Skip to content

OpenClaw with the Verde API

The prototype now has a separate OpenClawBackend that invokes the actual OpenClaw llm-task plugin through /tools/invoke. It does not call Verde directly and label that an OpenClaw run. The same four bounded role contracts, scientific QA, model-review veto, and human-review requirements remain in place.

Runtime boundary

The adapter targets the already installed openclaw-local:latest image containing OpenClaw 2026.5.18. Its installed documentation and plugin implementation were inspected before choosing this interface. The plugin runs prompt-only inference with disableTools: true, validates JSON Schema, and returns parsed JSON together with the actual provider/model. Other OpenClaw versions require explicit inspection and adapter tests; setup fails instead of assuming newer documentation applies.

The isolated dataset-prototype-openclaw container has:

  • A gateway published on 127.0.0.1:18891 with generated bearer authentication.
  • Only the llm-task plugin/tool allow-listed; no tools exposed to the inference.
  • Verde as its sole provider/model route, with no fallback models.
  • Four role identifiers and independent prompt-only role calls.
  • No Slack/channels, scheduled model heartbeat, existing working-group state, repository data, home directory, or Docker socket mounted.
  • Separate writable runtime state and an empty workspace under ignored run/; a read-only container filesystem, dropped capabilities and resource limits.

The container listens on its internal network interface so Docker port forwarding works; only the host loopback interface is published. Gateway authentication is operator-level for this isolated instance, not a per-role authorization system. Never publish its port or share its bearer token with untrusted clients.

Start, inspect, and use

Docker must be running, and the existing image must be installed. No image build, model download, or OpenClaw upgrade is performed automatically.

prototype openclaw start --verde-env /path/to/openclaw_container/.env
prototype openclaw status
prototype propose run/real-sentinel --backend openclaw \
  --runtime run/openclaw --output run/my-openclaw-proposal

run/real-sentinel is an acquired, checksummed snapshot from the real-data workflow. Use a fresh output directory each run. Allow the gateway a few seconds to finish startup before calling it. Status reports container state and configuration; it does not prove successful provider inference. The proposal command is the live contract test.

The trusted env file must contain:

VERDE_LLM_BASE_URL=https://llm-api.cyverse.ai/v1
VERDE_LLM_API_KEY=YOUR_LOCAL_KEY
VERDE_LLM_DEFAULT_MODEL=js2/gpt-oss-120b

These are configuration examples, not a request to place real keys in tracked files. The operator's process environment takes precedence. The reader selects only these three fields and never executes shell expressions from the file. Slack and other credentials are not copied. A private generated runtime env file contains only the credential and variables needed by this isolated gateway. run/ is ignored, its runtime directory has mode 0700, and credential files have mode 0600. Docker administrators remain trusted and can inspect container secrets.

The existing key and js2/gpt-oss-120b availability were verified using the configured Verde /models endpoint. Use the URL granted to your team: current AI-VERDE documentation also describes a newer generic base URL, so do not rewrite a working team configuration from an example.

To stop just this prototype runtime while retaining its artifacts:

prototype openclaw stop

Stop verifies the container's ownership label. The existing working-group containers, models, Slack setup, sibling repository and credentials are not reconfigured. Starting again requires an explicit stop first; old configuration digests are retained under the ignored runtime's config-history/ directory.

Compatibility findings

The installed version resolves bare js2/gpt-oss-120b as a provider-qualified reference. Task requests must instead use verde/js2/gpt-oss-120b. The response must identify provider verde and model js2/gpt-oss-120b; substitutions fail.

Verde's server rejected an empty tools: [] array. The installed OpenClaw model schema supports compat.supportsTools: false, which removes the unsupported tool schema surface while preserving tool-free inference. No dummy tool or execution capability is introduced to work around this error. Both failed live attempts were retained in separate proposal directories.

Evidence and limits

Each role records the exact submitted task envelope, schema, prompt/response content digests, role, configured and returned provider/model, runtime version, configuration digest, and returned plugin envelope. Prompt/response references are attached to recipe proposal evidence. Unknown roles, changed runtime policy, wrong provider/model, transport failures and invalid JSON/schema results fail without an Ollama or mock fallback.

This version of the OpenClaw plugin returns validated, normalized JSON, not raw provider text, token usage, or a model-weights digest. It can remove Markdown fences before parsing. The adapter labels this response scope explicitly, leaves usage unknown, and records the model digest as unavailable. It does not misrepresent a configuration/image digest as the model weights or claim an exact raw-provider transcript. Runtime provenance is additional to scientific provenance.

A completed model call is not approval to transform data. Inspect discovery, semantic mappings, proposed recipe, model QA, deterministic QA, review status and traces. All model proposals require human review. QA rejection/revision remains a veto bound to scientific recipe content.

Verified live result — 2026-10-21

The four roles completed through this gateway against an actual Sentinel snapshot. Every response reported provider verde, model js2/gpt-oss-120b, and passed both plugin and client contract validation. The local artifacts are under run/openclaw-verde-no-tools/, including all four traces and prompt/response references.

Model QA returned request_revision, raising calibration/target-context, license label, and alignment-evidence concerns. Some claims may reflect model confusion about the intentionally retained raw-DN target; they require expert evaluation, not silent acceptance as scientific fact. Deterministic QA enforced the veto and no harmonized data or representations were executed from this proposal. Runtime integration success is separate from scientific recipe acceptance.

The gateway is left running for subsequent work. prototype openclaw status shows its current container state; prototype openclaw stop stops only this isolated instance. The expanded offline suite passes 79 tests, and the strict documentation build succeeds.

Metadata specialists and context budget

The streaming module adds units, support, and rights specialists alongside semantic analysis, recipe construction, and adversarial QA. Their shared dialogue is metadata-only; see streaming harmonization.

A live multi-review run exposed the installed runtime's default 20,000-token compaction reserve within the configured 32,768-token context window. A roughly 13,182-token prompt then failed the gateway precheck, returning HTTP 500/empty output. Inspection of the installed schema and session-management documentation confirmed agents.defaults.compaction.reserveTokens and reserveTokensFloor. Both are now 8,192 for this tool-free one-shot runtime, above the task's 4,096-token output cap. The request byte limit still applies, and oversized contexts fail rather than silently removing scientific evidence. Runtime configuration history and failed task transcripts remain retained.

Deterministic backbone and model trust

VERDE interprets, selects, parameterizes, compares and proposes. It does not author scientific computation. New model plans select closed, source-bound capabilities; metadata coefficients and mathematical operations are resolved by tested code. Schema validation, scientific checks, existing vetoes and human decisions remain separate gates. Approved recipes execute without a model. See deterministic execution and model trust for the registry, IR contracts, examples, provenance and current limits.