fix(mcp): advertise whole-string ID patterns so connector clients accept returned IDs [v2]

[jameswyse]

Summary

MCP tool schemas advertise prefixed IDs with prefix-only patterns such as ^dpl_, which Effect's Schema.isStartsWith emits. That is valid JSON Schema, but the OpenAI connector used by Codex matches pattern against the whole string. It therefore rejects the IDs Executor returns: skills deployment and profile selection, and resume request IDs.

This PR replaces the prefix checks in the shared SDK Id helper and ElicitationRequestId with one whole-string check, ^<prefix>_[\s\S]+$. That check drives both runtime validation and the advertised schema.

-Schema.check(Schema.isStartsWith(`${prefix}_`), Schema.isMinLength(prefix.length + 2)),
+Schema.check(Schema.isPattern(new RegExp(`^${prefix}_[\\s\\S]+$`))),

• Runtime validation: unchanged. The new check accepts exactly the strings the old startsWith plus minLength checks accepted, which I compared across 78 boundary samples in six ID families.
• Error text: an invalid ID now reports the RegExp instead of "starting with".
• AccountId (acc_): unchanged. Released app protocol snapshots freeze its schema, and it never appears in an MCP tool input.

Linked issue

Fixes #2186

Verification

• bun run check (format, lint, typecheck, apps:protocols, e2e:check). v2 has no bun run test script.
• e2e, both run against the unfixed code first:
  • local-skills.spec.ts (local): expected 'dpl_1302…' to match /^(?:(?:^dpl_))$/u before the fix, passes after. The scenario now checks the returned deployment and profile against the advertised skills patterns as whole strings, then calls skills with both.
  • pat-mcp.spec.ts, "PAT MCP approvals bind continuations…" (self-host): expected 'apr_4c23…' to match /^(?:(?:^apr_)|(?:^elc_))$/u before the fix, passes after. Its existing approval and token-binding assertions still pass.
• Real connector A/B. Self-host images were built from 828a680 and from this branch, with the same data and the same temporary ChatGPT connected app, called from Codex. The connector's tools were refreshed after the image swap.

Call through the connector	v2 828a680	This branch
skills + returned revision (^[a-f0-9]{64}$)	accepted	accepted
skills + returned deployment	rejected, ^dpl_	accepted
skills + returned profile	rejected, ^ins_	accepted
skills document + deployment + profile	not run	accepted
resume + returned apr_ ID, decline	rejected, ^apr_	completed, ApprovalDenied
skills + deployment: "ins_0000"	not run	rejected, ^dpl_[\s\S]+$

A direct MCP client accepted the returned IDs on both images. The rejections came from the connector, not from Executor. elc_ uses the same pattern shape but wasn't exercised through the connector, because no scenario produces an input-required pause.

all

all because the shared Id helper validates every SDK ID across the HTTP API and MCP.

Checklist

• Added a changeset (bun run changeset), or this change needs none.
• Added or updated tests for the new behaviour.
• No secrets, credentials, or private data in the diff.

[jameswyse]

Closing this, since the change landed on v2 in e780e12 ("Export 72d2449") alongside the closure of #2186 via the export sync.

Thanks @RhysSullivan for picking it up so quickly!