fix(skills): Replace invalid ${CLAUDE_SKILL_ROOT} with bare relative paths#111048
Conversation
jonathanhefner
force-pushed
the
remove-claude_skill_root-references
branch
2 times, most recently
from
a2e955b to
7c2ebc4
Compare
jonathanhefner
changed the title
${CLAUDE_SKILL_ROOT} with bare relative paths${CLAUDE_SKILL_ROOT} with bare relative paths
`${CLAUDE_SKILL_ROOT}` is not a recognized Claude Code variable, so all
22 file references using it were passed through as literal strings and
never resolved. While `${CLAUDE_SKILL_DIR}` does work, it is scoped to
bash injection commands (backtick-bang syntax) where the working
directory may differ from the skill directory. For plain markdown file
references like these, bare relative paths (e.g.
`references/endpoint-patterns.md`) are the standard approach per the
Agent Skills spec.
jonathanhefner
force-pushed
the
remove-claude_skill_root-references
branch
from
7c2ebc4 to
cc7a998
Compare
jonathanhefner
changed the title
${CLAUDE_SKILL_ROOT} with bare relative paths${CLAUDE_SKILL_ROOT} with bare relative paths
|
i would run this through skill-writer - the harnesses are supposed to shape some of this but if you dont give them certain hints they will just call random locations (and if this is wrong, skill-writer needs adjusted so its not mishandled in future skills) |
|
@dcramer I'm not sure what you're referring to. 😅 I found a post of yours on X that mentioned these skills, and I noticed a few minor issues, so I decided to submit PRs for them (this PR and #111040). If there's an upstream process responsible for the issues, though, I agree that should be fixed instead. (By the way, thank you for sharing your insights about skills! ❤️) |
|
@jonathanhefner i totally didnt realize this was an external contribution :) let me just grab this and run our tools on it we have https://github.com/getsentry/skills and we have this pr-writer skill we use in there |
Update the backend and frontend bug-review skills to match the current diff-first review workflow and review harness expectations. Broaden the trigger language so the skills match how PR and branch audits are actually invoked, and document the same conventions in AGENTS.md so agent guidance stays in one place. Co-Authored-By: Codex <noreply@openai.com>
dcramer
added
the
Trigger: getsentry tests
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
| description: 'Sentry backend bug pattern review based on real production errors. Use when reviewing Python/Django backend code for common bug patterns. Trigger keywords: "backend bug review", "common errors", "error patterns", "sentry bugs".' | ||
| allowed-tools: Read, Grep, Glob, Bash | ||
| description: 'Review Sentry Python and Django changes for bug patterns drawn from real production issues. Use when reviewing a backend diff or PR, checking Warden findings, auditing the current branch, reviewing production-error patterns, or looking for common regressions in `src/` and `tests/`.' | ||
| allowed-tools: Read Grep Glob Bash |
There was a problem hiding this comment.
allowed-tools changed to undocumented space-separated format
Medium Severity
The allowed-tools value was changed from comma-separated (Read, Grep, Glob, Bash) to space-separated (Read Grep Glob Bash). Claude Code's own documentation specifies comma-separated format for this field. The space-separated format may cause the parser to not recognize individual tools, potentially resulting in the skill losing pre-approved tool access and prompting for permission on every tool invocation. The other skill in this repo (react-component-documentation/SKILL.md) still uses comma-separated format, making this inconsistent as well.
Additional Locations (1)
2 participants
${CLAUDE_SKILL_ROOT}is not a recognized Claude Code variable, so all 22 file references using it were passed through as literal strings and never resolved. While${CLAUDE_SKILL_DIR}does work, it is scoped to bash injection commands (backtick-bang syntax) where the working directory may differ from the skill directory. For plain markdown file references like these, bare relative paths (e.g.references/endpoint-patterns.md) are the standard approach per the Agent Skills spec.