Skip to content

AGENTS.md

Latest commit

 

History

History
86 lines (68 loc) · 20.8 KB

AGENTS.md

File metadata and controls

86 lines (68 loc) · 20.8 KB
Raw

Repository-level Agent Guide

Reply in the same language as the user.

This is a TypeScript monorepo built for agent-assisted development. Keep the root AGENTS.md limited to hot-path rules: the project map, hard constraints, and workflow requirements — things every task needs to know.

Working Principles

  • Think from first principles. Start from real requirements, code facts, and verification results; if the goal is unclear, discuss it with the user first.
  • Treat code, not documentation, as the source of truth. Unless the user explicitly says otherwise, do not read ordinary Markdown just to understand the implementation.
  • Before making code changes, read the relevant code and the most recent constraints, and follow the nearest AGENTS.md in the directory tree.
  • Keep changes focused. Do not slip in unrelated refactors along the way.
  • When committing, do not add any co-author attribution, and do not reveal the identity of the agent in commit messages, PR descriptions, or any explanatory text.

Project Map

  • apps/kimi-code: the CLI / TUI application. It consumes core capabilities through @moonshot-ai/kimi-code-sdk and must not depend directly on @moonshot-ai/agent-core. When writing or modifying its terminal UI, use the write-tui skill (.agents/skills/write-tui/SKILL.md).
  • apps/kimi-web: the browser web UI, a peer to the TUI. Vue 3 + Vite + vue-i18n; talks to the server over REST + WebSocket under /api/v1. It must not depend on @moonshot-ai/agent-core (wire types are re-implemented locally). Debug against the two engines via the root pnpm dev:v1 / pnpm dev:v2 backend scripts — the dev Sidebar shows the active backend and switches it at runtime. See apps/kimi-web/AGENTS.md.
  • apps/vis, apps/vis/server, apps/vis/web: visual debugging tools for sessions and replays.
  • apps/kimi-inspect: web inspector for the kap-server /api/v1/debug RPC surface — workspace/session browser, per-session chat, and Service panels (data + trigger buttons) for the Session and Agent scopes. A left icon rail (src/components/NavRail.tsx) switches top-level views: the Chat workspace, the Model Catalog (src/components/ModelCatalogView.tsx — every Provider with its Models and the default marker, via IModelCatalog / IModelService channel proxies), and App Services (src/components/AppServicesView.tsx — the app-scope Service reflection, full width; the Agent scope stays in the Chat view's right dock (src/components/RightPanel.tsx) across two tabs: the Agent tab (Inspector: agent switcher + a Plan lookup card — PlanCard in src/components/Inspector.tsx — querying GET /sessions/{id}/transcript/plan (one tool_call_id, or every plan of the agent) via src/transcript/api.ts's fetchTranscriptPlan — plus the agent Service panels) and the State tab (every key an Agent Service registered into the agent-state container, polled live via IAgentStateService.snapshot() — the same live diff-tree view as the session State tab, sharing StateCard from src/components/StateCard.tsx), while the Session scope has its own column right next to the session-list sidebar (src/components/SessionPane.tsx) with two tabs: Services (the pending-interactions card — src/components/InteractionsCard.tsx — plus the session Service panels) and State (every key a Session Service registered into the session-state container, read on demand via ISessionStateService.snapshot())). Expanding a Model opens the model inspector inside that view: provider/model config layers plus the resolved runtime view with per-value provenance (config / override / builtin / env / synthesized), served on demand by IModelCatalog.inspect — the same resolution pass the runtime's get serves, traced via ResolutionTraceCollector and assembled by kosong/model/inspection.ts. Built on its own old-klient-style channel layer (src/channel/: the VS Code ProxyChannel model — service-bound IChannel, HTTP ProxyChannel for calls routed to /api/v1/debug), typed by agent-core-v2 Service interfaces; GET /api/v1/debug/channels loads the whole wire protocol 1:1 (every scoped Service, no whitelist). There is no Service-event push channel: panels fetch/refresh on demand (Sidebar polls react-query on a 15 s interval), and a connection failure shows a blocking "Debug surface unavailable" screen instead of falling back anywhere. Session-level coarse status is the one exception: src/activity/ holds a second /api/v1/ws client (GlobalEventsWs) that subscribes to nothing and consumes the server-pushed global facts — event.session.work_changed updates a per-session activity map (SessionActivityHub + subscribe/version store, seeded on connect/reconnect from GET /api/v1/sessions), while event.session.created / session.meta.updated invalidate the ['sessions'] query; the Sidebar session rows render running / approval / question / failed badges from it via useSessionActivities. The Vite dev server proxies /api to a running kap-server (KIMI_SERVER_URL, default http://127.0.0.1:58627) and exposes GET /__inspect/servers (vite/serverDiscovery.ts), which scans the local kap-server instance registry (~/.kimi-code/server/instances + legacy lock) and the home token so the app can zero-config auto-connect and switch servers from the header dropdown at runtime. The per-session chat (src/components/ChatView.tsx) renders turn-granularly from the transcript surface instead of context memory: full state is read from GET /api/v1/sessions/{id}/transcript (initial load = newest page, refreshes re-read from the tail backwards), older history auto-pages with before_turn via an IntersectionObserver sentinel at the top of the scroll view, and each timeline item is wrapped in content-visibility: auto + contain-intrinsic-size so the browser virtualizes off-screen rendering natively (no windowing library); /api/v1/ws is an incremental channel (transcript.ops, grade block — the cheapest grade that still carries whole-state frame upserts, dropping per-token append frames; transcript.reset is ignored by the store, surfaced only to the audit recorder via the optional onReset handler). The channel tracks the op-batch watermark: a dedicated subscribe_v2 control frame carries the per-agent grades and the transcript_since cursor, a seq gap / reconnect / resync_required / append gap triggers a point-to-point catch-up (fetchTranscriptOps → GET .../transcript/ops?since_seq=), and any legacy/incomplete answer falls back to the full REST refresh. Convergence reuses @moonshot-ai/transcript's L2 reducer (src/transcript/: REST/WS clients + store; the data model and reducer come from the package, nothing is re-implemented locally). The Transcript audit panel (src/components/audit/, the Audit tab of the chat view's right dock — src/components/RightPanel.tsx, fed the trail by ChatView's onTrailChange) replays how the visible store was built: an AuditTrail (src/audit/) records every step — each REST page (request + replace/prepend), every WS frame (transcript.ops live/buffered/flushed/catchup, transcript.reset), loss signals, and prompt/cancel actions — with the resulting immutable AgentState per entry; the panel offers a draggable timeline plus a Diff tab (structural diff vs the previous entry: added/modified/removed colored, long strings tail-truncated, all fields kept), a full State view, and the raw Event payload.
  • packages/agent-core: the unified agent engine, including Agent, Session, profile, skills, tools, plan, permission, background, records, the in-process DI service layer (src/services/), and other core capabilities.
  • packages/node-sdk: the public TypeScript SDK and harness.
  • packages/kosong: the LLM / provider abstraction layer.
  • packages/kaos: the execution environment and file/process abstractions.
  • packages/oauth: Kimi OAuth and managed auth utilities.
  • packages/telemetry: shared client-side telemetry infrastructure.
  • packages/transcript: the isomorphic transcript rendering data layer — agent-granular L1 store, idempotent L2 operations, off/turn/block/delta L3 subscription granularity, framework-free L4 view registry, and turn-cursor pagination. Pure TypeScript (browser-safe, no engine imports) and the sole owner of all transcript contract types (src/contract/); consumed by packages/kap-server (engine events → transcript, REST + WS surface; live stores backfill history from the persisted per-agent wire records — main on first attach, any agent on demand, cold sessions rebuild any agent — with 0-based turn ordinals matching the engine's). The cold rebuild is a two-level fold over wire.jsonl as the single source of truth: history/groupTurns.ts (context messages → turn tree) plus history/foldFacts.ts (non-context records → tasks, interactions, todos, goal/plan/swarm meta, and end-appended markers/taskrefs; interactions left pending at shutdown fold to cancelled). Plan content is a recorded fact too: each ExitPlanMode review submission offloads the document to agents/<agentId>/plan/<planId>/v<N>.md and persists a reference-only plan.revision record ({id, version, path, sha256, bytes}), which projects — live and cold — to a plan.revision marker and the modes.plan badge ({reviewPath, version}). It also owns the op-batch sequencing contract (transcriptSeqSchema in contract/schema.ts): a per-(session, agent) monotonic batch seq on transcript.ops / transcript.reset / the REST transcript response, the transcript_since subscription cursor, and the GET .../transcript/ops catch-up response shape — every field optional so pre-seq peers fall back to loss-signal-driven refreshes. Beyond the timeline, the model carries wire-equivalent detail: steps carry usage / finishReason / timing (LLM latencies) / retry / interrupt reason, turns carry durationMs / error / usage, tool frames carry the streamed inputText and the latest progress, tasks carry subagent resultSummary / error / stateReason / usage, meta.agent mirrors the agent status slices (model / usage / context / permission / phase), a global prompts entity (op prompt.upsert) tracks the prompt queue, and hook.result lands as a 'hook' marker. These live-projected fields are NOT backfilled by the cold rebuild (known limitation).
  • packages/kap-server: the Kimi Code server, backed by the DI × Scope agent engine (@moonshot-ai/agent-core-v2). Exposes sessions over REST + WebSocket (/api/v1 + /api/v1/ws); bootstrapped from src/start.ts and consumed by apps/kimi-code. The RPC surface is /api/v1/debug/* — a reflection dispatcher over the ENTIRE scoped DI registry (every Service callable, no whitelist; src/transport/registerDebugRoutes.ts + serviceDispatcherRoutes.ts), mounted only with --debug-endpoints on a loopback bind and gated by the global bearer auth; repo dev scripts pass the flag. Its transcript surface implements the op-batch sequencing contract: TranscriptService.dispatchOps assigns every dispatched batch a per-agent consecutive seq and retains it in a bounded in-memory journal (TRANSCRIPT_OPS_JOURNAL_CAPACITY, dies with the live store); WS transcript.ops/transcript.reset payloads carry the seq/watermark, a transcript_since subscription cursor (carried, with the per-agent grades, by the subscribe_v2 control frame — the only transcript subscription channel; its agent-grained counterpart unsubscribe_v2 detaches listed agents' streams, or the whole session's when agent_ids is absent, letting the detached agents' legacy events flow again) replays journaled batches instead of a baseline reset when the journal covers it, and GET /sessions/{id}/transcript/ops?since_seq= serves point-to-point catch-up (complete: false = journal can't cover or session cold → caller falls back to a full refresh). Beside the paged route, GET /sessions/{id}/transcript/plan?agent_id=[&tool_call_id=] projects an agent's ExitPlanMode plan info (content / path / options / review outcome; tool_call_id narrows to one call, omitted lists every recoverable plan) from the first available fact — the linked approval interaction's persisted request display, the live tool frame's display, or the tool result output text. The baseline transcript.reset itself is items-empty (TRANSCRIPT_RESET_TAIL_TURNS = 0): it carries only global state + the watermark + has_more_older, because history always pages in over REST. When a WS connection subscribes to the transcript protocol (grade ≠ off for an agent), the broadcaster suppresses the transcript-projected session_event types for that connection × agent (TRANSCRIPT_PROJECTED_EVENT_TYPES + suppressedByTranscript in sessionEventBroadcaster.ts; cursor replay via getBufferedSince applies the same filter). Suppression is only a per-connection send view — the journal still records everything, and connections without transcript grades are unaffected. The session's work aggregate behind event.session.work_changed (busy / main_turn_active / pending_interaction / last_turn_reason) is owned by the core's ISessionActivityView (sessionActivity domain, Session scope): the broadcaster only schedules the wire emission around turn frames (busy:false lands after turn.ended), and resolveSessionFacts (src/routes/sessions.ts) reads the same view — never fold per-agent activity at the edge. Delivery split on /api/v1/ws: global events (session.meta.updated and the event.session.* / event.workspace.* / event.config.* families, including every activated session's event.session.work_changed) fan out to EVERY established connection — WsConnectionV1 registers itself via broadcaster.addGlobalTarget on construction and unregisters on close — while session/agent-grained events only reach connections subscribed to that session (subject to agent_filter and the transcript suppression above); transcript frames are a separate channel governed by the per-agent grades alone and bypass agent_filter entirely.
  • packages/klient: the client SDK — a contract-driven facade over agent-core-v2 with aggregated global.* / session(id).* / agent(id).* methods, zod validation on every call, and klient-level typed event forwarding. Transport is chosen once at creation via subpath entry (@moonshot-ai/klient/ipc|memory); both return the same Klient. The package also hosts the e2e suites: the legacy /api/v1 live suites (test/e2e/legacy/) and the docker e2e runner (pnpm --filter @moonshot-ai/klient docker:e2e). See packages/klient/AGENTS.md.
  • packages/server-e2e: live e2e tests and scenarios against a running server (KIMI_SERVER_URL, default http://127.0.0.1:58627). See packages/server-e2e/AGENTS.md.

Environment Requirements

  • Node.js: >=24.15.0 (from the root package.json engines; .nvmrc is 24.15.0, used by nvm / fnm / mise to pick the minimum recommended version).
  • pnpm: 10.33.0 (from the root package.json packageManager).
  • pnpm install will fail when the Node version is not satisfied, because .npmrc sets engine-strict=true.

Monorepo Workspace Maintenance

  • pnpm-workspace.yaml is the source of truth for workspace membership, but flake.nix also contains hardcoded workspacePaths and workspaceNames lists.
  • Whenever you add or remove a workspace package, you MUST update both pnpm-workspace.yaml and flake.nix — for every package, including leaf / test / e2e packages that nothing depends on.
    • pnpm-workspace.yaml uses globs (packages/*, apps/*), so most packages land there automatically; flake.nix is fully manual and is where omissions happen.
    • Missing a path in flake.nix's workspacePaths will silently drop files from the Nix build's src fileset.
    • Missing a name in flake.nix's workspaceNames will break pnpmConfigHook because dependencies for that workspace will not be fetched.
  • The automated "Check flake.nix workspace sync" (scripts/check-nix-workspace.mjs) only validates the transitive dependency closure of @moonshot-ai/kimi-code. A leaf package outside that closure (e.g. an e2e package nobody imports) slips through even when it is missing from flake.nix. A green check is therefore NOT proof that flake.nix is fully in sync — keep it updated by hand on every add/remove, do not rely on the check to catch omissions.

General Coding Rules

  • For optional object properties, pass undefined directly instead of using conditional spread.
    • YES: { user }
    • NO: { ...(user ? { user } : undefined) }
  • Optional object properties do not need to additionally allow undefined in the type.
    • YES: interface Options { user?: User }
    • NO: interface Options { user?: User | undefined }
  • Internal methods with only a single parameter should not be turned into options objects just for stylistic uniformity.
  • Except for a package's index.ts, other index.ts files should prefer export * from './module';.
  • The Agent class in packages/agent-core/src/agent must be usable on its own. The constructor must not force the caller to create a Session instance, nor require an agentId or session. It may accept an optional sessionId as a request-config hint — for example mapped to the provider's prompt_cache_key — but the instance must not hold sessionId, and must not depend on the Session lifecycle, metadata, or parent/child relationship logic.
  • Do not add too many new test files. Prefer adding tests to the existing test file of the corresponding component or module.
  • When a test fails because of a user modification, default to fixing the test first; do not change the implementation to satisfy an old test unless the implementation truly has a bug.
  • Do not sacrifice code quality for external compatibility unless the user explicitly asks for it. Breaking changes go through changesets and a major bump, gated by the rule below.

Experimental Features

  • Gate a not-yet-public feature behind an experimental flag. Add the flag to the registry at packages/agent-core/src/flags/registry.ts, then check it with flags.enabled('my-feature'). Flags are env-driven and default off: KIMI_CODE_EXPERIMENTAL_<NAME> toggles one, KIMI_CODE_EXPERIMENTAL_FLAG enables all. Release by flipping the entry's default to true.

Where to Update Instructions

  • Hard rules that affect almost every task: update the root AGENTS.md.
  • Rules that only affect a specific directory: update the nearest sub-directory AGENTS.md.
  • Keep instruction updates focused and supported by code facts.

Workflow Requirements

  • Prefer rg / rg --files when reading code.
  • When designing changes, follow existing boundaries and local patterns first.
  • In public text and test data, replace real internal identifiers with neutral placeholders such as example.com, example.test, and YOUR_API_KEY. Before opening a PR, ask a read-only agent to audit the diff for context-specific internal identifiers.
  • When creating a PR, the PR title must follow Conventional Commit style, e.g. chore: remove legacy format commands.
  • When an AI agent opens or updates a PR, fill in .github/pull_request_template.md — link the related issue or explain the problem, then describe what changed. Do not leave placeholder text or submit a generic summary of the diff.
  • Do not submit vague AI-generated PR text. The human author must understand the change well enough to explain the code, edge cases, and why the approach fits this repository.
  • After finishing a task and before submitting a PR, you must run the gen-changesets skill (see .agents/skills/gen-changesets/SKILL.md) and generate a changeset under .changeset/ according to its rules.
  • When generating a changeset, never decide on a major bump on your own. When you judge a change to meet the major criteria (breaking changes, incompatible user configuration, renamed or removed commands/arguments, changed behavior semantics, etc.), you must stop and explain it to the user and ask for confirmation. Only write major after the user has explicitly agreed. Otherwise default to minor (and fall back to patch if minor is unclear). See the "Hard rule: confirm with the user before writing major" section in .agents/skills/gen-changesets/SKILL.md for details.
  • Prefer importing via import ... from '#/...', which serves the same purpose as import ... from '@/...'.
  • Do not commit throwaway scratch or exploratory files. Never stage:
    • Agent working notes or handoff/summary documents (e.g. HANDOVER-*.md, HANDOFF-*.md, handoff.md).
    • Throwaway UI/UX prototypes or design mockups (e.g. *-designs.html, *-mockup.html, *-demo(s).html) at the repo root or under a design/ folder. The only tracked .html files should be Vite index.html entrypoints. Before committing or opening a PR, run git status and git diff --staged --stat and remove anything matching these patterns. Put scratch work under .tmp/ (gitignored) instead of the repo root or the source tree.