feat(server): add Kimi Code CLI provider over ACP

[maiconfontana]

Summary

Adds optional Kimi Code CLI support as a first-class built-in provider, following the same ACP adapter pattern as Cursor/Grok.

• Spawns kimi acp for sessions, prompts, interrupts, approvals, resume, and text generation
• Uses the live catalog / config options from Kimi (kimi-code/* models, thinking low/high/max)
• Projects agent_thought_chunk into thinking worklog activity and streams assistant text
• Wires contracts/settings, web picker + icon, install docs, and focused tests

Why this exists

Kimi Code CLI is a real terminal agent with official ACP support (kimi acp). T3 already wraps multiple BYO CLIs; this fills that gap without requiring a separate product surface.

Scope / CONTRIBUTING note

I read CONTRIBUTING.md. This is larger than the “small fixes only” bar you currently prefer, and I understand you may close or defer large external feature PRs. Opening this so maintainers can decide, request a split, or take the idea in-house if useful.

Happy to:

• open an issue first and close this if preferred
• split into smaller follow-ups (adapter-only, models/settings, UI chrome, docs)
• drop UI chrome / docs if you only want the server adapter path

How to test

1. Install Kimi Code CLI and run kimi → /login
2. Checkout this branch; start T3; confirm Kimi in the model picker / Settings
3. Enable Settings → Assistant output for token streaming if desired
4. Start a thread on kimi-code/kimi-for-coding (or K3)
5. Confirm thinking appears in the worklog and assistant text streams; approvals still work

Test plan

• Focused unit tests for Kimi ACP support / provider / adapter / text generation
• Live kimi acp probe: auth, model list, thought + message chunks
• Manual UI session against authenticated local Kimi

Made with Grok Build.

---

Note

Medium Risk
Large new provider surface (subprocess ACP, turn/prompt lifecycle) plus shared ingestion changes for reasoning streams; mistakes could affect non-Kimi ACP providers or worklog behavior.

Overview
Adds Kimi Code CLI as a sixth built-in provider, using the same ACP pattern as Cursor/Grok (kimi acp sessions, health checks, model discovery, and kimi upgrade maintenance).

Server: New KimiDriver, KimiAdapter (sessions, steer/interrupt prompt settlement, permissions, resume), KimiProvider snapshot/probe, and KimiTextGeneration for commit/PR/branch/title helpers. KimiAcpSupport normalizes kimi-code/* model ids and applies thinking/mode via ACP config options.

ACP + orchestration: Parses agent_thought_chunk as reasoning_text content deltas (Cursor/Grok adapters now forward streamKind). Reasoning streams skip assistant segment buffering; ingestion buffers per-turn thought text into Thinking worklog tasks (progress + completion when assistant text or turn ends).

Contracts & web: KimiSettings, defaults/aliases, registry entry; picker, settings meta, and KimiIcon. Docs/README/AGENTS updated for install and provider list.

^{Reviewed by Cursor Bugbot for commit a0b60ca. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Kimi Code CLI as a built-in provider over ACP

• Adds a full Kimi provider driver (KimiDriver.ts) with session adapter, text generation service, ACP runtime wiring, model discovery, and provider health checks.
• Introduces makeKimiTextGeneration (KimiTextGeneration.ts) for commit messages, PR content, branch names, and thread titles via ACP-backed JSON prompts with a 180s timeout.
• Extends ACP session parsing to handle agent_thought_chunk events, emitting ContentDelta with streamKind: "reasoning_text" alongside the existing assistant_text kind.
• The ingestion layer (ProviderRuntimeIngestion.ts) now buffers per-turn reasoning streams, surfacing a 'Thinking' task with progress updates (every ~280 chars, capped at 24,000 chars) and completing it when assistant text starts or the turn ends.
• Kimi appears in the provider picker with a 'new' badge, its own icon, and configurable settings (binary path, custom models) in the settings UI.

^{Macroscope summarized a0b60ca.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 29c8035f-d17b-44f8-a615-cdde73c4c3ff

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[maiconfontana]

Opened an issue first per CONTRIBUTING: #5244

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR adds a new provider (Kimi Code CLI) with ~3600 lines of new code including driver, adapter, ACP integration, text generation, and UI components. New feature additions of this scope introducing new user-facing capabilities warrant human review.

^{You can customize Macroscope's approvability policy. Learn more.}

[maiconfontana]

Addressed the automated review findings on c8a863902:

• Reasoning buffer (High): cap at 24k chars; only open a Thinking task once non-whitespace content arrives; only advance lastPublishedLength when progress is published; always complete an opened task; clear bufferedReasoningByTurnKey in clearTurnStateForSession
• Grouped models (Medium): flatten ACP select groups in buildKimiDiscoveredModelsFromConfigOptions
• Text gen policy (Medium): forward input.policy for branch name + thread title
• Docs (Low): built-in driver count 5 → 6

Focused Kimi provider/adapter/ACP tests pass. One pre-existing KimiTextGeneration mock model id flake (kimi-code/grok-mock-alt) is unrelated to these changes.

[macroscopeapp]

Reviewed the new Kimi provider modules (KimiDriver, KimiAdapter, KimiProvider, KimiAcpSupport, KimiTextGeneration) plus the touched ACP/ingestion code against the Effect service conventions. Dependencies are acquired from the environment (yield* Crypto.Crypto, ChildProcessSpawner, ServerSettingsService, Effect.service(ServerConfig)), Effect modules are imported as subpath namespaces, no ManagedRuntime/runPromise leaks into service construction, and error handling uses Schema.TaggedErrorClass types with Effect.catchTags. One finding below on a duplicated service shape.

^{Posted via Macroscope — Effect Service Conventions}

[maiconfontana]

Review pass complete (7eb874d56)

Addressed the remaining open bot findings after the first fixup. Summary of all review threads:

Fixed (valid)

Finding	Resolution
Reasoning buffer unbounded (High)	Cap at 24k chars
Whitespace Thinking task stale (Medium / Bugbot)	Only start on non-whitespace; only advance lastPublishedLength on published progress; always complete if started
clearTurnStateForSession missed reasoning cache (Medium)	Invalidate bufferedReasoningByTurnKey
Grouped ACP model entries dropped (Medium)	Flatten { options: [...] } in model discovery
Docs driver count 5→6 (Low)	Fixed
Text gen ignores modelSelection.options (Medium, new)	Pass selections: modelSelection.options into applyKimiAcpModelSelection
Structural ACP runtime type (Effect conventions)	Type as Pick/Partial of AcpSessionRuntime.Service
Kimi text-gen test mock model id	Mock agent accepts kimi-code/* ids; assertion expects prefixed id

False positive / out of scope

Finding	Why
Pass policy to branch/title generation	Incorrect. BranchNameGenerationInput / ThreadTitleGenerationInput have no policy field (only commit/PR do). Passing it fails typecheck. Prompt helpers accept optional policy for future use, but the service contract does not. Other providers (Claude/Codex/Cursor/Grok) also omit it. Reverted.

Verification

• Focused tests: Kimi provider, adapter, ACP support, text generation — 19/19 pass
• CodeRabbit auto-review is disabled on this repo (settings); happy to re-trigger with @coderabbitai review if maintainers want

Scope note (unchanged)

Still a large feature PR; issue #5244 remains the tracking entry. Ready for human review of the adapter surface.

[coderabbitai]

Tip

For best results, initiate chat on the files or code changes.

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

echo '== Fixup commit summary =='
git show --stat --oneline 7eb874d56

echo
echo '== Candidate files =='
fd -t f . | rg '(Kimi|ProviderRuntimeIngestion|AcpRuntimeModel|README|install|internals)' | head -200

echo
echo '== Relevant definitions and usages =='
rg -n -C 3 \
  'bufferedReasoningByTurnKey|lastPublishedLength|MAX_.*REASONING|24000|24_000|clearTurnStateForSession|buildKimiDiscoveredModelsFromConfigOptions|applyKimiAcpModelSelection|selections: modelSelection\.options|AcpSessionRuntime\.Service|kimi-code/grok-mock-alt|built-in.*driver|driver.*[56]' \
  . \
  -g '!node_modules/**' -g '!dist/**' -g '!build/**' | head -500

echo
echo '== Fixup diff for the reported areas =='
git show --format=fuller --find-renames 7eb874d56 -- \
  ':!node_modules' ':!dist' ':!build' | sed -n '1,1400p'

Length of output: 44358

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

echo '== Symbol map =='
ast-grep outline apps/server/src/provider/Layers/KimiProvider.ts \
  --items all --type function --match 'buildKimiDiscoveredModelsFromConfigOptions|buildKimiThinkingCapabilitiesFromConfigOptions'
ast-grep outline apps/server/src/textGeneration/KimiTextGeneration.ts \
  --items all --type function --match 'generateBranchName|generateThreadTitle|buildBranchNamePrompt|buildThreadTitlePrompt'
ast-grep outline apps/server/src/provider/Layers/KimiProvider.test.ts \
  --items all --type function --match 'discover|config|group'

echo
echo '== Grouped ACP option implementation and tests =='
sed -n '185,255p' apps/server/src/provider/Layers/KimiProvider.ts
rg -n -C 5 'group|options: \[|configOptions|discovered model' \
  apps/server/src/provider/Layers/KimiProvider.test.ts \
  apps/server/src/provider/acp/KimiAcpSupport.test.ts

echo
echo '== Branch/title generation input contracts and peer-provider calls =='
rg -n -C 5 \
  'interface BranchNameGenerationInput|type BranchNameGenerationInput|interface ThreadTitleGenerationInput|type ThreadTitleGenerationInput|generateBranchName|generateThreadTitle' \
  apps/server/src/textGeneration \
  packages \
  -g '*.ts' | head -550

echo
echo '== Reasoning ingestion tests =='
rg -n -C 4 \
  'reasoning|Thinking|24_000|24000|whitespace|clearTurnStateForSession' \
  apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.test.ts \
  apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.approval.test.ts

Length of output: 4442

---

🏁 Script executed:

#!/bin/bash
set -u

echo '== Branch and title input contracts =='
rg -n -C 6 \
  'BranchNameGenerationInput|ThreadTitleGenerationInput|generateBranchName|generateThreadTitle' \
  . -g '*.ts' -g '!node_modules/**' -g '!dist/**' -g '!build/**' | head -700 || true

echo
echo '== Kimi model-discovery tests =='
rg -n -C 6 \
  'configOptions|availableModels|group|models.*discover|discover.*models|model.*select' \
  apps/server/src/provider/Layers/KimiProvider.test.ts \
  apps/server/src/provider/acp/KimiAcpSupport.test.ts || true

echo
echo '== Reasoning-ingestion test coverage =='
rg -n -C 6 \
  'reasoning|Thinking|whitespace|24_000|24000|session\.exited|clear.*session' \
  apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.test.ts \
  apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.approval.test.ts || true

echo
echo '== Kimi text-generation option propagation test coverage =='
rg -n -C 6 \
  'options|thinking|set_config_option|setConfigOption|session/set_config_option' \
  apps/server/src/textGeneration/KimiTextGeneration.test.ts \
  apps/server/src/provider/acp/KimiAcpSupport.test.ts || true

Length of output: 84109

---

@maiconfontana Static verification supports the reported resolutions.

• The reasoning buffer is capped at 24,000 characters. Session cleanup invalidates its cache.
• Grouped ACP model entries are flattened before model discovery.
• Kimi text generation forwards modelSelection.options.
• The ACP runtime parameter now uses the AcpSessionRuntime service type.
• The mock model IDs and the test assertion use the kimi-code/* prefix.

The policy revert is correct. BranchNameGenerationInput and ThreadTitleGenerationInput do not define policy. The other providers also omit it for branch-name and thread-title generation.

I could not execute the focused tests in this environment. I found no remaining source-level issue in the addressed findings.

---

✏️ Learnings added

Learnt from: maiconfontana
URL: https://github.com/pingdotgg/t3code/pull/5243

Timestamp: 2026-08-02T21:05:47.216Z
Learning: In `apps/server/src/textGeneration/TextGeneration.ts`, `BranchNameGenerationInput` and `ThreadTitleGenerationInput` do not have a `policy` field. Text-generation providers, including Kimi, Claude, Cursor, and Grok, must not forward `input.policy` to `buildBranchNamePrompt` or `buildThreadTitlePrompt` unless the service contracts add that field.

Note: Learnings are effective only in the context of similar code segments. To apply general coding guidelines, please configure review instructions. You can manage existing learnings in the app.

_{You are interacting with an AI system.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit a0b60ca. Configure here.}

[cursor]

Stuck Thinking on session exit

Medium Severity

session.exited clears bufferedReasoningByTurnKey without completing an opened Thinking task. Kimi stopSession emits exit without turn.completed, so a mid-thought stop leaves a permanent in-progress Thinking entry in the worklog.

Additional Locations (2)

• apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.ts#L1374-L1380
• apps/server/src/provider/Layers/KimiAdapter.ts#L502-L521

 

^{Reviewed by Cursor Bugbot for commit a0b60ca. Configure here.}

[t3dotgg]

Note

🤖 GPT-5.6 Sol responding on behalf of Theo

Closing this PR after an automated pass over open pull requests. Outside Kimi provider integration. Three separate implementations compete for the same unsupported provider.