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
Priority: P2 — optional polish on the rules layer; the last thing built in the v2.0.0 pivot Effort: ~3–5 days AI substrate: Claude Code (claude -p) primary; Local / BYOK fallback — see #57 Depends on:#52 (rules layer), #57 (AI config), #46 (issue context — for the keystone check)
Right column of the risk-flagging pane. Optional, off by default. Calls the configured AI substrate (#57) — primarily the user's claude -p — for the fuzzy reasoning the rules layer can't do. Suggestion-only, never blocking, always cancelable.
Checks (run only when the AI layer is enabled)
Summary — "what does this change do, in 2 sentences."
Diff ⊨ acceptance criteria (keystone) — "does this diff satisfy the linked Linear issue's acceptance criteria?" Runs only when issue context ([feat] Issue context panel in commit/diff UI #46) is present. This is the most valuable check in the AI layer and the reason the substrate is claude -p rather than a 7B local model — it needs real reasoning over both the diff and the criteria. Output: per-criterion satisfied / partial / not-addressed + a one-line note.
Scope-creep — "what in this diff looks unrelated to the stated intent?" High value for sprawling agent-authored diffs.
Priority: P2 — optional polish on the rules layer; the last thing built in the v2.0.0 pivot
Effort: ~3–5 days
AI substrate: Claude Code (
claude -p) primary; Local / BYOK fallback — see #57Depends on: #52 (rules layer), #57 (AI config), #46 (issue context — for the keystone check)
Right column of the risk-flagging pane. Optional, off by default. Calls the configured AI substrate (#57) — primarily the user's
claude -p— for the fuzzy reasoning the rules layer can't do. Suggestion-only, never blocking, always cancelable.Checks (run only when the AI layer is enabled)
claude -prather than a 7B local model — it needs real reasoning over both the diff and the criteria. Output: per-criterion satisfied / partial / not-addressed + a one-line note.Scope
claude -p --output-format jsoncall per check (or one batched call returning all three) — no agent loop.Acceptance criteria
claude -pand returns typed results