# Orchestration guidance evidence

[Sanitized dev transcript and supplied installed-build baseline](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9393e2d08280d07c/issue-23191-dev-transcript.json)

## Provenance

- Fixed dev worktree: `/home/ubuntu/wt/issue-23191`, clean at [5f90bd3add](https://git.cafe/pingdotgg/t3code/commit/5f90bd3add7ede32d1fa75c2c298a324d4629d45), compared with [49e97d48f9](https://git.cafe/pingdotgg/t3code/commit/49e97d48f9742ba4c5ab79efc618fbf9066523ec).
- Original fix [ec336f235f](https://git.cafe/pingdotgg/t3code/commit/ec336f235f5df4081bbdc533d374695b0b94d687), original base [8acbefefc5](https://git.cafe/pingdotgg/t3code/commit/8acbefefc5). Only `docs/user/welcome-wizard.md` changed between the original fix and merged head; the guidance diff remains three files, four insertions, three deletions.
- Existing server started 2026-10-22 01:41:35.923 UTC, backend 15558, web 7518. Isolated home `/home/ubuntu/wt/issue-23191/.t3`. No server/browser launch or stop by evidence capture. No repository edits or PR actions.
- Provider: Codex; model `gpt-6.1-sol`; inherited options `[]`; runtime mode `full-access`. Outbound initialize identifies client `T3 Code`, version `0.1.1`. No provider binary version was captured.
- Parent launched through host CLI with `--base-dir /home/ubuntu/wt/issue-23191/.t3 --json --wait`; CLI completed successfully. It used `t3_thread_launch` internally. Both authorized read-only delegated children completed before this evidence handoff.

## Observations

1. Actual outbound decoded Codex `turn/start` payload contains the new async-first, retain-taskId, provider-client-serialization, and confirmed-launch-only rule in `t3_code_orchestration_01`. This is provider-bound protocol logging, not a model paraphrase or source string search.
2. Actual HTTP MCP `tools/list` from that dev server advertises the changed `delegate_task` description and mode schema description. Mode description is under `inputSchema.properties.mode.anyOf[0].description`. A short-lived local MCP client session was issued and revoked through the existing auth service using only isolated home state.
3. Both async launches confirmed `running`: 01:43:10.350 UTC and 01:43:10.929 UTC. Distinct taskIds and exact structured responses are retained in the transcript. First `task_status` began at 01:43:42.021 UTC, after both launch responses. Children executed `sleep 15`; final status responses confirmed `completed` and `PROBE_COMPLETE` at 01:43:42.235 UTC and 01:43:55.029 UTC.
4. All five authored orchestration sections arrived in sorted order. Their UTF-8 byte lengths were 2287, 1428, 2166, 1653, and 839. All actual additionalContext entries, including runtime/tools entries, were below 4000 bytes.
5. Primary-owned `final-tests.log` records the existing focused CodexDeveloperInstructions test: 1 file, 16 tests passed. Evidence capture did not rerun it.
6. Supplied installed-build baseline remains separately labeled in the transcript: the second blocking wait began 74 ms after the first returned; async launch confirmations took 795 ms total. This baseline is LIVE installed-build evidence, not fixed dev-build evidence.

## Limits and sanitization

The evidence establishes delivered guidance, advertised tool descriptions, async launch acceptance before status calls, and eventual completion of both children. It does not establish parallel CPU execution, an external-client behavioral regression test, or compliance with every instruction. No UI fixture applies: the visual matrix remains empty, and no screenshot was made. Baseline build/provider identity remains unestablished.

Transcript contains only allowlisted guidance, changed tool surfaces, probe calls/results, sleep commands, and safe provenance. Credentials, tokens, auth headers, unrelated conversation/database content, and the full capability catalog are omitted. Secret-pattern scan passed before upload. Raw logs and the auth script were not uploaded.

## FRICTION

- startup race | First host CLI attempt reached a watch-restart gap and returned “No running T3 Code server”; retry succeeded against the replacement process. | Confirm runtime file and listening port after the watch restart; use bounded retry, never start a competing server.
- CLI latency | Full CLI module imports were slow. | Reuse the running server and let one bounded CLI probe complete; avoid repeated speculative invocations.
- script service wiring | Outside-repo MCP capture initially lacked NetService. | Mirror the CLI runtime layers (NodeServices plus NetService), scope session acquisition/revocation, and retain no credentials.
- schema wrapper | Mode description is nested in optional/null `anyOf`. | Inspect the actual advertised schema wrapper, not only a top-level description field.