Agent-Trace is a Claude Code plugin that provides dependency-aware remediation. Before an agent edits risky code, it maps the blast radius using a hybrid graph + curated contracts model.
Three layers:
- Repo Universe — Machine-generated graph (files, symbols, imports, routes, tests) + human-curated overlays (invariants, ownership, source-of-truth)
- Agent Query — Script interface that traverses the graph to return blast radius, impacted tests, invariants, and confidence before edits
- Safe Remediation Skill — 7-step workflow that mandates querying before risky edits
| Command |
What It Does |
/trace |
Full safe-remediation workflow |
/map |
Build/update the repo universe |
/query |
Query impact of changing a target |
/validate-universe |
Validate repo universe integrity |
| Agent |
Role |
impact-analyzer |
Analyzes blast radius from query results |
contract-mapper |
Maps current contracts before edits |
manifest-builder |
Guides manifest building |
| Script |
Purpose |
build_manifest.py |
Scan codebase, generate nodes + edges |
query_impact.py |
Query blast radius for a target |
validate_universe.py |
Validate manifest integrity |
check_query_schema.py |
Validate query output schema |
common.py |
Shared utilities |
- Hybrid model: machine-generated graph + human-curated overlays — avoids stale inventories and semantics-blind code graphs
- Agent query is mandatory before risky edits — the skill enforces this
- Freshness and confidence metadata on every node and edge — agents can assess certainty
- File-based JSON storage (not graph DB) — simple, inspectable, CI-friendly
- Generated and curated facts always distinguishable via
source field — prevents false certainty
prompts/ — LLM-agnostic core prompts (all adapters wrap these)scripts/ — Python tooling (copied into target repos by install.sh)templates/ — Example curated overlaysreferences/ — Skill reference docsdocs/specs/ — Locked schema and contract specifications