fix(tracing): recognize Windows skill paths

[andrewjsu]

Summary

Recognize Windows path separators when inferring skill access from tool calls. The current detector matches /SKILL.md and skills/.../scripts/ only, so a successful Windows read such as Get-Content C:\Users\dev\.codex\skills\pdf\SKILL.md is missed. The same command with forward slashes is recognized.

This also affects Windows paths escaped inside exec JavaScript source, where two literal backslashes can appear in the captured tool input.

Change

• Accept one or more forward slashes or backslashes at the existing document/script separator positions.
• Add eight table-driven regressions: native and escaped Windows document paths, argument arrays, mixed separators, and normal/escaped/system/mixed skill-script paths.
• Keep tool allowlists, skill-name rules, prompt detection and existing path-inference semantics unchanged.

Verification

On Windows, Node 24.14.0 with the repository-pinned pnpm 9.5.0:

• Before the source change: the eight new regressions failed with [] instead of ["pdf"]; the five existing skill tests passed.
• After the source change: all 13 skill tests passed.
• pnpm run lint passed (formatting and TypeScript).
• Build passed. Full suite: 125 passed / 4 failed.
• Unmodified upstream 586910c81eeaf21d8f54c1c86c89a29a5466c1ef on the same machine: 117 passed / the same 4 failed. These are three hook-command.test.ts cases and the sessions-tree indexing case in trace.test.ts; no skips or assertions were weakened.

Local setup notes: used a process-local pnpm 9.5.0 launcher because nested scripts otherwise resolved the machine's pnpm 11; restored canonical Git line endings in this isolated checkout for Prettier. No lockfile or unrelated file changes are included.

Scope

This repairs separator recognition, not proof of successful/effective skill use. The detector still infers from input text and can tag failed reads or incidental path mentions; changing that meaning should be a separate change. Fixtures are synthetic and contain no private transcripts, credentials or account identifiers. The installed plugin and tracing configuration were not modified.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}