Skip to content

v2.0 — Wire the Critic node #4

Description

@Ekam-Bitt

The Critic sits between coder and reviewer and asks what the coder could not see from inside the work: which assumptions are load-bearing, which files were never read, whether a materially different fix is better, what else could break.

Already present: CriticOutput in contracts.py and prompts/critic.md. Wiring it is one graph edge — the seam was built in Phase 1A specifically so adding it would not be a schema migration.

Why deferred: introducing a second reasoning node before the repair loop is reliable means debugging two new failure modes at once. The repair loop has now fired live exactly once (the first click solve needed two attempts), which is not enough evidence that it is reliable.

Measure it as a rung: Stages in graph/router.py gates stages for ablation, so the Critic should be added there and evaluated the same way — does it raise solve rate enough to justify another model call per attempt?

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    deferredDocumented as intentionally unbuilt in v1.0enhancementPlanned work, deliberately deferred

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions