Ultradoc

ultradoc

Answer from retrieved source evidence and deliver a cited answer or reference manual. The bundled scripts/ultradoc.mjs is a zero-dependency Node engine. Resolve script paths relative to this skill directory, not the target checkout.

Choose the smallest sufficient route

Use node <skill-dir>/scripts/ultradoc.mjs --help for flags. A local-only request uses --sources code,docs; do not probe remote providers for it. Do not load the full workflow, overview, optional semantic setup or orchestration for a small question unless evidence exposes a gap.

Fast answer

For one fact settled by at most three literal code/docs claims:

1. Run ask --repo <path> --q "<identifier and question>" --sources code,docs --out <run-dir>. Read its retrieval notes, then the relevant excerpts in EVIDENCE.md. One named declaration may instead start with symbol, then persist the evidence needed by the answer/check workflow.
2. Read citation-format.md, write ANSWER.md in the run directory: answer first, one claim per sentence, each cited [E#]. Put the pinned commit from meta.json in an HTML comment. Put unresolved matters under ## Unknowns; never fill them from memory. Before describing an effect, inspect callback defaults and injected dependencies: distinguish computed/requested values from executed effects. A no-op implementation is evidence of no effect, not an unknown or a real timer/write/network call.
3. Run check --run <run-dir> --strict. Repair citation/coverage failures and rerun. If its coverage summary counts more than three claims, use the standard semantic verification below; a strict pass does not waive that gate. Present the answer, clickable source refs, commit and unknowns.

Stop when the excerpt settles the question and the gate passes. Do not execute unrelated project tests for a read-only explanation. Additional retrieval is for an identified evidence gap, not a required number of commands.

Any interpretation, disputed behavior or issue/PR claim requires the semantic verification below. A resolvable citation alone does not prove its claim.

Standard answer and semantic verification

1. Resolve the project, precise question and requested version. Ask only when target ambiguity prevents retrieval. Split multipart questions into explicit subquestions. Read retrieval-playbook.md when choosing sources/query variants; use ask to persist a pinned dossier.
2. Read retrieval notes before snippets: truncation, rate limits, regex-tier analysis and missing providers bound the answer. Decode unclear notes using reading-evidence.md and engine-evidence.md.
3. Drill only unresolved subquestions. Standard: one drill round, targeting two on-topic items per subquestion. Deep/contested: up to three rounds; stop when a round yields no new on-topic evidence. State remaining gaps explicitly.
4. Write the same answer contract as the fast path. Always run check --strict.
5. Run verify --run <run-dir>, judge every claim↔evidence pair skeptically as supported, partial, refuted or unsupported, with a concrete note. Cross-check issue/PR assertions against current code. Fold verdicts into one file, then run verify --apply <verdicts.json> --run <run-dir> and check --semantic --run <run-dir>. Missing verification fails closed. Fix, weaken, drop or re-retrieve unsupported claims and re-verify; never bypass the gate to obtain a pass.
6. Review answer-rubric.md, then present the answer in the user's language with source links, commit and coverage gaps.

Evidence rules

• Answer only from retrieved evidence. A familiar API or matching keyword is insufficient; the snippet must support the actual statement.
• Never call an empty, capped or failed retrieval proof of absence. Bound the conclusion to what was searched and disclose missing sources.
• check --strict applies to every answer and every generated DOC.md. Semantic verification is required except for the narrow fast path above.
• Keep going through recoverable errors. Rephrase an off-topic query; scope a truncated monorepo with --package; use available local sources offline and state the gap. Do not claim evidence from a failed provider.
• Retrieval and verification may run sequentially. Optional fan-out never changes the evidence or exit gates.

Open detail only when needed

• workflow.md: complete workflow, documentation, recovery table, source routing and optional semantic mode.
• tuning.md: caps, caching, cost and noisy retrieval.
• orchestration.md: independent worklists, subagent contracts and sequential eco mode; one writer folds results.
• provider-apis.md: issue/PR provider problems.
• web-discovery.md: web extraction and discovery.
• semantic-setup.md: optional local vector tiers.
• worked-example.md: complete worked answer.