Kitsoki MCP Debug

Use kitsoki mcp-test as the first validation surface for studio MCP work. It spawns the studio MCP server over stdio with the official Go MCP SDK client, so it exercises the same transport boundary an attached coding agent uses without requiring Claude/Codex to reload its MCP tool list.

Default Smoke

Run from the repo root:

GOCACHE=$PWD/.cache/go-build go run ./cmd/kitsoki mcp-test --stories-dir ./stories --timeout 20s

Expected behavior:

• the child server prints kitsoki: studio MCP server on stdio ... on stderr
• the JSON report has "ok": true
• tools includes studio.ping, studio.handles, story.validate, session.new, and render tools
• tool_runs contains successful studio.ping and studio.handles calls

Use a repo-local GOCACHE in sandboxed environments. Remove .cache/ before committing unless it is already ignored.

Test One Tool

Use --tool and a JSON object in --tool-args:

GOCACHE=$PWD/.cache/go-build go run ./cmd/kitsoki mcp-test \
  --stories-dir ./stories \
  --tool story.validate \
  --tool-args '{"dir":"stories/bugfix"}'

Other useful calls:

# read-only server shape, useful for meta-mode/Q&A surface checks
GOCACHE=$PWD/.cache/go-build go run ./cmd/kitsoki mcp-test --read-only

# workspace-bound authoring tools
GOCACHE=$PWD/.cache/go-build go run ./cmd/kitsoki mcp-test \
  --workspace stories/bugfix \
  --tool story.graph \
  --tool-args '{}'

# point at a built binary instead of go run's current executable
go build -o /tmp/kitsoki-mcp-test ./cmd/kitsoki
/tmp/kitsoki-mcp-test mcp-test --server-command /tmp/kitsoki-mcp-test --stories-dir ./stories

mcp-test defaults to spawning its current executable with generated mcp args. Use repeated --server-arg flags only when the default generated server args are not appropriate; they replace the generated mcp argument list.

No-LLM Boundary

The default smoke is no-LLM:

• mcp-test only initializes, lists tools, and calls deterministic tools
• kitsoki mcp defaults driving sessions to harness:replay
• do not use session.new with harness:"live" unless the user explicitly asks for a live integration test
• tests must use replay/cassettes or in-process SDK transports, not a real LLM

Local Test Targets

For changes in the CLI wrapper:

GOCACHE=$PWD/.cache/go-build go test ./cmd/kitsoki -run 'TestMCP|TestRunStudioMCPTest|TestCLI_TopLevelHelp'

For studio server/tool behavior:

GOCACHE=$PWD/.cache/go-build go test ./internal/mcp/studio

If go test ./cmd/kitsoki fails on sandboxed runs with ~/.kitsoki writes or Unix socket bind errors, treat that as unrelated unless the touched code is in those areas. Keep verification focused and report the sandbox blocker.

Debugging Failures

• mcp-test: connect: the child process did not initialize as an MCP server. Run with a longer --timeout, check the child stderr, and verify the server args start with mcp.
• missing tool in tools: inspect registration in internal/mcp/studio/server.go and the relevant register*Tools method.
• tool call returns "is_error": true: read the text/structured content in the JSON report; studio tools return typed error payloads such as NO_WORKSPACE, BAD_REQUEST, or UNKNOWN_HANDLE.
• story.* path surprises: pass --stories-dir ./stories for @kitsoki/<name> resolution and pass explicit dir or --workspace when testing workspace tools.
• image/render behavior: start with render.tui before render.tui_png or render.web; render.web may degrade when no browser-capable host is wired.

Visual MCP / Web QA

Use visual MCP when the claim is about what a real operator can see or point at in the web surface. For Slidey deck QA, the high-signal path is:

1. Open a no-LLM/replay session with the story cassette or flow seed.

2. Drive to the rendered deck or review state.

3. Call visual.open with kind:"web" and a query that deep-links the single-page app route:

   {
     "kind": "web",
     "session_id": "<session>",
     "query": {
       "route": "/s/<session>/chat",
       "visual_annotate": "1"
     }
   }

4. Call visual.observe and verify regions include the relevant surface: media, deck, and annotation for Slidey deck annotation work.

5. If visual.act is involved, verify the action is executable through the returned semantic handles. Treat non-actionable semantic text as an MCP/UI bug, not as a user workaround.

Use a local DB for repeatable command-line MCP visual smoke when the shared DB is read-only:

GOCACHE=/private/tmp/kitsoki-go-cache go run ./cmd/kitsoki mcp-test \
  --stories-dir ./stories \
  --timeout 20s \
  --server-arg mcp \
  --server-arg --stories-dir --server-arg ./stories \
  --server-arg --db --server-arg .artifacts/visual-mcp-smoke/sessions.db

Slidey-specific traps:

• visual.open must capture the session route, not just the runstatus observer landing page.
• The deck annotator should be open when testing anchor/refine flows; use visual_annotate=1 or the specific media handle.
• Baked demo decks rely on ignored generated files under stories/slidey-edit/baked/. If an isolated worktree shows a missing deck.html, check whether the main checkout has the ignored artifact before changing tracked story logic.
• Prompt args that pass a scene object must serialize it explicitly. If the prompt renders map[...], add a string world key such as scene_json and test the Starlark output.
• Sidecar-promoted videos are not live Slidey embeds. Keep the live picker gated to slideshow/deck media so mp4 + semantic sidecars still use the overlay path.

Monitoring / checking a RUNNING session

You cannot check on a job another connection/process is driving via a second session.status/studio.handles call:

• the MCP CLIENT serialises tool calls per stdio connection (a second call waits behind the in-flight session.drive), and
• kitsoki sessions are per-process — a different kitsoki mcp process can't see another's handles.

The studio SERVER is NOT the bottleneck — it's provably concurrent (guard test internal/mcp/studio/session_concurrent_readonly_test.go; ticket 2026-06-25T121622Z-… is wontfix). So don't go hunting for a server-side lock.

Instead read the trace from disk: kitsoki trace status <trace|id> [--json] — one-shot, cross-process, prints {state, turn, status, last_error, cost, idle} and flags ⚠ STALLED. This is the supported way to watch a live MCP-driven run.

Model selection is a session.new {profile} parameter, not a story edit: a profile that pins a model (e.g. codex-native → gpt-5.5, synthetic-claude → GLM) SUPERSEDES the story agent-def model: (internal/host/agents.go). The bare codex profile pins nothing and falls back to the agent-def — that's the trap.

Before committing, run git diff --check and confirm only intended files are staged. Leave unrelated untracked files alone.