pr-review workflow: intake, anatomy, lenses, reviewers, synthesis, output #128
Notifications
Due Date
No due date set.
Reference: sam/archie#128
Reference in new issue
Block a user
Each agent call recorded as a StepExecution; reviewer fan-out as children. See docs/prds/pr-review-agent.md.
Close reason
Stage wiring landed in
62b4fbf4(on top of the prior pass'sf20d2263,b76e63d6): PRSource contract + fake, and phases 1/2/3/4/7/9 wired as workflow.Stage functions in prreview_stages.go, registered as the pr-review workflow via legacyBuiltinWorkflows. Phase 4's reviewer StepExecution status distinction (StepFailed for turn-cap/deadline exhaustion vs StepSucceeded for a clean zero-findings run) implemented via a bespoke runReviewerAgent bypassing the generic RunAgentChild, which only maps Go errors. task check clean on the final tree. Phases 5/6/8 (verification, coverage/consistency, merge gate) remain archie-core-afbk.4/.5's scope.Imported from beads
archie-core-afbk.3· GitHub #1202 · type task · priority P1 · created 2026-09-25 · closed 2026-09-27sam wrote (originally 2026-09-27):
Partial progress (not closing -- scope remains open): commits
f20d2263(PRD refinements) andb76e63d6(phase 1/3 deterministic core: depth classification, dimension merge/cap, hallucination-check dimension), both task-check clean.Remaining for this bead: Stage wiring for intake/anatomy/lenses/reviewers/synthesis/output against the existing workflow engine (AgentStage, tc.RunAgentChild, StepExecution tree). Two concrete findings from this session that the next pass should start from:
tc.RunAgentChild already supports concurrent fan-out from inside one Stage.Run body (each call independently records a StepExecution child under the stage's own StepID) -- the execution-tree PRD's 'no parallel sibling steps' out-of-scope note is about the engine's top-level Stage list, not a stage's internal goroutines. Phases 3 (3 lenses), 4 (up to 8 reviewers) and 9 (polish pool of 8) do not need an engine change, just goroutines + a bounded semaphore inside each stage.
No existing capability (Forger/Trees) fetches an arbitrary external PR's diff text or materializes a .git-free snapshot of its head -- every existing path (Trees.Diff, Trees.Snapshot, StageReview's runReview) operates on the current task's own worktree the daemon is actively building, not a third-party PR being reviewed. Task.Owner/Repo/PRNumber already exist and are sufficient to identify a review target (no schema change needed there), but phase 1's diff statistics and the Isolation section's snapshot requirement need a narrow new contract -- recommend a PRSource interface on TaskContext (same pattern as Reviewer/Trees), implemented against real Forge/Trees calls by whichever bead wires triggers (afbk.7), with a test fake letting afbk.3's Stage code be built and tested today.