lab: wfbench — Workflow Editor Agent Benchmark harness (stakwork 58313 twin)

[Evanfeenstra]

Stacked on #1648 (needs meta/validate-workflow + the eval/* scoring steps). Design + step map: mcp/src/lab/plans/wfbench-harness.md.

What

wfbench-run is the vein twin of stakwork 58313 (+ nested 54419 / 55741 / 55740 / 58114 / 58414 / 57425 / 58312):

• roster (graph/* steps, params.namespace): EvalSet { id: task_slug } → EvalRequirement×N { id: <slug>-<criterion_id> } (HAS_REQUIREMENT, ordered) → EvalTrigger { id: <slug>-<runId>, workflow_id, workflow_version_id, workflow_input, project_id } → HAS_BASELINE_TRIGGER when the EvalSet had no trigger before, else HAS_TRIGGER. Runs in parallel with the author.
• author: core agent + agentTools: ["meta/*"] (editor tool only) builds wfbench-<slug>; the version it shipped is resolved as vpin || vactive with a published-vs-vbefore gate (the gaia-evolve-gen lesson).
• gate + rerun: wfbench/check-input-keys (the input.<key> references in the produced YAML vs the launch payload; mismatch or empty body = harness error, no launch) → meta/run-workflow (own runId = 58313's project_id) → wfbench/classify-run (a runtime-failed rerun is still judged).
• judge: wfbench/build-materials inlines workflow YAML + custom step sources + launch payload + run output + expected output into one block → wfbench-judge-criterion per criterion (agent schema mode, nothing to read; crash = { error } = honest FAIL) → eval/aggregate-scores.
• record: wfbench/build-eval-output → graph/create-batch-triplet: EvalTrigger -HAS_OUTPUT-> EvalTriggerOutput { id: <slug>-<runId>, result, verdict, score, max_score, n_passed, n_total, judge_model }, and per criterion EvalTriggerOutput / EvalRequirement -HAS_CRITERION_RESULT-> CriterionResult { id: <slug>-<runId>-<criterion_id>, … }.
• callback: wfbench/webhook-body = 58313's 4-way resolve_webhook_payload with the exact Hive keys (harness_error bodies carry no score fields); POSTed to webhook_url and returned as the run's output (58313's set_output diverged on error paths).

Graph attrs are limited to what the jarvis ontology declares (vein rejects the rest): EvalSet.project_id and the name on EvalTrigger/EvalTriggerOutput are omitted; 58312's HAS_CAUSE → Workflow_version is not written (no such relationship). Everything else matches and the smoke asserts it.

Not in v1: batch runner, stakwork-fetch/port-task, edit-existing. No stakwork credentials needed.

Verification

• npx tsc --noEmit clean.
• npx tsx src/lab/wfbench/smoke.ts — offline: seeds, discovers 22 steps, static-validates both workflows (0 errors, 0 warnings) through the authoring capability, drives all 10 pure steps, checks every node payload's attributes and every edge against JARVIS_ONTOLOGY. All pass.
• Not yet run live (needs Neo4j + ANTHROPIC_API_KEY).

[Evanfeenstra]

Added on this branch: core pack step (vein) and removed the three per-experiment pack-result copies.

Live run of wfbench-run against local Neo4j + Anthropic (before this last commit): 4/4 pass, 55s, author $0.12 / 8 tool calls. Roster, EvalTriggerOutput and CriterionResult chain verified by Cypher in the wfbench namespace — edges and ids match 58313/58312. The produced workflow used gaia/pack-result as its combiner, which is what motivated the core step.

Verification after the change: vein npm test 665/665 (incl. new pack.test.ts); mcp tsc clean after refresh-vein; smokes green: wfbench, harvey deliver / smoke / evolve-smoke, gaia smoke / evolve-smoke.