Drive healing from a coding agent¶

heal exposes its run store and fix engine over an MCP server, so a coding agent (Claude Code, etc.) can triage failures, inspect evidence, and apply fixes with full judgment — the missing half of in-run healing.

Start the server¶

heal mcp results/        # stdio transport over a run's store

What the agent can do¶

Tool	Purpose
`list_failures`	ids, statuses, failure classes
`get_failure_bundle(event_id)`	full evidence: diagnosis, RCA, DOM/screenshot excerpts, attempts, fix proposal
`get_fix_proposals`	proposals with blast radius and usage sites
`apply_fix(event_id, in_place, force)`	apply a fix — `shared` blast radius is never applied in place; the agent gets the patch + usage list to decide
`healing_history(db_path)`	cross-run hotspots

The agent skill¶

The repo ships an agent skill (skills/heal/SKILL.md) describing the triage → inspect → fix workflow: review failures, respect blast radius (apply local fixes, reason about shared ones), never force a dirty tree, and read the RCA to tell "test outdated" from "application changed". Point your coding agent at it.

Why delegate¶

The in-run engine can heal and propose, but it can't make the judgment calls a human or coding agent can: whether to change a variable or one usage, whether a shared fix is safe, whether an unhealed failure is a real app regression. MCP mode closes that loop.