Repository navigation
[ponytail] MCP server: generate inputSchema from zod, single reflect/explore renderer, dispatch and worker-pool trims #119
Description
Activity
PR: #137 (branch
ponytail/mcp-server, not merged).- Before: 131 files / 1689 tests passing (main at fa8219f). After, with origin/main merged in: 133 files / 1702 tests passing, build clean.
- All four findings done. No finding skipped. Two deliberate schema differences, both reported by the comparison test and in the CHANGELOG: unbounded
.int()fields advertiseintegerinstead ofnumber, andingest_turntool_calls[]items no longer claimadditionalProperties: false(never enforced). No schema was made.strict()because Hermes' warm-up recall sends an undeclaredlimit. - Removed test: dispatch.test.ts "extracts the session id from a recall_session result" (covered the deleted
extractSessionId). - Extra verification: built server on port 9917 answered
/health,initializeandtools/listwith 16 tools, everyinputSchemapopulated and closed (output in the PR). The 9907 daemon was not touched. - Follow-up noticed, out of scope:
interfaces/hermes-plugin/tests/test_stdio_e2e_real_server.pybridges to a healthy 9907 daemon (the provider spawnsserver.jswithout--standalone), so it is not hermetic and fails when the fixture memory already exists there.
- added a commit that references this issue
on Sep 21, 2026 Merged PR #137 into main with --no-ff (commit b5ef060). Post-merge on main: build clean, 133 test files / 1702 tests passing (was 131 / 1679; +23 from the schema-drift snapshot and renderer tests, -1 for the deleted extractSessionId case). server.ts 2220 to 1414 lines. Follow-up filed separately: the Hermes stdio e2e pytest needs ENGRAM_MCP_STANDALONE=1 to stay hermetic.
Cut
Shrink the MCP server module by generating tool input schemas from the zod schemas already defined in the same file, and by rendering each result shape once.
Findings
MCP_TOOL_DEFINITIONS(src/interfaces/mcp/server.ts:389) restate the zod schemas defined 250 lines above. zod is^4.6.5, so keep the name / description / annotations table and generateinputSchemawithz.toJSONSchema(Schema). Compare the generated schema against the hand-written one for every tool before deleting (a snapshot test that fails on drift is acceptable), and keep anydescriptionstrings the zod schemas lack by adding.describe()to them.formatReflectXmlandprintReflectResultrender the sameReflectResulttwice, and the explore result is formatted inline in two more places (src/interfaces/mcp/server.ts:1993). One renderer with a format switch (~120 lines).extractSessionIdregex-scrapes the session id back out of the XML this process just serialised (src/interfaces/mcp/dispatch.ts:131). Return it onToolResultinstead (~10 lines).hangMultiplier,readyTimeoutMs, generation counters, three custom Error classes,finishInflight(slot)whoseslotisvoid-ed,respawn(reason)whosereasonisvoid-ed (src/interfaces/mcp/worker-pool.ts:56, ~90 lines). Keep the timeout plus respawn-on-exit; keep the 2x-timeout kill documented inCLAUDE.md.Verification
npm run buildandnpm run test:rungreen before and after.tests/interfaces/mcp/tool-schemas*.test.ts(or equivalent) proves every tool's generatedinputSchemais equivalent to the previous hand-written one. The Hermes plugin checker (interfaces/hermes-plugin/tests) still sees 16 tools.