search-fastpath

Search Fastpath

Goal

Find the right files quickly without turning every task into an unbounded repository scan.

This skill is the retrieval lane. It should produce a compact evidence map that a coding or review lane can consume without repeating broad search.

Precision First

• Prefer exact paths, file names, symbols, routes, config keys, and copied error strings over regex.
• Use regex only when the user's description is vague, the exact symbol/path is unknown, or literal searches have failed and the pattern is genuinely structural.
• When using regex, record why literal lookup was insufficient and keep the search scoped to the smallest likely directory.
• Do not use regex for exact class names, method names, route strings, property keys, command names, filenames, or quoted log/error text; use fixed-string search instead.

Search Order

1. User-provided paths, current files, attachments, changed files.
2. Repo indexes: README*, .github/instructions/project.instructions.md, and target-local spec/INDEX.md / memories/repo/INDEX.md when present.
3. Exact file-name discovery: rg --files -g 'ExactName.ext', rg --files <likely-dir>, or find <dir> -name 'ExactName.ext'.
4. Literal content lookup with fixed strings: rg -F -n 'ExactSymbolOrError' <likely-dir>; add -w only when word-boundary matching is needed.
5. Scoped semantic search with 2-3 literal terms in the smallest likely directory when exact names are unknown.
6. Regex search only after the earlier steps fail or the task needs a structural pattern.
7. External sources only when repo evidence cannot answer the question and current instructions allow web use.

For code-generation tasks, stop search when you have: the target files, the local pattern to follow, the relevant test/check surface, and any public contract or frontend state affected by the change.

External Fetch Ladder

When a task needs web or repository evidence, use a source-aware ladder inspired by fetch-skill:

1. Prefer primary source URLs provided by the user.
2. Prefer raw Markdown / raw source over rendered HTML when available.
3. Extract the smallest relevant section, not the whole page.
4. Keep source_url, fetch_mode, and fallback_used in the result.
5. Treat social posts, mirrors, marketplaces, and blog summaries as secondary hints unless the user explicitly asked for them.
6. If output is summarized or truncated, keep a recovery URL or command.

Stop Rules

• Stop after enough evidence supports the next action.
• Stop after two searches with no new signal.
• Stop before repo-wide regex unless the task remains genuinely ambiguous after exact path, index, filename, and fixed-string attempts.
• Do not search directories excluded by PM or user.
• Do not inject long outputs; summarize and keep recovery commands.

Search Result To Packet Mapping

When this search output feeds context-packet-fastpath or a PM dispatch packet:

• selected_files -> files_involved or priority_files
• candidate_files -> search_hints
• commands, literal_attempts, regex_used, and retrieval_provenance -> packet retrieval_provenance
• gaps -> packet known_gaps, uncovered_items, or NEEDS_CONTEXT
• generated evidence map -> artifact_refs with kind: search_result
• budget overrun -> NEEDS_CONTEXT context_budget_exceeded with searched paths and the next smallest recommended shard

Record token-cost signals when practical: files_read, broad_searches, context_budget_used, artifact_reuse_count, and escalation_reason.

Output

## Search Result
- mode:
- commands:
- candidate_files:
- selected_files:
- literal_attempts:
- regex_used: yes|no, reason
- retrieval_provenance:
- artifact_refs:
- token_cost_metrics: files_read=, broad_searches=, context_budget_used=, artifact_reuse_count=, escalation_reason=
- gaps: