# Evidence index — takeover guidance alignment

PR [#20053](https://git.cafe/pingdotgg/t3code/pulls/20053), mirrored [#20053](https://github.com/pingdotgg/t3code/pull/20053).

BEFORE [eea9895](https://github.com/pingdotgg/t3code/commit/eea989536849e5b5049987ee24318b5a77326542); AFTER [c3eda49](https://github.com/pingdotgg/t3code/commit/c3eda496803eaedb003c5f01e9c7905209325e13). Branch takeover/pr-20053, checkout /home/ubuntu/wt/pr-20053. Five files, 43 insertions / 10 deletions. Matrix refreshed from older replacement head; this evidence uses the exact new head.

Executed 2026-10-24 on machine fs-40, Linux x86_64, Node v26.12.1, Vite+ test runner v5.0.5, installed Effect 4.0.5. Agent: gpt-6.1-sol / Codex harness. No live T3 state used.

## Result

RED: exit 1, exactly two new routing assertions fail and 21 tests pass with BEFORE production text. GREEN: exit 0, all 23 tests pass on AFTER. These are full runs of the two affected test files, not repo-wide tests. BEFORE experiment changes only the shared instruction module and orchestrator tool metadata to their immutable BEFORE contents; regression tests and all other sources remain AFTER. This is not a full BEFORE checkout. Restoration occurs in finally and is verified byte-for-byte before returning. Final git status --porcelain=v1 is empty; HEAD remains AFTER.

## Coverage against all 15 matrix rows

| Rows | Evidence | Scope |
|---|---|---|
| S1 | rendered-before-after.txt, short-transcript.txt | actual exports, complete section/joined string, unchanged sections 2–5 asserted |
| S2 | rendered-before-after.txt | actual Codex builder at each SHA, four browser/device pairs, exact keys/values/UTF-8 counts <=4000; equal orchestration entries asserted |
| S3 | source-gates.txt | adapter hasT3Mcp gate; no live Codex invocation; both modes share builder |
| S4 | source-gates.txt + rendered-before-after.txt | inspected Claude definedness gate; shared suffix assembly projection; no SDK run |
| S5 | rendered-before-after.txt + source-gates.txt | all four first-run/MCP helper combinations and real caller |
| S6 | rendered-before-after.txt | exhaustive 40 tuples per revision: two modes × MCP boolean × five prior states × normal/slash prompt; deduplicated actual outputs |
| S7 | rendered-before-after.txt + source-gates.txt | slash bypass asserted for all states; compaction reinjection follows inspected adapter dropping prior state; no actual compaction |
| S8 | rendered-before-after.txt + source-gates.txt | enabled/disabled actual shared helper, both legacy/v2 callers |
| S9 | rendered-before-after.txt + source-gates.txt | actual generated Pi literal decoded, shared equality asserted; section assignment/fallback rendered; credentials/hook gate inspected, no extension execution |
| T1 | rendered-before-after.txt + source-gates.txt | exact decoded metadata rendered as tools/list description projection; registration preserves Tool.getDescription; not HTTP/stdio exchange |
| T2 | rendered-before-after.txt + short-transcript.txt | full unchanged t3_thread_send description, cross-SHA equality asserted; modes not executed |
| D1 | rendered-before-after.txt + short-transcript.txt | exact source Markdown paragraph at each SHA |
| X1 | red.log, green.log, tests-summary.log | two new regression cases red against old production text; all 23 green on replacement |
| N1 | source-gates.txt | N/A: no UI or interaction change; dispatcher explicitly required text-only evidence |
| N2 | source-gates.txt | N/A runtime dimensions: no wire, authorization, lifecycle or transport change; no live provider/model execution |

All AFTER positive helper outputs are checked for both independent review and same-child follow-up language, and absence of obsolete prohibition. Tools metadata is checked separately. Actual model compliance, independent review quality, conversation memory and lifecycle behavior are outside this proof. The projection omits tool schemas and other metadata and is explicitly labelled throughout.

## Reconstruction

Keep artifacts in /tmp/to-20053. From the stated checkout at AFTER:

```sh
node /tmp/to-20053/render.mjs > /tmp/to-20053/rendered-before-after.txt 2> /tmp/to-20053/render.stderr.log
node /tmp/to-20053/source-gates.mjs > /tmp/to-20053/source-gates.txt
node /tmp/to-20053/tests.mjs > /tmp/to-20053/tests-summary.log 2>&1
```

The last command temporarily replaces two production files; run with exclusive source access. It restores in finally. The harness itself records exact commands, provenance, stdout/stderr and exit statuses. No full repository copy is required. render.mjs and source-gates.mjs read immutable Git objects. Dependency registration source is inspected from the installed Effect dependency; lockfile equality across revisions is asserted.

## Evidence URLs

[render.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2375d80760bdadc1/render.mjs)
[source-gates.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/44daaddacc453872/source-gates.mjs)
[tests.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4602f7e1e33302b2/tests.mjs)
[rendered-before-after.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d084fe97f2f43c88/rendered-before-after.txt)
[short-transcript.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/db177fdacfe30147/short-transcript.txt)
[source-gates.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d45fcb3a6e91bf8f/source-gates.txt)
[red.log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/25ba3a3ffdc1a0fe/red.log)
[green.log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/37da3cb77ea99409/green.log)
[tests-summary.log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5542eceff804629f/tests-summary.log)
[render.stderr.log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bc0389bae2abc33b/render.stderr.log)
[proof-matrix.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/44676772cc144a1e/proof-matrix.md)

## Changes and limitations

Only scratch artifacts under /tmp/to-20053 were created. No commit, push, PR comment, watcher, delegated task, browser, dev server or provider run. The existing PR was registered with this thread as required by the harness; no forge content changed.

FRICTION: minor. Used authorized two-file restoration instead of copying the repository. Node emitted its standard experimental stripTypeScriptTypes warning (see render.stderr.log). Initial inspection corrected src/server and effect/src/ai source paths. No missing dependencies or test blockers.

## SHA-256 artifacts

- render.mjs: `2ea609faa78ebfb5e80d10fb1acd2d901c0a029471120ea6c3eea81491fa71b3`
- source-gates.mjs: `1447e18d810898866804049d6e8fc728b144c9d72be883b74727ad065ec80eae`
- tests.mjs: `de8544438b572e129b42921c57158a8c7102e25e273e3ed2760fb0129c2684e6`
- rendered-before-after.txt: `3cb4562c79106bc68d87662165cc3ab12f9c9dd1210c93d452201420f61632e4`
- short-transcript.txt: `4c610f3256af97cb593f5ac9ac3169c4d6bdd88c6fe9f22a8ff8d6f23a5bf255`
- source-gates.txt: `670c7a7b4b65db06f9297e391760440ce202bbb45738db615b8a92858a5fc914`
- red.log: `299fed60d71ac062561526274968dffd596678eea6c0fcb48a68dd63b481092c`
- green.log: `33253590f024bfe3f2af468a520e81d08ac8729e52451a6f1e879dc7deb3d854`
- tests-summary.log: `102d4c5092ee9e058e11bd971b5233155edeba3cf1ae1e8e5c14462caddf011f`
- render.stderr.log: `754fbc659ca0f7a78c1eaf7309d46457f6172cdba9b2fe9f4a76485ad99fd77f`
- proof-matrix.md: `944b0dd8339e656615d6acde268a0394aa311622ebd6af04e1fda9ce21c35e40`