Define a specialized Claude Code subagent instead of overloading the main thread

Define a Claude Code subagent as a single markdown file for .claude/agents/, to be invoked either automatically when its description matches the moment, or explicitly by name.

JOB THIS SUBAGENT DOES, AND ONLY THIS JOB
Run the full test suite after any code change, parse failures, and report which specific test names failed and why, without attempting to fix anything.

WHEN IT SHOULD BE INVOKED
Automatically whenever the main conversation has just finished editing code and is about to claim a task is complete; also invokable explicitly by name.

WHY IT NEEDS ITS OWN CONTEXT RATHER THAN RUNNING IN THE MAIN THREAD
Test output for a large suite can run to thousands of lines; parsing that in the main conversation would consume context the actual coding task still needs, so it should happen in a disposable subagent context instead.

STRUCTURE THE FILE AS
1. YAML frontmatter: name (short, kebab-case, matching the filename), description (written as a clear trigger condition the main agent could match against, not a vague summary — state the situation that should cause delegation to this subagent), tools (an explicit, narrow list; grant only what this specific job needs, never the full default set out of convenience), and model, set to a faster, cheaper model than the main conversation, since parsing test output is high-volume and low reasoning difficulty if the job is simple and high-volume enough to justify a cheaper or faster model than the main conversation is using.
2. A system-prompt body written in second person, addressed to the subagent itself, stating its one job, its boundaries, and explicitly what it must hand back to the main conversation rather than silently deciding on its own.
3. An output contract: the exact shape the subagent must return to the calling conversation, since the subagent's own reasoning and any tool output it generated does not automatically appear in the main thread — only what it explicitly returns does.

CONSTRAINTS
- Do not give this subagent authority the described job does not need. A subagent that only reads code and reports findings should not have Write or Bash access, even if that access would make some future job easier — scope it to today's job, not a hypothetical future one.
- Make the description field specific enough that the main agent would not mistakenly delegate an unrelated task to this subagent, and would not fail to delegate a task that actually belongs to it. Ambiguous descriptions cause both failure modes at once.
- State explicitly, inside the subagent's own body, that it should not consume its own context speculatively exploring the repo beyond what the job requires — an isolated subagent with unbounded scope defeats the reason for scoping it narrowly in the first place.
- If this subagent's job overlaps at all with an existing subagent already defined in .claude/agents/, name the overlap and propose either merging the two responsibilities or drawing a sharper boundary between them, rather than leaving two subagents with descriptions vague enough to both plausibly match the same situation.

OUTPUT
The complete file content, frontmatter and body, ready to save as-is. After it, one sentence on what the main conversation must still do that this subagent explicitly will not — the actual boundary of delegation.