| name | agentflow-search |
| description | Find code across many files from a natural-language description, returning ranked locations instead of loading every file into the main session. Use when the user asks where something lives or how something is implemented across a directory or codebase ("where do we validate tokens", "find the retry logic", "which files touch the pricing config") and you want the search to happen in an isolated context. Dispatches a Haiku worker that greps/reads candidate files and returns ranked {file, line range, relevance, snippet} hits. The native Explore agent is a fine alternative; this skill is the cheaper Haiku-pinned path with a structured JSON result. |
agentflow-search
Locate code matching a natural-language query across a directory tree and return
ranked hits — without pulling all the candidate files into the primary session.
A Haiku-pinned worker does the grepping/reading in an isolated context.
Honest note: the native Explore agent already isolates search in a subagent.
This skill's edge is the cheaper Haiku pin and a structured JSON result format.
When it fires
- "Where do we validate auth tokens?"
- "Find everything that touches the pricing config."
- "Which files implement the retry/backoff logic?"
How to run it
Dispatch the agentflow-haiku-worker subagent (Agent/Task tool, model haiku)
with the prompt below. Fill in {query}, {paths} (default the cwd), and
{max_results} (default 10). Return the worker's JSON verbatim.
You are a code search engine. Search the given paths for code matching the query.
Use Grep/Glob to find candidate files, then Read the most promising ones to
confirm relevance.
Return results as a JSON array of {file, line_start, line_end, relevance, snippet}
objects, ranked by relevance (0.0-1.0). Return ONLY the JSON array — no markdown
fences, no explanation. If no matches found, return [].
Return at most {max_results} results.
Query: {query}
Paths: {paths}
Other runtimes (Codex, and tools without Claude subagents)
The worker above is a Claude Code agent definition; elsewhere only the dispatch
changes — the methodology prompt is identical:
- Codex CLI:
spawn_agent with the filled prompt (requires
multi_agent = true in ~/.codex/config.toml), wait_agent for the result,
then close_agent.
- Other tools: use the runtime's isolated-subagent mechanism; if none exists,
run the prompt inline and honor its output rules exactly.
No model pin exists outside Claude Code — pick the cheapest/fastest model if the
runtime lets you choose. Since the AgentFlow worker agent is absent, prepend this
line to the prompt: "You are a disposable worker in an isolated context. Follow
the methodology exactly and return only the result — no preamble, no commentary."
