Script a Claude Code headless run that a CI pipeline can actually trust

This prompt is invoked headlessly in print mode inside a CI pipeline, with no human present to answer a permission prompt or clarify an ambiguous instruction. Because of that, resolve ambiguity using the stated defaults below rather than asking a question that would never be answered, and never wait on approval before completing the task.

TASK
Run the full lint and typecheck suite against the current branch, and if there are auto-fixable lint violations, apply the fixes and commit them with message 'chore: auto-fix lint violations'.

DEFAULTS TO USE WHEN THE TASK IS AMBIGUOUS
If a lint rule violation is not auto-fixable, leave it unfixed and list it rather than guessing at a manual fix.

TOOLS AVAILABLE IN THIS RUN
Bash(npm run lint:*), Bash(npm run typecheck:*), Bash(git commit:*), Edit

OUTPUT CONTRACT — THE PIPELINE PARSES THIS PROGRAMMATICALLY
Produce exactly one JSON object as the final output, matching this shape, and nothing after it:
{ status: "success" | "blocked" | "needs_human", lint_violations_fixed: number, lint_violations_remaining: string[], typecheck_errors: string[], reason: string | null }
Do not wrap it in a markdown code fence, do not add commentary before or after it, and do not emit partial JSON if the task fails partway — emit a complete object with a status field describing the failure instead.

EXIT CONDITIONS
- If the task completes successfully against its stated definition, set status to success and populate every other field.
- If the task cannot be completed for a reason the pipeline should surface to a human later, such as a missing file or a test that fails for reasons outside this task's scope, set status to blocked and put the specific reason in the reason field — do not attempt an unrequested workaround just to force a success status.
- If a step in the task would normally require a permission prompt that cannot be answered in this unattended context, such as installing a new dependency, do not attempt it silently and do not fail silently either — set status to needs_human and name the exact action that needed approval.

CONSTRAINTS
- Do not produce output that reads well to a human but cannot be parsed by a script; the JSON object is the actual interface here, and a pipeline that fails to parse it fails the whole run regardless of whether the underlying work was correct.
- Do not retry a failed step more than 1 times before reporting blocked; an unattended run that loops indefinitely on a step that will not succeed wastes CI time with no path to resolution without a human anyway.
- Log a brief plain-text trace of what was attempted, in order, before the final JSON object, so a human reading CI logs later has a narrative to check against the machine-readable result.
- Only use the tools listed in Bash(npm run lint:*), Bash(npm run typecheck:*), Bash(git commit:*), Edit; if the task as described would require a tool outside that list to complete fully, report that gap through the needs_human status rather than attempting the closest available substitute silently.

Begin the task now.