source-deep-dive

Codex compatibility note:

• Invoke $skill-name: load the registered .agents/skills/<name>/SKILL.md and execute it with Codex tools; no separate Skill tool or Claude mode switch is needed. Author .claude/**; mirrors are generated.
• Track steps before execution and keep progress/evidence current. Explicit skill requests execute their protocol. Use the native question tool for required questions.
• Map foreign tool names to available host capabilities. A name mismatch needs no approval; an actually unavailable required capability stops with its step/evidence and a user question before adaptation.
• Subagent authorization: when a skill is user-invoked or AI-detected and its protocol requires subagents, that skill activation authorizes use of the required spawn_agent subagent(s) for that task.
• Preserve protocol order unless the user approves a deviation. Workflows use $start-workflow: gates stay fixed; other steps may flex with logged reasons and evidence.

[BLOCKING MUST ATTENTION] Execute declared steps in order; NEVER skip, reorder, or merge without explicit user approval. [BLOCKING MUST ATTENTION] Update task tracking before/after each step or sub-skill: in_progress at start, completed at end. [BLOCKING MUST ATTENTION] Completed steps need brief evidence; skipped steps need explicit reason; if Task tools are unavailable, maintain an equivalent synchronized step tracker.

Quick Summary

Goal: Deep-dive the prior source map into a cross-validated, source-cited evidence base (tmp/research/_evidence-{slug}.md) where every finding has source trace and confidence, discrepancies stay explicit, and no unverified single-source claim becomes fact.

Summary:

• Purpose/input: Consume prior tmp/research/_sources-{slug}.md; prioritize Tier 1-2, high-relevance, gap-covering sources; NEVER start a fresh search.
• Ordered path: (1) load/prioritize source map → (2) fetch 5-8 sources with WebFetch (hard cap 8) → (3) extract claims, data, quotes, methodology, publication date, author credentials, source type → (4) cross-validate → (5) write evidence base.
• Confidence gate: 2+ independent sources agree = high confidence; disagreement = both positions + discrepancy; one source = single source, unverified; declare 95/80/60/<60% for every finding.
• Handoff/routes: Write tmp/research/_evidence-{slug}.md with inline citations, ## Unresolved Discrepancies, and ## Gaps Remaining; a direct run uses ask user question tool for the post-completion choice only (no routing question: a direct call is an explicit skill request).

Workflow:

1. Read source map — Load tmp/research/_sources-{slug}.md; prioritize Tier 1-2, high relevance, and gap coverage.
2. Fetch top sources — Run WebFetch for prioritized URLs; maximum 8 calls.
3. Extract findings — Capture claims, data points, quotes, methodology, publication date, author credentials, and source type.
4. Cross-validate — Compare findings; distinguish agreement, discrepancy, and unique-source claims; assign confidence.
5. Build evidence base — Write structured findings, citations, confidence, discrepancies, and gaps to the output path.

Key Rules:

• MUST ATTENTION Maximum 8 WebFetch calls per invocation; prioritize Tier 1-2 sources covering gaps.
• MUST ATTENTION Every finding cites a specific source and confidence percentage; factual claims need 2+ independent sources.
• MUST ATTENTION Conflicting claims → present both + flag discrepancy; one source → single source, unverified.
• MUST ATTENTION Output the intermediate evidence base, not final synthesis; preserve discrepancy and gap sections.

Critical thinking: Be skeptical; apply critical/sequential thinking; trace every claim and state confidence (>80% to act).