Deep Research

Deep Research (via DeepAPI)

We use DeepAPI (deepapi.co) for all deep research. This fully replaced the retired perplexity-deep-research skill that called OpenRouter/Perplexity directly.

API key

• Key lives in ~/.zshrc as DEEPAPI_API_KEY.
• Gotcha: do NOT source ~/.zshrc — it breaks the shell (exit 126). Use the env var if set, else read just the key line:

KEY=${DEEPAPI_API_KEY:-$(rg -o 'DEEPAPI_API_KEY=\S+' ~/.zshrc | head -1 | cut -d= -f2)}
BASE=${DEEPAPI_API_BASE_URL:-https://deepapi.co}

• Key missing → stop and ask the user. Never print or log the key.

Step 1 — Build the research prompt

Write ONE self-contained paragraph following the research-prompt skill:

• Lead with the single question + the decision/end use it informs.
• Embed all context — no back-and-forth needed.
• Number 3-6 inline sub-questions (1, 2, 3…). One mission per prompt.
• State include/avoid constraints; prefer primary sources; separate fact from inference.

Field limits: query ≤ 4000 chars (the paragraph goes here), optional context ≤ 8000, optional instructions ≤ 2000. Do NOT pass model or provider fields — the API rejects provider controls.

Step 2 — Run it

One call = one cited answer (~700 words max, finishes or fails within ~60s server-side).

IDK=$(uuidgen)   # keep this; retries must reuse the SAME Idempotency-Key
jq -n --rawfile p /tmp/dr_prompt.txt '{query:$p, maxCostUsd:"0.10"}' > /tmp/dr_body.json
curl -s --max-time 120 "$BASE/v1/research/deep" \
  -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $IDK" \
  -d @/tmp/dr_body.json > /tmp/dr_result.json

Default spend cap is maxCostUsd: "0.10" per call; raise it only if the user approves.

Step 3 — Read the report + sources

jq -r '.status'                    /tmp/dr_result.json   # succeeded | failed
jq -r '.output.answer'             /tmp/dr_result.json   # the report
jq -r '.output.sources[]?.url'     /tmp/dr_result.json   # source URLs

Save the report to a markdown file for the user and list citation URLs beneath it. Don't report research costs unless the user asks.

If output.sources comes back empty while the answer shows [n] citation markers, that's a DeepAPI regression (fixed 2026-07-05) — still deliver the report, but tell the user.

Bigger topics — multi-call reports

One call is capped at ~700 words. For a full deep-research report, fire one call per numbered sub-question (each with its own Idempotency-Key), then synthesize all answers + sources into a single markdown file.

Failure modes

• HTTP 402 insufficient_credits → stop; the user tops up at deepapi.co/credits; then retry with the SAME Idempotency-Key (safe — replays don't double-charge).
• HTTP 429 rate_limit_exceeded → wait Retry-After seconds, retry once.
• status: failed / HTTP 502 → report requestId + error.message to the user. Do not retry in a loop.
• Replayed request (same Idempotency-Key) returns HTTP 200 with replayed: true and no new charge.
• Envelope/auth mechanics and all other endpoints: see the deepapi skill.