You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Tiered sensors: deterministic checks every change, inferential review at the end #250
Verify-before-stop currently recognises only tests and builds. Extend sensors to include linters, type checkers and structural checks, and separate them into two tiers:
Computational (fast, deterministic): lint/typecheck/test commands, configurable, runnable after edits and before stop.
Inferential (LLM review/judge): slower, run once at the end (reuse existing review).
Acceptance:
Config for a list of computational check commands (pass = exit 0), counted by verify-on-stop.
Clear ordering: cheap checks first, review last.
Off by default; existing behaviour unchanged. Tests with fake models. Docs updated.
Slice 1 is in #255: the computational tier (verify_on_stop.checks / --verify-check, run in order at stop, cheapest first, counted by verify-on-stop, verify.check JSON event, docs).
Remaining work:
Inferential tier: optionally run the existing review (the --review uncommitted-changes flow) once at the end, only after every computational check passes, so ordering is cheap checks first and review last. Needs a config key/flag (off by default), a decision on how review findings feed back (re-prompt vs. report only), JSON event and docs.
Activity
vicgupta commentedon Oct 30, 2026
Slice 1 is in #255: the computational tier (verify_on_stop.checks / --verify-check, run in order at stop, cheapest first, counted by verify-on-stop, verify.check JSON event, docs).
Remaining work:
Merge pull request #255 from vicgupta/amf/issue-250
Verify-on-stop: inferential review tier after computational checks (#250