[ytvee]

Problem and result

Kimi discovered WebDev skills but lacked a complete native integration. This release candidate packages 23 skills, nine bounded roles, startup policy, opt-in capture, and a Kimi proposal/text-evaluation engine.

The follow-up audit found functional and privacy problems in the initial PR. They are corrected in this branch; authenticated runtime acceptance remains required before publication.

Audit findings and fixes

Finding	Fix
Project .agents/AGENTS.md is not a Kimi startup instruction location	Add preview/apply/remove helper for a managed pointer in .kimi-code/AGENTS.md; preserve surrounding host instructions, use atomic writes and unique backups.
Installing a working .agents as a plugin copies project facts, memory and attachments into the user-wide plugin cache	Require a clean release payload in staging; document separate project/plugin modes and safe migration without deleting project state.
Local roles can shadow upgraded plugin roles; skills can be duplicated	Document one runtime installation mode and explicit vendor reconciliation.
UserPromptSubmit carries content parts; failures use PostToolUseFailure	Capture text parts only, subscribe to failures, normalize native IDs/errors, mark missing assistant text and potentially truncated tool output.
Removing or updating a managed TOML block could change the scope of adjacent settings	Compare parsed configuration semantics and refuse changes to unrelated data; retain existing hooks and unique private backups.
Separate evaluator launches could use different models or misreport the active TUI model	Add --kimi-model, pin one configured alias across both A/B arms, snapshot config/environment fingerprints, reject drift, record configured rather than provider-observed identity.
Relative/empty KIMI_CODE_HOME differed from subprocess/Kimi semantics	Honor the default for empty values and pass an absolute home to the isolated subprocess.

The evaluator denies tools and subagents, uses JSON-encoded file-based evidence with template-variable protection, rejects error/tool stream records, and does not pass large evidence through argv. Visual attachments remain metadata-only. Role names do not select models; existing Kimi secondary-model settings control routing.

Evidence reviewed

Official Kimi documentation and source at MoonshotAI/kimi-code 244819a:

• Project instruction discovery
• Plugin copy behavior
• Hook payload generation
• CLI reference, hooks, agents, plugins.

Verification

• Full python scripts/validate_skill_pack.py: source, graph, budgets, generated targets, archives and cross-client contracts passed.
• Existing learning suite: 104 tests, one skip. Extended coverage for text parts, failed tools, truncation, TOML preservation, startup pointer, model pinning/drift, error streams and 200 KiB literal-template evidence.
• Both changed skill validators, Markdown links and git diff --check passed.
• Actual helper CLI preview/install/reinstall/remove passed in temporary paths containing spaces; no live user configuration changed.
• Independent read-only review completed; reported path-resolution findings fixed.
• GitHub Actions on audited implementation head 60362bd: quality-ci link-check, and skill-pack-ci all passed, including native helper regressions on Ubuntu and Windows. Audited implementation tree: d4482889153f78abb4f2d111bbe6524594dc71c8.
• Rebuilt all 12 release archives and validated v1.10.0 release notes after the fixes.

Installation documentation follow-up

Added a task-based MCP checklist (Context7, Playwright, GitHub, conditional Figma), CLI/VS Code configuration steps and a credential-free example. Documented Desktop browser, Browser Extension and Computer Use setup, Windows/macOS requirements, model/permission choices, main-session browser verification, and current subagent limitations. Official provider documentation is linked in the guide.

Documentation head: d771261. Full local pack validation, install-guide validation, links and whitespace checks passed. quality-ci and link-check passed for this documentation-only commit; skill-pack-ci is still running. No client configuration or credentials were changed.

Remaining release gates

Authenticated current Kimi v2: startup/skill discovery, a real bounded role handoff, capture on/off and isolation, proposal generation and same-model A/B. Live Kimi execution on Windows also needs observed acceptance; helper regression tests already pass there. Synthetic checks are not live Kimi certification.

Kimi Stop exposes no final assistant response; tool output is capped upstream. These are documented client limits, not full-transcript capture. Text benchmarks cannot certify real coding behavior.

Scope and rollout

Version 1.10.0 metadata and release packaging are prepared; PR remains draft. No merge, tag or release publication. No new dependency or test infrastructure. Existing host instructions, model credentials and capture consent are preserved. Normal frontend work or installing the package must not enable learning/capture.