Kimi Code Configured Assets

Use this note for Kimi Code-specific configured-asset locations and evidence boundaries. Start from Agent Customize Routing for owner selection and Global Coding-Agent Assets for the shared inventory contract.

Static Inventory Route

<node> <better-harness-root>/scripts/agent-customize/cli.mjs inventory \
  --provider kimi \
  --workspace <absolute-target-path>

<node> <better-harness-root>/scripts/coding-agent-practices/asset-baseline.mjs kimi \
  --workspace <absolute-target-path> \
  --include-user-home \
  --format json

Use --kimi-home <path> for an isolated configuration root (default: ~/.kimi-code).

Source Map

Asset	User	Selected project	Installed Plugin
Instructions / Rules	None	AGENTS.md, CLAUDE.md (compat)	Plugin systemPrompt/systemPromptPath (plugin metadata only, never merged into rules)
Skills	~/.kimi-code/skills/**/SKILL.md	.kimi-code/skills/**/SKILL.md, .kimi/skills/**/SKILL.md (probed)	Plugin-declared skills roots under ~/.kimi-code/plugins/managed/<id>/ (root SKILL.md when undeclared)
MCP	~/.kimi-code/mcp.json#mcpServers	None	Plugin-declared mcpServers from kimi.plugin.json
Agents / Commands / Hooks	Custom agents (not yet inventoried)	Custom agents (not yet inventoried)	Plugin agents, commands, and hooks from kimi.plugin.json
Memory	No Kimi Code equivalent	No Kimi Code equivalent	No Kimi Code equivalent

~/.kimi-code/config.toml holds model/provider settings, not customizable assets; it is surfaced only as a diagnostics flag and never parsed into inventory items.

Installed plugins are indexed in ~/.kimi-code/plugins/installed.json; each record points at a managed copy under ~/.kimi-code/plugins/managed/<id>/. The provider reads kimi.plugin.json first and falls back to .kimi-plugin/plugin.json, inventories assets only for records with enabled: true (disabled plugins stay listed without component assets), and skips manifest-declared paths that escape the plugin root. Collection follows symlinks, so any collected file whose realpath escapes the plugin root (for example through a symlink inside it) is dropped from the inventory as well.

Skill Layout And Invocation

Repository-root skills/ is the Better Harness distribution layout, not a Kimi Code project-level convention. The provider deliberately does not probe it, so a workspace-only inventory can report skills: 0 for a repository whose skills are physically present in that root directory. A skill counts only after installation into user scope (~/.kimi-code/skills) or project scope (.kimi-code/skills, .kimi/skills).

Skills are invoked with /skill:<name> (for example /skill:better-harness) or automatically by the model, so the skill surface doubles as the main invocation surface. Plugins can additionally register namespaced slash commands (<plugin>:<command>) from a commands/ directory declared in kimi.plugin.json; commands declared by enabled plugins are inventoried into the Commands collection.

Session Evidence Boundary

Session evidence comes from scripts/session-analysis/platforms/kimi.mjs:

• Transcripts live at ~/.kimi-code/sessions/<wd_*>/ses{sion}_*/agents/<agent>/wire.jsonl; state.json next to them supplies title and created/updated timestamps.
• The workspace-to-wd_* mapping is resolved through ~/.kimi-code/workspaces.json (exact root match) and ~/.kimi-code/session_index.jsonl (per-session workDir). Only when both indexes are absent does the adapter fall back to wd_<name>_* directory prefixes, and it records a kimi-workspace-index-absent warning.
• Wire records are normalized from context.append_loop_event (tool.call/tool.result/content.part), turn.prompt/turn.steer, context.append_message (protocol 1.0), and usage.record token usage. Unknown record types degrade to bounded metadata.* events for forward compatibility (metadata.protocol_version is currently 1.4).
• User text, command text, and message content stay gated behind the shared includeUserText/includeCommandText/includeContent privacy flags.
• ~/.kimi-code/credentials and server.token are never read.

MCP And Privacy Boundary

The static collector reads mcp.json only. Output may include server name, transport, command, safe argument metadata, environment key names, and direct-secret-key warnings. It must not contain environment values, header values, URL credentials, or authentication state.