| name | bug-capture |
| description | Capture a reported bug as a high-context GitHub issue after targeted diagnosis. Use when a user reports a bug, shares reproduction steps or screenshots, and wants the issue tracker to contain enough context for a later agent to fix it quickly. |
Bug Capture
Capture a real bug as a GitHub issue with enough evidence for a later agent to
fix it quickly: workflow, screenshot evidence, diagnosis, likely fix, and
verification path.
Rules
- Investigate, but do not fix: no product-code edits, branches, commits, or
leftover temporary files.
- Do not overstate certainty. Say when a bug was reproduced, when it was not
reproduced, and how confident you are in the diagnosis.
- Spend tokens on evidence that changes the issue. Summarize long logs, files,
and screenshots instead of pasting them.
Process
1. Gather
Use existing context first: reporter description, screen/route/feature,
workflow, actual result, expected result, environment/account state, screenshot,
logs, and issue links.
For screenshots:
- Inspect attached screenshots and summarize what matters.
- Use
[screenshot] in the issue unless the reporter provides a real URL or
safe path to reference.
- Keep image observations in the summary or diagnosis, not in a long screenshot
section.
2. Investigate
Inspect before asking questions.
- Read only relevant domain docs (
CONTEXT.md, feature docs, ADRs).
- Search for the reported screen, route, component, API, event, or error text.
- Check nearby tests and recent local patterns.
- Reproduce or run the narrowest practical UI/API/test flow when available.
- Use sub-agents only for independent questions; ask each for concise findings.
Record concrete evidence: whether you reproduced it, commands/run results,
browser observations, likely affected files/modules, and confidence.
3. Clarify
Ask questions only after investigation, only for facts you cannot infer.
- Ask one question at a time.
- Ask at most 3 questions total.
- Include your recommended/default answer.
- Focus on missing screen/step, reproducibility, expected behavior, or relevant
account/auth/Google/self-hosted state.
If key facts are still missing, publish a useful issue with an information-needed
label.
4. Prepare GitHub
- Use GitHub Issues through
gh, inferred from the current repo remote.
- Apply the
bug type to the issue (do NOT create a bug label)
5. Dedupe
Search open issues with screen, route, workflow, error, and symptom keywords.
- Strong duplicate: comment on it with the new workflow, screenshot placeholder,
diagnosis, likely fix, and verification path.
- Weak match: create a new issue and mention the possible relation.
- No match: create a new issue.
6. Publish
Use title format:
<area>: <observable broken behavior>
Use this body:
## Bug report
Short summary of what is broken and who is affected.
## Where it happened
Route, screen, feature, environment, or account state.
## Workflow
1. Reporter step.
2. Reporter step.
3. What happened.
## Expected outcome
What should have happened.
## Screenshot
[screenshot]
## Diagnosis
Reproduction status, strongest evidence, likely affected files/modules,
confidence, and related issue if any.
## Likely fix
Concrete handoff prompt for the fixing agent: where to start and what behavior
to preserve.
## Verification
Shortest practical flow, test, or command proving the fix.
If publishing fails, report the failure and provide the draft body.
7. Report
Return the issue URL, whether it was new or a duplicate comment, any missing
information, and a short plain-English diagnosis/likely-fix summary.