Renders a completed Claude Code dynamic Workflow run (wf_<runId>.json) as a self-contained, offline HTML page with a run→phase→agent tree and a parallelism timeline. Trigger when the user asks to "visualize the workflow run", "render wf trajectory", "trace a workflow", or "show the last workflow execution".
No-Python remote drop-in for plan-validator - runs the adversarial plan review on remote models over the Interactions API (curl + ADC from caller subagents; configurable roster; Vertex fallback), with a synthesis model casting an extra vote and nominating the first domino. Symptoms - "run the interactions plan panel", "remote plan validation without Python", "validate this plan via the Interactions API", "GEAP plan validation, no local scripts", geap-plan-validator wanted but no venv/Python available.
No-Python remote drop-in for spec-validator - runs the adversarial spec review on remote models over the Interactions API (curl + ADC from caller subagents; configurable roster; Vertex fallback), with a synthesis model casting an extra vote. Symptoms - "run the interactions spec panel", "remote spec validation without Python", "validate this spec via the Interactions API", "GEAP spec validation, no local scripts", geap-spec-validator wanted but no venv/Python available.
Use when a drafted implementation plan spans territories no single agent can hold at once — the spec's intent, multiple subsystems of the real codebase, and the delivery pipeline — BEFORE plan-validator, to improve the plan by deliberation rather than attack. Dispatches delegate agents each assigned a different territory to deep-read and speak for, who deliberate over bounded rounds until they converge on a single jointly revised plan, negotiating the trade-offs (migration strategy, group boundaries, scope) that a validator can only flag, never decide. Symptoms - "deliberate on this plan", "improve this plan from multiple perspectives", "the plan touches three subsystems", "we need to pick a migration strategy", plan.md spans code no one context window can deep-read whole, resolving the unconfirmed tail of a plan-validator run.
Use when a drafted spec depends on knowledge that is siloed across stakeholders, documents, or repos — BEFORE adversarial validation — to improve the spec by deliberation rather than attack. Dispatches delegate agents with deliberately DISJOINT context bundles (product, engineering, ops/security) who deliberate over bounded rounds, relayed verbatim by the orchestrator, until they converge on a single jointly revised spec. Symptoms - "deliberate on this spec", "improve this spec from multiple perspectives", "get product/eng/security input on the spec", "the constraints live in different places", spec touches systems whose limits no single context window can hold, resolving the unconfirmed tail of a spec-validator run.
Remote drop-in alternative to plan-validator - runs the adversarial implementation-plan review on Vertex AI foundation models (3 configurable skeptics + a synthesis model) instead of local subagents, attacking the plan text for ordering defects, hidden assumptions, and missing failure handling. Symptoms - "run the remote plan panel", "GEAP plan validation", "validate this plan with cloud/Vertex models", "get a second opinion on this plan from other models", plan-validator requested with remote or Gemini/Claude-on-Vertex models.
Remote drop-in alternative to spec-validator - runs the adversarial spec review on Vertex AI foundation models (3 configurable skeptics + a synthesis model) instead of local subagents, for an independent perspective from different model families. Symptoms - "run the remote spec panel", "GEAP spec validation", "validate this spec with cloud/Vertex models", "get a second opinion on this spec from other models", spec-validator requested with remote or Gemini/Claude-on-Vertex models.
Use after code is written and before merge, to confirm the implementation actually does what it claims. Dispatches independent skeptic agents that read the diff and surrounding code with a default-to-reject posture, hunt for real defects (or refute explicit acceptance claims), and assign a corrected severity — keeping only findings confirmed by a 2-of-3 majority. Symptoms - "validate this implementation", "did this actually work", "review this diff adversarially", "verify these findings are real", after completing a feature/task, before merging to main.