| name | neograph-harness-authoring |
| description | Compile and run bounded PR review, bug triage, or research synthesis panels through the NeoGraph Harness MCP tools. Use when a task benefits from multiple schema-constrained workers, strict preflight validation, read-only workspace policy, durable host-brokered resume, compact polling, and URI-linked detailed artifacts. |
NeoGraph Harness
Goal
Run a multi-worker panel without giving the main conversation every worker
trace or allowing an unvalidated workflow to execute.
Procedure
- Call
neograph_schema; use only presets and fields returned by this build.
- Build one request with a precise objective, acceptance criteria, bounded
budgets, and a JSON output schema for every worker.
- For review work, use
pr_review_panel, set policy.read_only to true, and
set policy.evidence_required to the evidence fields required in every
finding schema.
- Give each worker only the tool IDs it needs. Mark read-only tools and list
path-bearing string arguments in
path_arguments. Set explicit
policy.workspace_roots whenever such arguments exist.
- Call
neograph_compile. If ok is false, fix diagnostics by phase,
path, and source; never call start with a rejected request.
- Call
neograph_start with the retained artifact_id.
- Poll
neograph_get with run_id. If status is awaiting_tool_results or
input_required, fulfill only the returned pending call, then call
neograph_resume with the same run_id, exact call_id, and a result that
conforms to result_schema. Treat an identical duplicate as acknowledged;
never substitute a different call ID.
- Continue polling until status is terminal. Keep the compact result in the
main context.
- Dereference a returned
neograph://runs/... URI through neograph_get
with its run_id only when the final answer needs worker details or the
execution trace.
- Report partial, zero-findings, timeout, cancelled, expired, max-step, and failed
outcomes exactly; do not turn them into a generic success.
Anti-Patterns
- Do not skip
neograph_compile for an inline request.
- Do not attach a broad tool catalog to every worker.
- Do not configure path-bearing tools without workspace roots.
- Do not treat malformed or empty worker output as an empty findings list.
- Do not fetch detailed traces before the compact result proves they are needed.
- Do not add write-capable tools to a read-only review.
- Do not retry a host result with modified data after its call ID was consumed.
- Do not assume MCP Tasks is core protocol support. Use stable
neograph_get
polling unless the server and the individual request explicitly opt into the
experimental io.modelcontextprotocol/tasks extension.
Example
For a PR review, collect the diff with the host's repository tools, place it in
the task objective, and use two workers with distinct correctness and security
instructions. Require file, line, and evidence in each finding. See
docs/HARNESS_MCP.md for the complete request and host setup commands.