feat: add minimal Kimi Code integration by jackie-cqz · Pull Request #1964 · oceanbase/powercontext

jackie-cqz · 2026-10-12T17:09:25Z

Which issue or RFC does this PR close?

Closes #1963.

Rationale for this change

Kimi Code users need a working PowerContext memory integration. This adds an experimental minimal profile through Kimi's native plugin, HTTP MCP and UserPromptSubmit hook.

What changes are included in this PR?

Add powercontext setup kimi-code and readonly powercontext doctor kimi-code. Setup installs from a local checkout into Kimi's managed plugin directory, preserves unrelated configuration and existing disablement, and retains the previous plugin for recovery if rollback fails.
Expose search_memory, remember_memory and get_artifact_revision. The plugin uses an explicitly configured existing Scope and runtime bearer-token environment variable.
Prepare and inject historical context, and optionally capture text prompts as Source. Automatic work has a four-second budget, continues ordinary tasks on failure, emits redacted diagnostics and does not retry writes with uncertain outcomes.
Keep setup and runtime endpoint validation consistent for the full IPv4 loopback range (127.0.0.0/8), and include repository-compliant license headers in the Jinja guidance templates. Regression tests cover automatic context/capture on 127.0.0.2, readonly diagnostics across the loopback range, and rejection of non-loopback plaintext endpoints.
Add generated routing guidance, English/Chinese documentation, capability declarations, behavior tests and opt-in real-host acceptance.

Are there any user-facing changes?

Users install the plugin from a local checkout and open a new Kimi session. One Kimi home uses one configured Scope. Explicit memory operations work without Generation; automatic memory creation requires the Server's Generation configuration and scheduler. This experimental profile does not include work handoff or expand the first-class setup wizard. No HTTP API or existing host contract changes.

How was this change tested?

124 tests passed with a real Kimi Code 2.1.1 npm CLI, real PowerContext Server and SQLite; inference in this repeatable test is controlled:

KIMI_CLI_BIN=<built main.mjs> uv run pytest tests/e2e/test_kimi_code_host.py tests/test_cli_kimi_code.py tests/test_cli_zcode.py tests/test_integration_manifest.py tests/test_guidance_installation.py tests/test_guidance_generation.py -q -p no:cacheprovider --basetemp=output/pytest-kimi-review-fix

Ten hook behavior tests passed:

node --test integrations/kimi-code/plugins/powercontext/tests/plugin.test.mjs

Manual live acceptance on Windows with Kimi Code 2.1.1, Node 24.15.0 and the official DeepSeek OpenAI Chat Completions API (deepseek-v4-flash) passed native MCP save/exact revision read/search, real Source-to-Memory Generation with provenance, and a fresh-session final answer from injected memory without tool calls. Kimi and Server Generation both used live inference. The run used an isolated Kimi home, Scope and SQLite database.
Full pre-commit checks passed with SKIP=ty-check uv run --locked --no-sync prek run -a; type checking was run separately with ty check --python-platform linux and passed. Changed-file Windows type checks also passed. The default whole-repository Windows type check has existing errors involving POSIX-only APIs in unchanged code.
uv lock --locked, scripts/generate_integration_manifest_docs.py --check, scripts/integration_guidance_skills.py --check-generated, scripts/release_version.py --check and git diff --cached --check passed.

Other host versions/platforms, HTTPS ingress and TUI reload have not been qualified. The documentation site build was not run locally. Specs, credentials and local acceptance output are excluded from this PR.

AI usage statement

OpenAI Codex (GPT-6) assisted implementation, documentation and validation.

Copilot