2

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

Screenshots / Logs

[alt-glitch]

This was generated by AI during triage.

Related to merged #17455/#14018 and #56727: this reverses the earlier Kimi-thinking suppression based on a claimed changed endpoint contract. It is related work, not a duplicate; please validate the current multi-turn replay contract before choosing it.

[webtecnica]

Implemented in #67231 — same approach: _supports_adaptive_thinking() returns True for Kimi-family models, blanket thinking suppression removed. All 53 tests pass.

[teknium1]

Thanks for revisiting the Kimi contract. The current-main premise is real: agent/anthropic_adapter.py:2647-2648 still suppresses Anthropic thinking for Kimi-family endpoints.

Problems

• The stated signed-thinking replay guarantee is not implemented by this diff. agent/anthropic_adapter.py:2297-2308 preserves only unsigned Kimi thinking blocks and removes any block with signature or data at :2304-2306; this PR changes neither branch.
• The changed tests cover request construction and effort mapping, but not the stated signed-history or no-thinking-block replay cases. Existing replay coverage at tests/agent/test_kimi_coding_anthropic_thinking.py:175-212 covers only an unsigned reasoning_content-derived block.

Suggested changes

• Add conversion tests for signed, unsigned, and absent-thinking prior tool-call histories.
• Validate and encode the required signed-block behavior before removing the suppression; preserve the current signature-stripping safety rule unless the verified Kimi contract requires a different Kimi-specific path.

Automated hermes-sweeper review.

[teknium1]

Merged via PR #67606 — your commit was cherry-picked onto current main with your authorship preserved in git log (rebase-merge). Thanks for revisiting the Kimi thinking contract and for the thorough live verification across effort levels and replay shapes; combined with the replay-side fix in ddd81e9, Kimi-family endpoints now get full adaptive thinking on the Anthropic wire.