deep-research

A subscription-only port of NVIDIA AI-Q's deep-research core (Apache-2.0 — see NOTICE). The main agent is the orchestrator: it plans, fans out parallel Task researchers that use native WebSearch/WebFetch, gap-fills, synthesizes a long-form cited report, and then runs a deterministic citation verifier so every [n] maps to a real, captured source URL.

$SKILL_DIR below is this skill's directory; prompts live in prompts/, the verifier in scripts/verify_citations.py. Read each prompts/*.md when you reach its step.

When to use

Researched reports, landscape/market/literature scans, cited comparisons, "deep research on …", "research X and tell me the alternatives". Skip it for trivial chit-chat or pure factual one-liners (the router will route those to a direct answer / shallow pass).

Flow

0. Route & disambiguate — read prompts/router.md. Classify meta vs research, then shallow vs deep. For a plainly trivial question, answer directly. If the query is genuinely ambiguous and you can't pause for the user, state ONE interpretation line at the top of the report and proceed.

1. Set up a per-run working dir — create a fresh run directory: <vault>/<topic>/runs/<run-id>/. Inside it create:

• <run-dir>/notes/
• <run-dir>/yt-report/
• <run-dir>/report_draft.md
• <run-dir>/report_final.md
• <run-dir>/audit.json

Prior runs are never verifier input. The lane must not write to persistent shared notes/; only the current run's <run-dir>/notes/ is scanned by the verifier.

2. Plan — read prompts/planner.md. Run 2-4 scoping WebSearch calls, then produce the plan object: task analysis, report title, TOC (≤8 sections), constraints (acceptance criteria), and 4-6 self-contained queries. Keep the plan in your context.

3. Research (parallel fan-out) — read prompts/researcher.md. Group the web queries into 2-3-per-bundle and launch up to 6 Task (general-purpose) subagents in a single message so they run concurrently. Give each subagent: the researcher prompt, its bundled questions, the relevant constraints, and an instruction to write its notes to `<run-dir>/notes/researcher_<k>.md` (a dedicated notes subdir — keep researcher notes separate from the draft; the verifier only scans researcher_*.md files, so this naming is REQUIRED) and also return them. Each researcher uses only WebSearch/WebFetch, ≤8 calls, broad→narrow, and lists ONLY real captured URLs.

3b. YouTube lane — if any planned queries are marked youtube_suited and gemini is available, run one Bash -> gemini task using prompts/researcher_youtube.md. Pass the tagged queries, <run-dir>, fan-out cap <=4, search/fetch cap <=8 per subagent, and links-per-video cap <=5. Gemini writes raw notes to <run-dir>/yt-report/researcher_yt_<k>.md and opened logs to <run-dir>/yt-report/opened_yt_<k>.json. If gemini is absent, skip the YouTube lane, log the skip, and continue web-only.

3c. Codex trust check — detect codex before checking YouTube outputs. If absent, continue to the deterministic gate with opened-log enforcement and emit the integrity warning Codex trust-check skipped. If present, run Codex read-only over <run-dir>/yt-report/ using prompts/codex_check.md; it may only write <run-dir>/yt-report/codex_annotations.json. Codex must not edit Gemini files, add sources, write notes, or synthesize.

3d. Filter YouTube notes — before synthesis and verification, run the deterministic filter for each raw YouTube note:

python3 "$SKILL_DIR/scripts/filter_yt_notes.py" \
    --run-dir <run-dir> \
    --index <k>

If Codex was detected as present for the run, add --codex-present. The filter writes <run-dir>/notes/researcher_yt_<k>.md and <run-dir>/yt-report/filter_integrity_yt_<k>.json. Opened-log exact-match enforcement always applies. Absent Codex emits Codex trust-check skipped; present-but-failed Codex annotations exclude all YouTube notes. drop URLs are removed mechanically at paragraph/list-item/sentence/marker scope. flag URLs are kept and surfaced in integrity metadata. No semantic claim rewriting is promised beyond those mechanics.

4. Gap-fill (bounded, merit) — read all researcher_*.md. Check each TOC section/constraint for coverage. If a section came back empty or weak, dispatch ONE more researcher to fill that specific gap, then proceed. Do not loop — "try once to fix, then proceed".

5. Synthesize — read prompts/synthesis.md. Read EVERY note file, then write a 3000-5000+ word report that follows the TOC, with inline [n] citations and a ## Sources section. Cite ONLY URLs present in the notes; never from memory; no bare URLs in the body. Write the draft to <run-dir>/report_draft.md.

6. Verify (deterministic — the core merit) — read prompts/citations.md. Run:

python3 "$SKILL_DIR/scripts/verify_citations.py" \
    --report <run-dir>/report_draft.md \
    --notes  <run-dir>/notes \
    --out    <run-dir>/report_final.md \
    --audit  <run-dir>/audit.json

The verifier rebuilds the source registry from the notes' real URLs and keeps a [n] only if its URL resolves into that registry (repairing fuzzy-but-real URLs, dropping fabricated ones, deduping, sanitizing, renumbering). URL matching is strict — exact / child-path / query-subset only; a report URL that is merely a prefix of a captured one is rejected (truncated citations are dropped, not guessed), and every URL in the final report must resolve into the registry. Then:

• exit 3 / `captured_sources: 0` → present under the `⚠️ UNVERIFIED — model knowledge

only` banner (or re-run); never as a normal report.

• exit 4 → sources were captured but the verified report has no surviving [n] (or no

## Sources, or a bare/unresolved URL remained): the script emits an ⚠️ UNVERIFIED banner and refuses to manufacture a citation. Do ONE targeted re-dispatch/rewrite citing only captured sources, then re-verify; never present exit-4 output as sourced.

• read audit.json and tell the user which citations (if any) were dropped and why

(citations_dropped, skipped_report_files, gate_failures).

• verify the report against each planner constraint; note any unmet.
• apply the done-gate (length, ≥2 ## headers, ## Sources present, ≥1 verified

citation, no giving-up phrasing). The script enforces the sources / ≥1-citation part (exit 4); the rest is on you. If it fails, ONE targeted re-dispatch, then re-verify.

7. Present report_final.md to the user, plus a one-line integrity note (sources captured, citations verified, any dropped). Read every <run-dir>/yt-report/filter_integrity_yt_<k>.json and aggregate its warnings, flagged_urls, and video_skips into that note so flag/low-trust kept URLs and YouTube degradation warnings are user-facing, including YouTube degradation warnings that would otherwise be hidden in metadata.

Dependency/degradation

Hard rules (the AI-Q merits — do not weaken)

1. Citation integrity is deterministic. Cite only URLs the researchers recorded from

WebSearch/WebFetch results; the script — not your judgment — is the gate (exit 3 = zero sources, exit 4 = zero surviving citations ⇒ the loud UNVERIFIED banner, never a normal report). The registry is built from the notes, so the script catches synthesis-time fabrication; for stronger provenance you can pass --registry sources.json.

1. Gap-fill once for weak sections before synthesizing.
2. Constraint verification against the planner's acceptance criteria at the end.
3. Disambiguate with a bias to proceed: state an interpretation rather than blocking.
4. YouTube trust boundary is weaker. The YouTube lane trust model is

Gemini-fetched -> opened-log filtered -> optional Codex-checked -> deterministically gated. This is weaker than Claude WebSearch/WebFetch capture because Gemini's opened log is self-reported. Opened-log enforcement lives in filter_yt_notes.py; Codex is read-only and the verifier still receives only <run-dir>/notes/.

Knobs

≤6 researchers · ≤8 searches each · 2 loops (round 1 + one gap-fill) · 4-6 plan queries · 2-3 queries per dispatch · ≤8 TOC sections · 3000-5000+ words (deep). See references/aiq-mapping.md.

Always produce a report

Never stop to ask permission mid-run, and never emit "I can't produce a report". A partial report with honestly-acknowledged gaps beats stopping — except the zero-sources case, which must be labelled UNVERIFIED.