Past Research Search

past-research-search

Keep a local, searchable record of past tavily-research runs, so they can be found later by topic instead of by remembering a request_id.

Known issue: research status/poll currently require API-key auth, not OAuth

TODO: remove this section once Tavily's MCP server exposes a tavily_get_research tool — tracked as a gap report filed 2026-08-28.

As of 2026-08, re-fetching a run by request_id (the last step below) fails when the CLI is authenticated via tvly login (OAuth) — that path routes through Tavily's MCP server (mcp.tavily.com), which doesn't currently expose a tavily_get_research tool at all. research status returns Unknown tool: 'tavily_get_research'; research poll silently retries until it times out instead of surfacing that error. The local logging/lookup steps in this skill are unaffected — only the "re-fetch the full result" step needs this.

Fix: authenticate with a raw API key instead of OAuth before using the re-fetch step:

tvly login --api-key tvly-YOUR_KEY
# or: export TAVILY_API_KEY=tvly-YOUR_KEY

Store the log and saved reports in the user's working project or a configured writable data directory, not inside the installed power. If the log does not exist yet, report that no runs have been recorded locally.

Why this exists

tavily-research already supports checking a specific run by ID (tvly research status <request_id>, tvly research poll <request_id>) — that's "I have the ID, what's the status." This skill covers the gap: "I don't have the ID anymore, I just remember roughly what I asked." Tavily's API has no cross-session history endpoint to query, so this is a local logging convention, not a new Tavily capability.

How it works

1. After every tavily-research run, append a record to a local log: timestamp, the query, the model used, the request_id, and where the output was saved.
2. To look something up later, search that local log by keyword instead of trying to recall the ID.

Quick start

Log a run right after it completes:

python3 << 'PYEOF'
import json, os
from datetime import datetime, timezone

log_path = ".tavily/runs.jsonl"
os.makedirs(os.path.dirname(log_path), exist_ok=True)

record = {
    "timestamp": datetime.now(timezone.utc).isoformat(),
    "query": "Competitive landscape for AI code assistants",
    "model": "pro",
    "request_id": "req_abc123",     # from the research() response
    "output_file": "reports/ai_code_assistants.md",
}

with open(log_path, "a") as f:
    f.write(json.dumps(record) + "\n")

print("Logged.")
PYEOF

Look up a past run by keyword:

python3 << 'PYEOF'
import json

with open(".tavily/runs.jsonl") as f:
    records = [json.loads(line) for line in f if line.strip()]

keyword = "code assistants"
matches = [r for r in records if keyword.lower() in r["query"].lower()]

for m in matches:
    print(f"[{m['timestamp']}] [{m['request_id']}] {m['query']}")
    print(f"  model={m['model']}  output={m.get('output_file', 'n/a')}")
PYEOF

Re-fetch the full result for a match, if still within Tavily's retention window:

tvly research poll req_abc123 --json

Options

Tips

• Log at the moment of running, not after the fact. It's cheap to append one line right after a research call completes; it's much harder to reconstruct after the fact.
• Always pair with -o <file> on the original tvly research call — the log should point at a durable local copy of the report, not rely solely on Tavily's hosted retention.
• This can be a habit inside tavily-research rather than a separate step — if it's simpler for a given workflow, just always log after research calls without treating it as a distinct invocation.

See also

• tavily-research — the skill this one logs runs for