kimi-code/packages/server-e2e/scenarios

Scenarios

Each .ts file under this directory is an executable wire-level test of a single user-facing flow against a running server.

Running

Default base URL is http://127.0.0.1:58627. Start a server (pnpm dev:server from the repo root) before invoking scenarios.

Scenarios import directly from the package's src/*.ts so they need a TS loader. tsx (a workspace devDependency) works out of the box:

# Single scenario
npx tsx packages/server-e2e/scenarios/01-create-and-send.ts

# All scenarios (sequential; first failure exits non-zero)
pnpm --filter @moonshot-ai/server-e2e test:scenarios

# Custom server URL
KIMI_SERVER_URL=http://127.0.0.1:8080 npx tsx packages/server-e2e/scenarios/02-tool-call-with-approval.ts

test:scenarios writes reports/latest/index.html with the scenario timeline, including stdout/stderr milestones, HTTP request / response envelopes, and WebSocket frames. Set KIMI_SERVER_E2E_REPORT_DIR=/tmp/server-e2e-report to write it somewhere else.

Catalog

File	What it does
_template.ts	Copy-paste starting point. No assertions; smoke-tests the lifecycle.
01-create-and-send.ts	Happy path: create session → submit prompt → assert assistant replied with the expected token.
02-tool-call-with-approval.ts	Drives a Bash tool call; built-in approval handler auto-approves; asserts canary round-trips through tool_result AND assistant text; asserts session ends in idle.
03-refresh-replay.ts	"User refreshes the browser" worst case: Phase 0 probes (/healthz, /meta, /auth) → WS handshake → prompt to populate the ring buffer → fresh WS with last_seq_by_session (caught-up first, then 0 for full replay) → REST snapshot → steady-state follow-up prompt. Asserts replay ordering (seq 1..N) and that no resync_required fires while the buffer covers the gap.
04-stateless-controls.ts	Per-request stateless prompt controls diff-dispatch into the session shadow, including plan mode and permission mode transitions.
05-workspace.ts	Workspace registry + folder picker happy path: fs:home → fs:browse $HOME → POST /workspaces { root } → POST /sessions { workspace_id } → GET /sessions?workspace_id= → prompt round-trip (skipped without provider auth) → DELETE /workspaces/{id} (verifies the session survives).
06-model-catalog.ts	Model/provider catalog reads and current-default :set_default round-trip.
07-session-children.ts	Direct child session creation, child prompt execution, direct-child listing, and missing-parent 40401.
08-pending-recovery.ts	Pending approval and pending question recovery APIs, including resolve and pending-list cleanup.
09-image-file-prompts.ts	Uploaded file prompt references: missing file, non-image validation, and PNG prompt submission.
10-prompt-queue-steer.ts	Prompt queue steer: debug-inject an active prompt, queue two prompts, steer them through POST /prompts:steer, and assert REST, WS, content, and queue-drain behavior.
11-terminal.ts	Terminal flow: create/list/get a session terminal, attach over WS, write input, observe output, resize, close, and assert terminal exit/final state.
12-send-and-cancel.ts	Send prompt + cancel prompt: happy-path completion, abort a queued prompt by id, session-level abort of an active prompt, and scheduler recovery after aborts.

Writing a new scenario

Copy _template.ts and fill in the TODO block.

Conventions:

• Exit 0 on pass; non-zero on any assertion failure or unhandled rejection.
• Always try { ... } finally { close + delete session }.
• Print ▶ for milestones and ✓ / ✗ for the final outcome — test:scenarios greps for those prefixes when surfacing CI logs.
• Default timeouts to 60s; tool-call scenarios may want 120s.
• Use client.onApprovalRequested / client.onQuestionAsked to auto-resolve reverse-RPC requests — bypassing them risks 60s server-side timeouts that look like flaky scenarios.