# Advisor usage evidence

Issue [#24129](https://github.com/pingdotgg/t3code/issues/24129); existing PRs: GitHub [#24139](https://github.com/pingdotgg/t3code/pull/24139), GitCafe [#24140](https://git.cafe/pingdotgg/t3code/pulls/24140).

Primary proof: [captured runtime DOM transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/290dba0493f563e7/runtime-transcript.json), [exact reporter fixture](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4aefb97e053599e2/session.jsonl), and [focused red-test log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ef4f145a3eff6643/red-test.log).

## Provenance and commands

Worktree: `/home/ubuntu/wt/issue-24129`. Local commit IDs below are recorded as command output, not asserted to be published forge objects:

```text
main:             27e0c684c585f955252d65f06206dd11a143bff5
before/red only:  b9d069f4ad1f7ca65540714ec19bd12155455505
after/local:      f13a1fabed5fbd7fa1731ef57e3aab610cda26e5
existing PR head: 7d5880c5c666da50dc51559cbf5b1b18f45cf19b
```

The before commit adds only the regression test to main. The after commit applies the existing PR's production/parser/cache files unchanged. Verified read-only:

```sh
git diff --name-only 27e0c684c5 f13a1fabed
git diff --exit-code 7d5880c5c666da50dc51559cbf5b1b18f45cf19b f13a1fabed -- apps/server/src/provider/Drivers/claudeUsage.ts apps/server/src/usage/usageScanCache.ts
```

Production-file comparison: exit 0, no differences. All five changed files are server files (two production files and three test files); no client/render files changed.

Lead's previously executed runtime command, reproduced here for audit only:

```sh
T3CODE_DEV_SEED=0 CLAUDE_CONFIG_DIR=/tmp/issue-24129/claude vp run dev
```

This evidence step executed only the focused red test on the before checkout:

```sh
NO_COLOR=1 vp test run apps/server/src/provider/Drivers/claudeUsage.test.ts -t 'Claude advisor iterations' > /tmp/issue-24129/red-test.log 2>&1
```

Result: exit 1; one failed test, three unrelated tests excluded by the name filter. Exact failure: `AssertionError: expected [ { provider: 'claude', …(7) } ] to deeply equal [ { provider: 'claude', …(7) }, …(1) ]`. The diff shows the missing `claude-fable-5-1` record with 54,204 uncached input tokens and 4,229 output tokens. Failure occurs at `claudeUsage.test.ts:168`; the subsequent projected-parser assertion is not reached on red.

## Observable runtime result

The lead's authoritative captures used the same isolated fixture file, `/tmp/issue-24129/claude/projects/fixture/session.jsonl`, for both states. Conditions: real web app `/usage`, Claude Code filter, Cost, 30 days; Chromium, 1280×800 CSS pixels, DPR 2, light theme, no device emulation.

- Before: total $0.13; only `claude-opus-5-5` at $0.13; 126K processed tokens; no Fable row.
- After: total $0.88; `claude-fable-5-1` $0.75 plus unchanged Opus $0.13; 184K processed tokens; one session in both states.

The one-line fixture contains the same executor/advisor usage payload as the regression test, with synthetic fixture session/request/message identifiers. Both public artifacts were downloaded and verified byte-identical to the local files during this evidence step:

```text
session.jsonl SHA256:
947461dd08e8975b20f6c694fbfcd8a11c5033d52cfb10135d8bd299006d83b9
runtime-transcript.json SHA256:
f6a8213f1596fc09a7ab5a7fe8845b15e2aeaf237c646efdd22c828088b8c1f4
red-test.log SHA256:
96026a010b3f9151edb33b5cfb2dacd6c7014f4c0de9c90431d53e6693693598
```

Optional supporting captures remain locally at `/tmp/issue-24129/before-main.png` and `/tmp/issue-24129/after-existing-fix.png`. The runtime transcript and red failure are primary proof; screenshots are supplementary.

## Scope, limitations, and friction

Proof matrix: **0 rows**, `[]`, under the lead's explicit server-only exception. No changed client/render files require a visual matrix row. Independent proof audit pass 1 remains pending; this index does not claim audit approval.

The green agent reportedly passed 57 focused tests, server typecheck, and lint. Those checks were not rerun or independently certified in this evidence step. Runtime/browser evidence was captured previously by the lead; this step inspected the saved artifacts and did not launch a browser, server, or provider. No fresh provider call is needed to validate this offline transcript reader; this does not validate a new live provider emission or additional desktop/mobile/remote runs.

FRICTION: a default Python urllib download received HTTP 403; repeating with a browser User-Agent succeeded for both public artifacts. No remaining blocker. Uploads contain inspected test output and evidence metadata, with no credentials. Checkout and tracked files remained unchanged; no push, comment, merge, or public forge write was performed. Only evidence files were created under the lead-owned `/tmp/issue-24129`; no separate scratch files were created.