PR #26686 real dev verification

main: 5732bbee397f0a2b0e61d0c2e432f01dca6df3dc
PR: e9d21fab026b0910cdb97ce30e2c29abf127f4aa
Linux, Node v24.22.1, Google Chrome 151, Playwright from repo dependencies.

Both revisions installed with vp i and run one at a time with vp run dev.
main: /home/user/main/.t3, web 8693, server 16733.
PR: /home/user/t3code/.t3, web 5733, server 13773.
No source modifications, mocks, fixtures, replay provider, or synthetic payloads.
Real project /home/user/t3code added through app UI (main) and t3 project add (PR).

Real provider path:
t3 thread start -> running server -> acpRegistry local command -> official @agentclientprotocol/codex-acp@2.5.0 -> Codex gpt-6.1-sol via local proxy.
Official @modelcontextprotocol/server-filesystem read_text_file read the real file:
/home/user/t3code/.repos/effect-smol/packages/effect/src/http-api/internal/httpApiScalar.ts
The same prompt and file were used on both revisions. Codex's isolated home configured the real filesystem MCP server.
The adapter emitted a bounded approximately 1 MiB tool-result notification.

Observer: temporary Node module loader outside git wrapped the actual makeStrictNdJsonRpcParser decode calls. It summed input bytes, read count and performance.now decode time until messages decoded. Disk logging happened after timing. No messages or boundaries were synthesized or altered. Both servers used --max-old-space-size=8192 after main's unrelated stream stress crash.
These are one live observation per revision, not a statistical benchmark or end-to-end latency claim.

main parser observation: [{"pid":78403,"time":"2026-11-01T04:39:42.051Z","bytes":1048617,"reads":21,"decodeMs":38.37074300003587,"messages":1}]
PR parser observation: [{"pid":83500,"time":"2026-11-01T04:43:18.236Z","bytes":1048617,"reads":20,"decodeMs":4.5869069999607746,"messages":1}]

Actual CLI result on main:
{"threadId":"mcp:ce492e2d-0a7a-4b36-be53-1a03fa03f4c8","runId":"run:thread:mcp%3Ace492e2d-0a7a-4b36-be53-1a03fa03f4c8:ordinal:1","status":"completed","reply":"large source read complete ✓ café 東京.","error":null}

Actual CLI result on PR:
{"threadId":"mcp:ed0c7419-5baa-4e73-8358-b53332c210c2","runId":"run:thread:mcp%3Aed0c7419-5baa-4e73-8358-b53332c210c2:ordinal:1","status":"completed","reply":"large source read complete ✓ café 東京","error":null}


Neighboring behavior:
- Two follow-up messages submitted through the real web composer in the same ACP session. Real git status and branch commands completed, unicode responses streamed, submit control became available again. Recording captures the second continuation.
- Native Codex gpt-6.1-sol on both revisions executed git rev-parse --show-toplevel and returned the requested unicode text. Native Codex bypasses effect-acp, so this is adjacent coverage only.
PR native CLI result:
{"threadId":"mcp:7125337e-5711-4255-bd99-8e2ddbaf9199","runId":"run:thread:mcp%3A7125337e-5711-4255-bd99-8e2ddbaf9199:ordinal:1","status":"completed","reply":"native codex check ✓ café 東京","error":null}


Checks executed remotely through blacksmith-check:
vp i && vp test run packages/effect-acp/src
6 files, 64 tests passed, including strict malformed input and fragmented frames.
Existing CI: 31 successful, 0 failing, 0 pending as reported by cafe pr status.

Limitations and findings:
- No live 10 MiB ACP line: Codex/adapter capped the returned result around 1 MiB. The claimed 10 MiB allocation and latency numbers were not reproduced.
- Initial git log -n 2000 --format=fuller --stat on main streamed many small ACP messages instead of one large frame. The main server reached its default Node heap limit and disconnected. This separate stress failure was not investigated or fixed. Bounded file reads and the PR comparison used the same 8 GiB heap ceiling.
- Native desktop/mobile clients, macOS/Windows, Cursor, Grok and Antigravity not exercised.
- Malformed protocol input was checked by existing tests, not injected into the real provider.