deep-research

name

deep-research

description

Run end-to-end deep research using OpenAI Responses API deep-research models plus Gemini Deep Research plus Exa Deep Research, then synthesize results into a comprehensive, easy-to-read, citation-backed Markdown report by default. Use for due diligence, investment research, market/competitor analysis, technical investigations, policy analysis, and any high-stakes question where cross-checking sources across multiple engines improves confidence; generate alternate formats only when the user explicitly requests them.

Deep Research

Produce decision-grade research outputs with multi-engine evidence gathering:

• OpenAI deep research via bundled script (scripts/run_openai_deep_research.py)
• Gemini deep research via bundled script (scripts/run_gemini_deep_research.py)
• Exa Research API via bundled script (scripts/run_exa_research.py)

Default output is always a comprehensive Markdown report. If the user asks for additional formats, provide them after the Markdown report.

Non-Negotiable Output Contract

1. Deliver a comprehensive Markdown report first.
2. Make the Markdown easy to scan: clear headings, concise tables, explicit assumptions.
3. Attach citations/links for all load-bearing claims.
4. Clearly separate evidence from inference.
5. Only add alternate outputs (JSON, CSV-style table, Notion-ready, slide-outline, short brief) when explicitly requested.

Use references/output-contract.md for the exact structure.

Speed Policy (Parallel-First)

Optimize for shortest end-to-end time without reducing research quality.

1. Run OpenAI, Gemini, and Exa research in parallel whenever all are required.
2. Parallelize independent sub-questions with multiple agents (market, technical, legal, risks, benchmarks).
3. Start synthesis only after minimum evidence is available from all active engines, then continue ingesting late-arriving evidence in parallel.
4. Keep at least one high-level coordinator agent to merge claims, resolve conflicts, and enforce citation quality.
5. Use sequential execution only for truly dependent steps.

Workflow

1. Scope and clarify

• Define audience, decision to support, time horizon, and constraints.
• Confirm required output formats. If unspecified, use Markdown only.

2. Prepare research brief

• Create a prompt file in the current task artifacts folder.
• Include required deliverables, confidence handling, and citation requirements.

3. Launch research engines in parallel

• OpenAI track: run scripts/run_openai_deep_research.py (prefer submit_and_check).
• Gemini track: run scripts/run_gemini_deep_research.py (prefer submit_and_check).
• Exa track: run scripts/run_exa_research.py (prefer submit_and_check).
• Persist all track artifacts as they complete.

Default quality posture:

• OpenAI/Azure requests default to --reasoning-effort medium, which is the proven supported setting for the default o3-deep-research route in this environment.
• Gemini defaults to deep-research-max-preview-04-2026 for maximum comprehensiveness.
• Exa defaults to exa-research-pro for highest reasoning capability.
• Use high only when you have verified the selected model/deployment supports it; otherwise prefer the default medium.

4. Triangulate and resolve conflicts

• Cross-check overlapping claims between OpenAI, Gemini, and Exa outputs.
• Prefer higher-trust sources when claims conflict.
• Mark unresolved conflicts and confidence level.

5. Synthesize and deliver

• Produce the comprehensive Markdown report first.
• If requested, append alternate output(s) after the Markdown report.

Credential Loading

Prefer durable botfiles secrets over scraping project-local env files:

• Azure OpenAI deep-research: ~/pro/botfiles/secrets/local/codex-azure.rc
• Gemini and Exa direct providers: ~/pro/botfiles/secrets/local/deep-research.rc
• Direct OpenAI fallback: ~/pro/botfiles/secrets/local/codex-openai.rc, plus explicit direct opt-in

.botenv sources those files for normal agent shells. The bundled runners also load the relevant botfiles secret files directly when the environment was not preloaded, while still allowing an explicit --env-file <path> override and nearest .env fallback.

OpenAI Execution

Use the bundled script for deterministic runs and durable artifacts. Keep the final synthesized memo/report/brief in top-level task-progress-artifacts/, and send raw engine-specific outputs to task-progress-artifacts/scratchpad/openai/.

uv run ~/.codex/skills/deep-research/scripts/run_openai_deep_research.py \
  --action submit_and_check \
  --prompt-file <path/to/research-prompt.md> \
  --outdir <path/to/task-progress-artifacts/scratchpad/openai>

Azure routing:

• When AZURE_OPENAI_DEEP_RESEARCH_ENDPOINT or AZURE_OPENAI_DEEP_RESEARCH_BASE_URL is set, the runner uses Azure Responses.
• Set AZURE_OPENAI_DEEP_RESEARCH_API_KEY when the deep-research deployment lives on a different Azure OpenAI resource than AZURE_OPENAI_API_KEY; otherwise the script falls back to AZURE_OPENAI_API_KEY.
• In Azure mode, --models values must be Azure deployment names. If omitted, the runner uses AZURE_OPENAI_DEEP_RESEARCH_DEPLOYMENTS or falls back to o3-deep-research.
• Direct OpenAI billing is disabled by default. Use it only with explicit opt-in via --allow-direct-openai or OPENAI_DEEP_RESEARCH_ALLOW_DIRECT=1 plus OPENAI_API_KEY.

Useful variants:

# Submit only
uv run ~/.codex/skills/deep-research/scripts/run_openai_deep_research.py \
  --action submit \
  --prompt-file <path/to/research-prompt.md> \
  --outdir <path/to/task-progress-artifacts/scratchpad/openai>

# Check an existing response id
uv run ~/.codex/skills/deep-research/scripts/run_openai_deep_research.py \
  --action check \
  --response-id <resp_id> \
  --outdir <path/to/task-progress-artifacts/scratchpad/openai>

Model guidance:

• Azure fallback order comes from AZURE_OPENAI_DEEP_RESEARCH_DEPLOYMENTS or defaults to o3-deep-research.
• Direct OpenAI opt-in fallback order is o3-deep-research,o4-mini-deep-research.
• Override with --models when needed.
• Reasoning effort defaults to medium, the proven supported setting for the default o3-deep-research route. Override with --reasoning-effort high only after verifying the selected model/deployment supports it.
• Use --max-tool-calls <n> for low-cost smoke tests or latency-bounded checks.

Gemini Execution

Use the bundled script for deterministic runs and durable artifacts. For best first-pass reliability, use an explicit long timeout and retry budget. Keep raw Gemini outputs under task-progress-artifacts/scratchpad/gemini/; reserve top-level task artifacts for the synthesized outputs you want humans to review quickly. The runner defaults to the high-quality Deep Research Max agent when supported and sends the documented Deep Research agent_config (type=deep-research, thinking_summaries=auto, collaborative_planning=false) unless --no-agent-config is passed.

uv run ~/.codex/skills/deep-research/scripts/run_gemini_deep_research.py \
  --action submit_and_check \
  --prompt-file <path/to/research-prompt.md> \
  --outdir <path/to/task-progress-artifacts/scratchpad/gemini> \
  --timeout-minutes 180 \
  --max-timeout-retries 2

If you want unlimited waiting, use no-timeout mode:

uv run ~/.codex/skills/deep-research/scripts/run_gemini_deep_research.py \
  --action submit_and_check \
  --prompt-file <path/to/research-prompt.md> \
  --outdir <path/to/task-progress-artifacts/scratchpad/gemini> \
  --timeout-minutes 0

Useful variants:

# Submit only
uv run ~/.codex/skills/deep-research/scripts/run_gemini_deep_research.py \
  --action submit \
  --prompt-file <path/to/research-prompt.md> \
  --outdir <path/to/task-progress-artifacts/scratchpad/gemini>

# Check an existing interaction id
uv run ~/.codex/skills/deep-research/scripts/run_gemini_deep_research.py \
  --action check \
  --interaction-id <interaction_id> \
  --outdir <path/to/task-progress-artifacts/scratchpad/gemini>

Timeout behavior:

• --timeout-minutes <= 0 means no timeout.
• submit_and_check automatically retries timed-out jobs once by default (--max-timeout-retries 1).
• Increase retries for unstable runs (for example --max-timeout-retries 2).
• Do not use --foreground with deep-research agents; Gemini Interactions API requires background=true.

Recovery playbook:

1. If a run exits on timeout, re-check the latest interaction id first.
2. If still in_progress, keep polling with --action check.
3. Resubmit only when needed (or let submit_and_check retries handle it automatically).

Agent guidance:

• Default agent is deep-research-max-preview-04-2026 using Interactions API revision 2026-05-20.
• Use --agent deep-research-preview-04-2026 when speed/cost matters more than maximum comprehensiveness.
• Use --api-revision only when Google publishes a newer Interactions API revision or you need legacy compatibility.
• Script reads GEMINI_API_KEY from environment, explicit --env-file, nearest .env, or ~/pro/botfiles/secrets/local/deep-research.rc (falls back to GOOGLE_API_KEY).
• Submit actions save a header-free gemini-submit-request-*.json payload alongside the API response so agent/config drift can be audited without exposing API keys.
• Failed, cancelled, or expired interactions write gemini-terminal-summary-*.md instead of gemini-report-*.md. A gemini-report-*.md file means the interaction reached completed.
• Completed interactions write final model_output text only; thought summaries and stored prompts are excluded from gemini-report-*.md.
• If an older artifact has a report that only repeats the original prompt, treat it as a pre-fix runner artifact and inspect the matching gemini-check-*.json terminal status.

Exa Execution

Use the bundled Exa Research API runner for an additional independent pass.

uv run ~/.codex/skills/deep-research/scripts/run_exa_research.py \
  --action submit_and_check \
  --prompt-file <path/to/research-prompt.md> \
  --outdir <path/to/task-progress-artifacts/scratchpad/exa>

Model guidance:

• exa-research-pro is the default for highest quality and strongest reasoning.
• exa-research is acceptable when speed/cost matters more than maximum depth.
• exa-research-fast is acceptable for low-cost smoke tests.

The script reads EXA_API_KEY from environment, explicit --env-file, nearest .env, or ~/pro/botfiles/secrets/local/deep-research.rc. If EXA_API_KEY is unavailable but Exa MCP tools are available in the current agent session, the deprecated MCP deep researcher tools may be used as a temporary degraded fallback; record that in task artifacts.

3. Persist the returned report in task artifacts.
   • Default raw Exa outputs to task-progress-artifacts/scratchpad/exa/.
   • Promote only the final synthesized report/memo/brief to top-level task-progress-artifacts/.

Source Quality and Citation Rules

Apply references/source-quality.md.

Minimum quality bar:

• Prefer primary and institutional sources.
• Treat forums, SEO farms, and uncited aggregators as weak evidence.
• Label confidence for every major conclusion.
• Add a "What would change this recommendation" section for decision-oriented work.

Optional Output Modes

If requested, provide one or more of these after the Markdown report:

• Executive short brief (1-2 pages)
• Structured JSON summary
• Comparison table only
• Notion-ready page body
• Action checklist

Keep citation IDs consistent across formats.

Reference Map

• references/workflow.md - end-to-end operating procedure
• references/output-contract.md - Markdown-first report template and optional formats
• references/source-quality.md - source trust tiers, conflict resolution, confidence rules
• scripts/run_openai_deep_research.py - reusable OpenAI deep research runner
• scripts/run_gemini_deep_research.py - reusable Gemini deep research runner
• scripts/run_exa_research.py - reusable Exa Research API runner