/cursor-workflows

Use this skill when the request asks for a cursor-workflow, autocursor gather, or cursor gather: read-heavy reconnaissance that can run through headless cursor-agent instead of spending Claude orchestration tokens.

The implementation lives in lib/autocursor.py and is intentionally portable: no golems imports, no worktree mutation, and no v0 edit flows. Worktree isolation is deferred until mutating workflows exist.

Primitives

• agent(prompt, *, schema=None, label=None, timeout=900, resume=None, model=None) runs one headless cursor-agent -p --force --approve-mcps --output-format json <prompt>. model=None means Cursor auto model; never pass --model unless explicitly set. With schema, AutoCursor appends a JSON instruction, validates harness-side, retries malformed output, and records raw NDJSON logs to disk.
• parallel(thunks, *, concurrency=8) runs a ThreadPoolExecutor barrier. Failed thunks return None. MAX_CHILDREN caps local concurrency.
• pipeline(items, *stages) flows each item through stages independently. A failed item becomes None.
• phase(title) prints an observability marker.
• loop_until_dry(round_fn, *, dry_rounds=2, max_rounds=10) keeps gathering until consecutive rounds add no new stable-keyed findings.

Inner-Loop Gather

For broad read-heavy recon, do not run one cursor-agent prompt and stop. Build an inner loop:

1. run parallel gatherers over the initial files or topics;
2. synthesize stable findings and derive follow-up questions;
3. run red-team / contradiction / missing-surface passes from those findings;
4. repeat with loop_until_dry() until consecutive rounds add no new stable-keyed findings or the max-round cap is reached.

Use this for cross-skill, workflow, policy, or architecture remediation where second-round drift is likely. Keep the workflow read-only; the lead applies patches and owns verification.

Analyze

analyze/analyze.py takes gathered findings, clusters/ranks/deduplicates them, then produces a structured synthesis through autocursor.agent(schema=...). This workflow is pure local analysis: no web, no search backend, and no external API.

Quick-Deep-Research

research/research.py is the subscription-only research workflow from SPEC §8: cursor-native web search (NO exa, NO paid API) builds a web graph of sources, entities, and claims; cross-reference/verify flags conflicts instead of silently merging them; synthesis emits a cited report. Treat it as a cheaper, thorough+fast alternative to Gemini Deep Research, not as SOTA deep research.

All reasoning and search in these workflows must stay behind the NO-API law: use the cursor-agent subscription path only, with no Anthropic/OpenAI/Gemini/Exa keys or paid external APIs. See docs/plan/cursor-workflows-harness/SPEC.md §8 for the full mandate.

Example

from lib.autocursor import agent, parallel
 
schema = {
    "type": "object",
    "required": ["file", "findings"],
    "properties": {
        "file": {"type": "string"},
        "findings": {"type": "array"},
    },
}
 
results = parallel([
    lambda path=path: agent(f"Inspect {path} for TODOs and risks", schema=schema, label=path)
    for path in ["README.md", "AGENTS.md"]
])

Run the fuller example with:

python3 skills/golem-powers/cursor-workflows/scripts/example-gather.py README.md AGENTS.md