feat: expose responseLevel over MCP — Phase 4

[thymikee]

What

Phase 4: let MCP agents opt into the leveled response views that #942 added, mirroring how includeCost is exposed at the MCP boundary.

src/mcp/command-tools.ts:

• readClientConfig reads a responseLevel tool arg, validates it against RESPONSE_LEVELS ('digest' | 'default' | 'full'), and sets client.responseLevel only for valid values; rejects other non-undefined values with a clear error. (Extracted a small readResponseLevel helper to keep readClientConfig under the fallow complexity threshold.)
• stripMcpConfigFields strips responseLevel so it never leaks into the command input.
• withMcpConfigSchema advertises responseLevel: { type: 'string', enum: ['digest','default','full'], … } in every tool's inputSchema.

Default-off is byte-identical

With responseLevel absent, client.responseLevel is never set and the request shape is unchanged — same additive, MCP-boundary-only pattern as includeCost. Covered by a test asserting the absent path yields an empty config.

Verification

• tsc --noEmit 0; oxfmt + oxlint --deny-warnings clean; fallow audit --base origin/main clean; vitest run src/mcp 25 pass; Layering Guard empty.

Built by a delegated worker; rebased onto main after #942 merged. Tests: digest→config + arg stripped; absent→byte-identical; invalid rejected; inputSchema advertises the enum.

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.4 MB	1.4 MB	+552 B
JS gzip	455.5 kB	455.7 kB	+171 B
npm tarball	561.0 kB	561.2 kB	+190 B
npm unpacked	2.0 MB	2.0 MB	+552 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	28.1 ms	28.0 ms	-0.1 ms
CLI --help	49.3 ms	49.1 ms	-0.2 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/server.js	+9.3 kB	+3.7 kB

[thymikee]

Blocking this one for the MCP optimized text path.

When an MCP caller sets responseLevel: digest, structuredContent receives the digest payload from the daemon, but renderToolText() still sends that same digest object through formatCliOutput() for optimized text. For snapshot, the CLI formatter expects nodes; the digest view intentionally drops nodes and returns nodeCount/refs, so formatSnapshotText() falls back to an empty array and prints Snapshot: 0 nodes. That makes the default MCP content[0].text contradict the actual digest payload whenever the screen is non-empty.

Please add a shipped-path MCP test for snapshot with mcpOutputFormat: optimized plus responseLevel: digest, and either render non-default response views as JSON text or teach the snapshot formatter how to render the digest shape. Until then agents using the optimized MCP text get misleading state even though the structured content is correct.

[thymikee]

Fixed (pushed). You're right — the optimized text path mis-rendered the digest. renderToolText now bypasses the optimized CLI formatters whenever responseLevel is non-default (digest/full) and emits the leveled payload verbatim as JSON, so the text matches structuredContent instead of printing 'Snapshot: 0 nodes'. Added the requested shipped-path test: snapshot with mcpOutputFormat: optimized + responseLevel: digest → content[0].text is the JSON digest (and asserts it never emits the '0 nodes' line). tsc/oxlint/fallow clean, 26 MCP tests pass.

[thymikee]

Re-reviewed the fix at 7437029.

The previous MCP optimized-text blocker is addressed: responseLevel is validated/stripped at the MCP boundary, passed into the client config, and renderToolText() now emits JSON for non-default response levels instead of passing digest/full payloads through CLI formatters that expect default shapes. The added shipped-path test covers snapshot with mcpOutputFormat: optimized + responseLevel: digest and asserts the text is the JSON digest, not Snapshot: 0 nodes.

All 21 checks are green. I do not see an actionable blocker now.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-06-30 06:59 UTC