# Exact-file skill selection proof

Target: [#19702](https://github.com/pingdotgg/t3code/pull/19702). Original contributor: [floklein](https://github.com/floklein). Maintainer rebuild preserves the contributor commits and merges current main. Head SHA and review completion are recorded in the PR body.

## Before and after

The original focused set shrank from 678 cases in 13 files to 676 in 12. Including the existing mobile context suite subsequently needed to prove chip tapping, the final comparable union is 693 → 692 cases in 14 → 13 files. PR-specific cases are 14 → 14: the helper reductions funded one meaningful mobile parity regression, replacing helper-only assertions with actual manager start/steer, outbound Codex frames, rich-document editing/copy/rehydration, and mobile selection. Against exact main production sources, 10 behavior assertions failed and two compatibility cases passed; restoring the rebuild made all selected cases pass. The enabled agent-only collision correction separately failed its strengthened existing assertion before the predicate fix, then passed without adding cases.

## Live Linux proof

All data below is synthetic. The app used the designated worktree's isolated `.t3` home, never the developer's live database. Two different local `SKILL.md` files had the same `name: proof-dup`, with instructions to return `PROOF_ALPHA` and `PROOF_BETA` respectively. Both forbid commands and modifications.

- Actual Codex CLI 0.163.0 app-server discovery reported both files. Independent read-only ephemeral threads received structured skill inputs for alpha and beta. Each returned its own exact marker.
- Actual Claude Code 2.1.303, Opus 5.5, ran two read-only synthetic fixture reviews. The prompt was expanded by this head's real non-Codex binding helper; each invocation returned its own exact marker. Only Read was available, with no extra MCP servers.
- The running web app discovered both local files through Codex. The dollar menu offered both choices. Picking beta created a chip with beta's exact path; View instructions opened the beta file. Reload retained beta in the draft's actual Tiptap document.
- Submitting beta through the full T3 app, server and Codex adapter returned `PROOF_BETA`. A follow-up picked alpha and returned `PROOF_ALPHA`. Reloading the completed thread retained both results and two distinct exact-file links in persisted message history.
- The browser exposed end-truncated paths that hid the differing file directories. The final layout allows skill descriptions to wrap on web and mobile. Matching main/head screenshots accompany the PR.

The provider runs above exercise real CLIs and the integrated web run exercises the actual T3 pipeline. Test replay additionally proves actual outbound start and steer frames and repeated-reference deduplication. No real user workspace or live developer thread was used.

## Static checks and limits

The original 676-case focused set passes. The added mobile-context suite has 16 passing cases (15 existing plus one new regression), for 692 final cases in 13 files. Mobile chip tapping uses the same exact-file parser as dispatch; its existing details sheet opens the retained path and supports removal even without matching catalog metadata. The new assertion failed before that fix, then passed with the 13-case mobile-menu suite (29 total). Targeted lint/formatting, and shared, client-runtime, Cursor, Grok, web, mobile and server typechecks pass. Fresh review reports and any final reruns are identified in the PR body.

Linux browser, Codex and Claude were verified live. Cursor, Grok, Antigravity, OpenCode, Pi and ACP dispatch were traced through the common manager and tested at relevant boundaries; their live runtimes were not launched. Windows paths were tested as data; Windows, macOS, Electron, iOS and Android were not run. Native URI skills remain name-based because no local file exists. Human approval of provider behavior remains required. No merge was performed.

Tests: GPT-6 Astra; implementation: GPT-6.1 Sol; independent reviews: GPT-6 Sol and Claude Opus 5.5. Codex and Claude Code harnesses in T3 Code.