repowiseCollapses search → read → reason into one round-trip. The tool
retrieves the most relevant wiki pages, enriches the top hits with
symbol docstrings and source excerpts, hands them to the LLM, and
returns a synthesized answer with citations and a confidence label.
If confidence is low, it returns best_guesses, a justified,
per-file candidate list with excerpts, instead of an empty answer,
letting the agent pick one to verify instead of skim the field.
It reads the generated documentation and its
search index; see MCP tools for where it fits among the
other tools.
When to call
- Always first on any code question, before
search_codebaseor manual file exploration. - If
confidenceis"medium"or"low", follow up withsearch_codebaseandget_contexton thefallback_targets. - Works best when the question names explicit identifiers (a class, function, or module name).
Parameters
| Parameter | Type | Description |
|---|---|---|
questionrequired | string | The developer's question, in natural language. |
scope | string | Optional path prefix to restrict retrieval (e.g. "src/auth/"). |
repo | string | Repository alias. Usually omitted; pass when the question is about a specific repo in a workspace. "all" is not supported. |
include | string[] | ["evidence"] returns the expanded evidence projection (up to 32,000 characters instead of 24,000). ["diagnostics"] adds retrieval_quality. |
Returns
| Field | Description |
|---|---|
answer | Synthesized answer with inline file citations |
citations | File paths the answer references |
confidence | "high", "medium", or "low", the one trust signal. It is graded from the retrieval and from whether the top hit covers what the question asks. high may be cited directly |
degraded | Present when no answer was synthesized: "no-llm-provider" or "synthesis-failed". The reply is ranked evidence, never high confidence. Retry synthesis-failed; with no provider the evidence is the answer |
grounding | How a grounded answer was produced: "symbol_body", "exact_symbol" or "extracted". Cite it directly |
retrieval_quality | Only with include=["diagnostics"]: "high", "partial", or "weak", the rating of the retrieval alone. It feeds confidence |
candidate_files | Ranked file paths the citations do not already name: up to 3 at high confidence, 5 otherwise. On low, read these before searching again |
fallback_targets | Top file paths from retrieval; agent should get_context on these for verification |
candidates | The ranked shortlist of files retrieval resolved, one {path, lines?} entry each, up to 5. Present whenever retrieval resolved anything, including high-confidence answers where retrieval is deliberately empty. It's where to look next, not evidence the answer is right |
retrieval | Top 5 wiki hits with title, target_path, score, summary, and (for top 2 file pages) symbols with docstrings and source excerpts |
symbol_bodies | When the answer names a function/method/class, the full live body of each question-named symbol, inlined so no get_symbol follow-up is needed |
best_guesses | On a low-confidence, gated response: a per-file candidate list, each with file, why_relevant, score, and (when available) an excerpt, so the agent can pick one to verify instead of skimming several |
episodes | At most one dated fact recorded about this checkout, present only below high confidence and when its scope intersects the answer |
note | Context string explaining why synthesis was skipped or hedged |
_meta | Timing, cache hit, answer hint |
Example
get_answer("how does the authentication flow work?")
get_answer(
"how are database migrations handled?",
scope="backend/",
)Things worth knowing
- Confidence is checked, not self-reported. A hedge in the model's own answer ("the evidence does not identify") lowers it. So do an answer that names none of the question's identifiers, a number the retrieved material does not contain, or a mechanism the cited code does not show.
- Question-aware symbol promotion: symbols matching identifiers
extracted from the question get longer docstrings (400 chars) and a
40-line source body excerpt, so "how does
Xwork?" questions can be answered without hedging. - Intersection retrieval: relational questions ("how does X talk to Y?") are split into two queries; hits appearing in both get a 2× boost.
- Caching: questions are normalized and hashed; repeat questions hit the answer cache instantly. Hedged cached answers are bypassed for re-synthesis when the symbol pipeline is updated.
- No LLM provider configured? The reply carries
degraded: "no-llm-provider": ranked hits and snippets atmediumconfidence, orlowwhen retrieval is weak. Local source, symbols and full-text search still answer. candidatesvsretrieval.retrievalis evidence: enriched hits to re-read when the prose needs checking, and it shrinks as confidence rises.candidatesis navigation: a flatter, longer shortlist of every file retrieval resolved, present even on high-confidence answers.
get_answer is the most useful tool in the set. Most "what does
this codebase do?" or "where is X handled?" questions are answered in
a single call.