Skip to content
Documentation Impact-aware proof & cache

Impact-aware proof and cache

Running the entire repository test suite for every small patch is wasteful. AgentDiff computes the minimum strong proof from the impact graph and caches deterministic results.

Proof levels

Level When What runs
static no test commands derivable compile / vet / type-check only
targeted normal source change static checks + tests covering affected modules
full high-risk change the complete repository proof

High-risk changes always widen to full:

  • dependency files (uv.lock, package.json, go.mod, Cargo.lock, …)
  • CI workflows (.github/)
  • Dockerfiles, Makefiles, CMakeLists
  • build-system files (pyproject.toml, nx.json, …)
  • agent instruction files (AGENTS.md, CLAUDE.md, .codex/, …)
  • security paths (.env*, keys, credentials)

The decision is pure path/import classification:

agentdiff impact --paths src/auth.py     # targeted
agentdiff impact --paths uv.lock         # full

Impact graph

The graph is compiled by agentdiff bootstrap (.agentdiff/repo-graph.json) with static import analysis for Python, JavaScript/TypeScript, Go, and Rust:

changed file → affected modules → affected tests → affected build targets

Existing project systems are integrated where present: npm/pnpm workspaces, Nx, Bazel, Cargo workspaces, Go workspaces, uv workspaces.

Deterministic proof cache

Proof results are cached under .agentdiff/cache/proof and are keyed by every input that can influence the verdict:

  • base digest (sealed pre-run source snapshot)
  • patch digest (run-independent content digest of the exact mutation set)
  • dependency lock digest (trust lock lockfile digests)
  • runtime image digest
  • proof plan digest (exact argv phases)
  • target (static / targeted / full)

A cache hit is only possible when every input is byte-identical; any change is a miss. Entries carry their own SHA-256 integrity manifest, and a tampered entry is treated as a miss. Cache hits are surfaced in the proof result:

Proof verdict: PROVEN
Cache:         HIT (from run 20260819T…)

agentdiff proof cache-status

Safe optimization, not blind optimization

  • High-risk changes never use a cached targeted result; they widen to full.
  • The cache never skips integrity verification or promotion gates.
  • Promotion still requires a PROVEN proof bound to the current run's immutable manifest — a cached verdict is re-sealed per run with the same phase digests.