ClawCures

Disease-campaign orchestration that separates planning from execution:

• OpenClaw plans through /v1/responses.
• refua-mcp executes typed scientific tools, with built-in web_search/web_fetch for evidence collection.

What You Get

• Single-run planner/executor flow.
• Autonomous planner/critic loop with policy checks.
• Offline plan validation and dry-run modes.
• Portfolio ranking for disease-program prioritization.
• Structured promising_cures output with full ADMET property maps and assessment text.

Requirements

• Python >=3.11,<3.14 (Python 3.14+ is not supported by refua-mcp).
• A running OpenClaw Gateway if you want live planning.
• Optional: refua-mcp runtime dependencies for tool execution.

Quick Start

1. Install

cd /path/to/ClawCures
python3.13 -m venv .venv
source .venv/bin/activate
pip install -e .

2. Configure OpenClaw access

• Set your gateway URL with REFUA_CAMPAIGN_OPENCLAW_BASE_URL.
• Ensure OpenClaw responses endpoint is enabled.
• Provide auth with one of:
  • REFUA_CAMPAIGN_OPENCLAW_TOKEN
  • OPENCLAW_GATEWAY_TOKEN
  • OPENCLAW_GATEWAY_PASSWORD (for password mode)

3. Run an offline smoke test (no OpenClaw call)

ClawCures run \
  --objective "Offline validation" \
  --plan-file examples/plan_template.json \
  --dry-run

4. Simplest live invocation (all-disease cure mission)

ClawCures run

ClawCures run now loops continuously by default (--max-cycles 0) until you stop it. To run a single cycle and exit, use --max-cycles 1. In loop mode, memory is preserved automatically by reusing a stable session key, enabling OpenClaw response storage, and carrying concise cycle summaries forward into subsequent planning objectives.

By default, ClawCures run sets the objective to: "Find cures for all diseases by prioritizing the highest-burden conditions and researching the best drug design strategies for each."

If planner JSON generation fails repeatedly (common on small local models), ClawCures now auto-attempts repair passes and then falls back to a deterministic all-disease bootstrap validation plan. For all-disease objectives, this fallback now includes deterministic web_search target-discovery queries before validation calls.

5. Run a live planning dry-run

ClawCures run \
  --dry-run

Environment Variables

Variable	Default	Purpose
REFUA_CAMPAIGN_OPENCLAW_BASE_URL	http://127.0.0.1:18789	OpenClaw Gateway base URL
REFUA_CAMPAIGN_OPENCLAW_MODEL	openclaw:main	OpenClaw model/agent ID
REFUA_CAMPAIGN_TIMEOUT_SECONDS	180	API timeout
REFUA_CAMPAIGN_OPENCLAW_TOKEN	unset	Bearer token override
OPENCLAW_GATEWAY_TOKEN	unset	Gateway token fallback
OPENCLAW_GATEWAY_PASSWORD	unset	Password-mode fallback token
REFUA_CAMPAIGN_SESSION_KEY	unset	Optional stable OpenClaw user key for cross-turn memory
REFUA_CAMPAIGN_STORE_RESPONSES	unset	Optional bool (true/false) for OpenClaw response storage
REFUA_CAMPAIGN_AGENT_MODEL_MAP_JSON	unset	Optional JSON object for phase/domain model routing (e.g. {"planner:oncology":"openclaw:oncology-planner","critic":"openclaw:critic"})
BRAVE_API_KEY	unset	Optional key for higher-quality web_search results
CLAWCURES_ALLOW_PRIVATE_WEB_FETCH	unset	Set true to allow web_fetch against localhost/private IPs

CLI Commands

Command	What it does
ClawCures print-default-prompt	Print bundled mission prompt
ClawCures list-tools	Show available execution tools (refua-mcp + web tools)
ClawCures run ...	Continuous planner + execution loop (--max-cycles 0 default)
ClawCures run-autonomous ...	Planner/critic multi-round loop
ClawCures validate-plan ...	Policy-check a local JSON plan
ClawCures rank-portfolio ...	Rank disease programs from JSON input

Common Usage

Run one plan + execute cycle:

ClawCures run \
  --max-cycles 1 \
  --output artifacts/kras_campaign_run.json

Run with OpenClaw native function calling (no intermediate JSON plan parsing):

ClawCures run \
  --max-cycles 1 \
  --native-tool-loop \
  --session-key mission-all-disease-v1 \
  --store-responses \
  --stream \
  --native-discovery-bootstrap-rounds 2 \
  --auto-web-fetch \
  --agent-model-map-json '{"planner:oncology":"openclaw:oncology-planner","critic":"openclaw:critic"}' \
  --evidence-file docs/RESEARCH.md \
  --output artifacts/native_tool_loop_run.json

The run JSON now includes:

• promising_cures: ranked therapeutic candidates extracted from tool outputs
• promising_cures_summary: aggregate counts and ADMET coverage
• interesting_targets: web-derived disease target hypotheses ranked by evidence density
• interesting_targets_summary: aggregate counts and top target list

Each cure includes:

• metrics (binding/admet/affinity/potency signals)
• admet.key_metrics (admet_score, safety_score, adme_score, rdkit_score when available)
• admet.properties (full ADMET scalar property map from discovered outputs)
• assessment (risk/opportunity summary)

Offline autonomous policy check:

ClawCures run-autonomous \
  --objective "Offline policy check" \
  --plan-file examples/plan_template.json \
  --dry-run

Rank programs:

ClawCures rank-portfolio \
  --input examples/portfolio_input.json

Release Packaging

Build and validate release artifacts in a version-specific directory:

./scripts/build_release_artifacts.sh

Upload only from that release directory (not from dist/*):

python -m twine upload dist/release-<version>/*

OpenClaw Compatibility

Verified against official OpenClaw docs and latest stable release v2026.2.15 (released 2026-02-16):

• API interface: POST /v1/responses
• Request fields used by this project: model, input, instructions
• Auth model: Authorization: Bearer ... token/password
• Endpoint behavior note: OpenClaw currently ignores metadata for prompt construction, so this project does not rely on metadata for critic payloads

Primary references:

• https://docs.openclaw.ai/openapi/openapi/responses_api/
• https://docs.openclaw.ai/openclaw/faq/
• https://github.com/openclaw/openclaw/releases/tag/v2026.2.15

Troubleshooting

• connection refused to /v1/responses
  • OpenClaw Gateway is not running or REFUA_CAMPAIGN_OPENCLAW_BASE_URL is wrong.
• 401 or auth errors
  • Token/password env var is missing or mismatched with gateway auth mode.
• requires a different Python: ... not in '<3.14,>=3.11'
  • Use Python 3.11, 3.12, or 3.13.

Notes

• Tool plans are strict JSON for reproducibility.
• All tool calls go through a strict allowlist.
• web_search and web_fetch are included in the allowlist for target/evidence discovery.
• web_fetch blocks localhost/private-network targets by default for safety.
• --native-tool-loop uses OpenClaw function calls directly, executing tools turn-by-turn.
• --session-key + --store-responses wire OpenClaw session memory across campaign turns.
• --stream requests OpenClaw streaming responses; --stream-to-stderr mirrors streamed deltas to stderr.
• --agent-model-map-file/--agent-model-map-json enable disease/phase agent routing.
• --evidence-file injects local literature text into OpenClaw input for evidence-grounded planning.
• --native-discovery-bootstrap-rounds forces early native-loop rounds to web discovery tools only (web_search, web_fetch).
• --auto-web-fetch follows discovered web_search URLs with bounded web_fetch calls for richer target evidence.
• --native-tool-fail-fast disables recoverable tool-error feedback and fails immediately on tool exceptions.
• Mission framing is aspirational; never claim cures without evidence.
• For local-model reliability, planner output is auto-repaired and canonicalized (args/tool aliases) before execution.
• ClawCures run defaults to an infinite loop. Use --max-cycles 1 for a single cycle.
• Loop mode carries forward compact cross-cycle memory (prior findings, failures, and state hints) to reduce repeated work.
• Architecture details: docs/ARCHITECTURE.md
• Research notes: docs/RESEARCH.md