Skip to content

Follow up Kimi Code upstream updates: kimi.plugin.json plugin target (0.31.0 adds agents + dual-engine systemPrompt; only command surface) #2502

Description

@dyoshikawa

Summary

Kimi Code 0.31.0 (2026-07-30) widened the kimi.plugin.json plugin manifest to contribute agents and systemPrompt/systemPromptPath on both agent engines, alongside the existing skills/commands/mcpServers/hooks. A plugin target would now cover rules, commands, subagents, skills, mcp and hooks in one artifact — and the manifest remains the only command surface Kimi Code has, so commands stays unreachable for the plain kimi-code target.

Baseline: #2425 (closed) was verified at 0.29.1; its gaps were implemented in #2438, and it deliberately deferred the plugin manifest as "a new target rather than a gap in the existing one". 0.31.0 changes that calculus.

Recent Releases

  • 0.31.0 (2026-07-30) — plugin-contributed agents (an agents/ directory or explicit paths, auto-discovered and delegable as sub-agents, PR #2365); plugin systemPrompt / systemPromptPath (inline or file-referenced instructions, max 32 KB, "effective on both agent engines", PR #2314); Markdown custom agents on agent-core; /secondary_model. — releases
  • 0.30.0 (2026-07-29) — [status_line] in tui.toml (TUI presentation; no rulesync feature dimension).
  • 0.29.2 (2026-07-27) — no config-surface change.

Gaps

commands / rules / subagents / skills / mcp / hooks — kimi.plugin.json manifest surface unmapped

  • Upstream: as of 0.31.0, a plugin manifest contributes skills, commands, mcpServers, hooks, agents, and systemPrompt/systemPromptPath. — plugins doc
  • rulesync: the kimi-code target writes only the plain .kimi-code/ and ~/.kimi-code/ trees; commands is reported unsupported and there is no way to produce one (the plugin manifest is Kimi Code's only command surface). Nothing under src/ emits kimi.plugin.json.
  • Note: systemPrompt is additive instruction injection working on the TUI — unlike $KIMI_CODE_HOME/SYSTEM.md, which wholesale replaces the built-in system prompt and was correctly excluded as a rules surface in Follow up Kimi Code upstream updates: config.toml [mcp] default timeouts (0.29.1) and [tools] global enable/disable (0.29.0) #2425.

Proposed Follow-up

Add a kimi-code-plugin target following the claudecode-plugin / antigravity-plugin precedent, emitting kimi.plugin.json plus skills/, agents/, commands/, and hook scripts. Project scope only — plugins install per-user into $KIMI_CODE_HOME/plugins/managed/, so the generated tree is a source checkout the user installs from, not a live config path. Map rulesync rules to systemPromptPath rather than inline systemPrompt to stay under the 32 KB cap. Coordinate with the cross-tool plugin-packaging design (#329/#382) that already tracks the Copilot CLI / Codex plugin counterparts.

Non-gaps (verified at 0.31.0)

  • Hook events: still the same 16, all mapped in src/types/hooks.ts.
  • [tools] remains registered only on the v2 engine (packages/agent-core/src/config/schema.ts has no tools key at tag 0.31.0), so Follow up Kimi Code upstream updates: config.toml [mcp] default timeouts (0.29.1) and [tools] global enable/disable (0.29.0) #2425's caveat holds and rulesync's emission stays forward-compatible.
  • [status_line] (0.30.0): TUI presentation; rulesync has no statusline feature for any target.
  • model_preference / [secondary_model]: still experiment-gated; round-trips via looseObject.
  • Markdown custom agents on agent-core (0.31.0) retroactively validates the .kimi-code/agents/*.md output rulesync already produces.
  • Dry-runs clean: project 48 files; global 50 including ~/.kimi-code/config.toml for hooks and permissions.

References

Metadata

Metadata

Assignees

No one assigned

    Labels

    consideringenhancementNew feature or requestmaintainer-scrapRough notes for AI implementation. Not for human eyes.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions