# Evidence index — Claude custom parent capabilities

PR [#22792](https://git.cafe/pingdotgg/t3code/pulls/22792), mirror [#22792](https://github.com/pingdotgg/t3code/pull/22792).

Base: [74541bab17](https://git.cafe/pingdotgg/t3code/commit/74541bab17270c35f4418d4e543d85628632ed8f) (`74541bab17270c35f4418d4e543d85628632ed8f`). Lead's real runtime after: [95d01097c4](https://git.cafe/pingdotgg/t3code/commit/95d01097c4bfe841543640f0e3864080eb22af15) (`95d01097c4bfe841543640f0e3864080eb22af15`). New local, unpushed test-only head: [b3f248adfd](https://git.cafe/pingdotgg/t3code/commit/b3f248adfd45d18371df6566b5e831f529312746) (`b3f248adfd45d18371df6566b5e831f529312746`).

Environment: Linux x86_64, Node v24.21.0; existing dependencies. This agent ran only scripted SDK-boundary tests. The supplied real isolated server/Claude Code 2.1.303 transcripts were read, not recreated. No browser/server started. No forge writes, pushes or PR edits; app-owned PR registration only.

## Artifacts

- [Core before runtime](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1d9ea38e27bb5670/before-runtime.json) — base SHA, saved descriptors, persisted explicit selection, query-open boundary and terminal states. Use the explicit-false fresh open, not the initial selection without options.
- [Core after runtime](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ddeec11fb2ef14d8/after-runtime.json) — unchanged runtime capture at [95d01097c4](https://git.cafe/pingdotgg/t3code/commit/95d01097c4bfe841543640f0e3864080eb22af15) (`95d01097c4bfe841543640f0e3864080eb22af15`); false/true/false successful turns.
- [Full scoped tests](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/88a56c8fdf1458cc/scoped-tests.log) — 223 passed, three files, exit 0.
- [Focused SDK fixture cases](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/52c25b7aeba6db7f/gap-tests.log) — 11 passed, 205 deliberately excluded by name filter; no upstream completion claim.
- [Checks](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/96261419fdd7d8fb/checks.log) — file lint/format and server package typecheck exit 0, whitespace check clean.
- [Revised matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a891857cb302ff18/matrix-revised.md) — all 10 original rows retained. [Original matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/832e74d821af5412/matrix.md).

## Reproduce

At the new test head in the existing worktree, with existing node_modules:

```sh
node_modules/.bin/vp test run apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts apps/server/src/claudeModelOptions.test.ts apps/server/src/provider/ClaudeModelCatalog.test.ts
node_modules/.bin/vp test run apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts -t 'custom Claude parent capability boundaries' --reporter verbose
node_modules/.bin/vp lint apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts
node_modules/.bin/vp fmt --check apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts
# From apps/server:
../../node_modules/.bin/vp run typecheck
```

## What else changed / runtime applicability

Exactly one file added assertions: apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts (+215 lines). Added 6 parent effort/catalog cases, one five-turn boolean sequence, and 4 send/steer prompt cases, with one module-level Effect decoder. No production edits. No scratch files, plans or evidence committed.

`git diff 95d01097c4bfe841543640f0e3864080eb22af15 b3f248adfd45d18371df6566b5e831f529312746 -- . ':!apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts'` is empty: the exact entire production tree is unchanged. Adapter blob at both heads is `08cb71ba6cf18b3b9aaff5442d4ea54975a7cd27`. Thus lead's core runtime artifact stays pinned to its actual capture SHA and applies to identical production code at the new test head; it was not rerun at the new SHA. Working tree clean after commit.

## Limitations and friction

No required row unverified; 10/10 rows captured by combined real runtime/scoped evidence. Unverified outside scope: Haiku upstream completion (known failure), new custom-specific upstream continuation/subagent sessions, other platforms and dynamic catalog refresh. No base rerun here; lead's two base regression failures remain separately reported evidence.

Ambient vp absent, used node_modules/.bin/vp. Initial fixture expectations were corrected for existing [1m] default suffix, automatic summary setting and omitted live effort field; production untouched. Inline fixture decoder initially triggered lint/Effect typecheck diagnostics, fixed with a module-level decodeEffect. Final checks pass (routine existing Effect suggestions omitted from sanitized check summary). Root disk 7.3G before/after, above 2GB floor. No installs, temporary worktrees or owned long-running processes; scoped temp directories cleaned by Effect scopes. Local evidence staging only in /tmp/pr22792-evidence. Uploaded logs contain fixture/test output and check summaries, no credentials or raw developer logs.

VERDICT: SUFFICIENT for evidence readiness under the expressly accepted server-only scope.