# Issue 19821 — evidence index

Server/provider guidance change: independent review rounds use fresh delegation; follow-ups requiring the child's conversation use thread-send. No visible differences. Visual fixture matrix EMPTY; required/captured visual rows **0/0**. No render fixture, browser, still, recording, or server launch was needed or performed in this evidence step.

## Exact focused regression proof

Command for both runs:

```sh
vp test run apps/server/src/mcp/toolkits/orchestrator/tools.test.ts apps/server/src/provider/CodexDeveloperInstructions.test.ts
```

- [Red test log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ceaf0ca117354e85/red-test.log): detached throwaway worktree `/tmp/issue19821-red-evidence`, revision [23673c96f8](https://git.cafe/pingdotgg/t3code/commit/23673c96f8937d7c18a5c323daa78c4537d9c697). **4 text assertion failures in 1 test; 17 passed / 18 total; exit 1.** Both injected guidance and delegate-task description fail the independent-review and same-conversation follow-up assertions. These are actual vp test failures, not assertion replay.
- [Green test log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5f64bc5da67f2dd0/green-test.log): detached throwaway worktree `/tmp/issue19821-green-evidence`, requested head [e7417481e5](https://git.cafe/pingdotgg/t3code/commit/e7417481e5a2a2f130551092f05e1aced2072b08). **2 files passed; 18 tests passed; exit 0.** Includes Codex instruction byte-cap and ordering coverage.

Explicit base: [8acbefefc5](https://git.cafe/pingdotgg/t3code/commit/8acbefefc51b4d281dcd04108c82baa1ab902b0e). Red revision differs from base only by the regression tests. The red log records an empty diff against base for both implementation sources. Source was never swapped in the shared checkout.

Dependencies were reused from the installed checkout. Each throwaway worktree has local node_modules directories with external dependencies linked to the existing installation and `@t3tools` workspace links redirected to that throwaway worktree's own packages. No install was necessary. An initial red run also produced the expected four failures; the published red log is the final repeat launched after dependency redirection was complete. Test runner: vp 1.3.0 / Vitest 5.0.5, host Node v24.21.0 (vp reports configured engines.node v26.12.1).

## Runtime evidence

- [Before: Codex and Claude](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b532b8d0f3933a0a/19821-runtime-before.md). Actual live-service behavior under broad old guidance. **Running build was not mapped to the explicit baseline**; this is observational context, not a controlled source-pinned A/B.
- [After: Codex](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7a2d25b7de4eb00f/19821-runtime-after.md). Requested head, Codex gpt-6.1-sol/options []; delegate then two thread-send follow-ups; same child conversation and distinct tasks; marker recovered without being included in round 2. **Watcher restart interrupted the first attempt; automatic recovery completed the probe.** Rendered transcript contains reported tool calls rather than a raw MCP event trace.
- [Sanitized Codex child projection](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9d4ab79101f8c990/19821-child-projection.json). Supplied persisted-message projection with synthetic marker prompts and replies, three completed runs, and one childThreadId. Complements the rendered transcript; it is not a raw tool-event trace.
- [After: Claude](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/049d941e688045fd/19821-runtime-claude-after.md). Lead-supplied completed head probe, Claude Opus 5.5/options fastMode=false. Delegate then two thread-send follow-ups in the same child; omitted marker recovered and round 2 recalled. Includes read-only persisted child-message projection. All child runs completed before the later merge; additional empty parent wakes after merge were cancelled. This evidence step retrieved the artifact but did not independently rerun either provider.

## Postmerge and scope

Lead reports main merged cleanly as [091f756fef](https://git.cafe/pingdotgg/t3code/commit/091f756fef57e3a09a8e08807cc1213390fcdfa6) atop [a50cd989ae](https://git.cafe/pingdotgg/t3code/commit/a50cd989ae), with the exact four-file fix unchanged relative to the new base. [Lead postmerge focused test log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/fa8cb885e6e2ec67/19821-post-merge-tests.log): 2 files / 18 tests green. This is additional lead-supplied evidence; the independently captured green log above remains pinned to the requested premerge head.

[Proof audit and empty fixture matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6de44814cc81e116/issue-19821-proof-audit-pass1.md). No client render paths change. Other provider consumers are source-covered, not runtime-proven here. No connection-mode, UI, or full-suite claim is made.

Shared checkout was already at the merged head when this evidence step began; it remained clean and was not modified. No fixtures/scripts were added to the PR. Logs and helper material remain under `/tmp`; only temporary detached worktrees were created. No forge commands, comments, PR changes, or other forge writes were performed. Lead reports the existing dev server was stopped cleanly; this step did not manage it.

## FRICTION

- **brief:** No ambiguity after preserving the empty visual matrix and explicit source pins.
- **runbook:** Text-only regression and persisted synthetic conversation evidence fit this change; visual rows remain 0/0.
- **infra:** Shared checkout had advanced to merged main. Both test runs used isolated detached worktrees. Dependency reuse required redirecting workspace links before the final red launch. Existing after-Codex runtime recovered from a watcher restart.
- **policy:** No browser or computer use, server launch, active-checkout mutation, or forge write was needed.
- **tool:** Actual vp tests available; no lower-fidelity assertion replay used. Public artifacts uploaded with `ghdrop upload -m`.
- **forge:** Read-only local git inspection only; no forge interaction. Merge/postmerge state supplied by lead.
- **model:** Codex and Claude after probes now both supplied and passed; other providers not runtime exercised.
- **other:** Before live build lacks baseline mapping. Runtime transcripts/projections retain their stated provenance; no controlled live A/B claim.