Spec Kitty

└─ kitty-specs
   └─ Rejection feedback reaches the implementer

Mission Run:

📚 Docs ↗

Data Model — Rejection feedback reaches the implementer

This mission touches no new persistent schema; it corrects the routing and gating of EXISTING entities. The "model" here is the set of domain entities from spec.md and the state edges they flow across.

Entities

Rejection edge (value)

A transition of a WP out of review. Two members handle feedback and must behave identically:

in_review → planned (full restart) — the reference oracle; already records feedback correctly.
in_review → in_progress (re-implement) — the defective edge (#4899).

Classified by the new pure predicate is_review_rejection_edge(old_lane, target_lane): resolve_lane_alias(old) == IN_REVIEW and resolve_lane_alias(target) in (PLANNED, IN_PROGRESS), unioned with target == PLANNED (to preserve any-source →planned rollback). Arbiter-forward edges (in_review → {approved, done}) are deliberately excluded (NFR-003).

Review-cycle record (durable, committed artifact)

ReviewCycleArtifact under tasks/<wp_slug>/review-cycle-<N>.md, written by persist_rejected_review_cycle_for_rollback (tasks_verdict_persistence.py:930) and committed via the CoordCommitRouter. Fields consumed by parity (FR-003): cycle_number, reviewer, verdict, feedback body, reference. Partition: WORK_PACKAGE_TASK (routed via _review_cycle_wp_dir) — same on both topologies. Invariant (FR-008): cycle_number increments exactly once per rejection.

Feedback reference (pointer left on the event)

event.review_ref on the emitted status.events.jsonl transition. Two shapes:

The re-implement edge wrongly emits this today. Partition of the event log: STATUS_STATE → COORD on coord topologies (the F4 defect surface).

Resolvable pointer — review-cycle://… (real feedback location). Required outcome of a rejection edge.
Non-resolvable synthetic marker — review:<WP> etc. (is_non_resolvable_review_ref skips these).

Fix-mode prompt (regenerated implementer prompt)

Produced by generate_fix_prompt from the latest ReviewCycleArtifact (workflow_executor.py::implement_try_render_fix_mode_prompt). Must carry the reviewer's feedback text (FR-005). On generation failure the system must surface a visible warning (FR-006) rather than silently substitute a feedback-less full prompt.

State edges (before → after this mission)

stateDiagram-v2
  [*] --> in_review
  in_review --> planned: reject-restart (records feedback — oracle, unchanged)
  in_review --> in_progress: re-implement — BEFORE: feedback SILENTLY LOST\nAFTER: records feedback == planned edge
  in_review --> approved: arbiter-forward (NFR-003 — untouched)
  in_review --> done: arbiter-forward (NFR-003 — untouched)
  planned --> in_progress
  in_progress --> for_review

Field-level parity contract (FR-003, SC-001)

For identical reviewer feedback text F, the record produced by in_review → in_progress must equal the record produced by in_review → planned by value:

Field	Parity requirement
committed artifact	both produce a committed `review-cycle-<N>.md`
feedback location (body)	both populated; resolving the reference yields text == `F` (not empty, not a placeholder)
`event.review_ref`	both a resolvable `review-cycle://…` pointer (not synthetic)
reviewer / verdict / `cycle_number`	equal field-for-field

A record whose fields all exist but whose feedback content is empty/placeholder does NOT satisfy parity (spec US1 scenario 3).

Read-path partition model (F4 / FR-007)

Read	Kind	Partition (coord)	Correct source
review-cycle artifact	`WORK_PACKAGE_TASK`	PRIMARY	`_review_cycle_wp_dir` (already correct)
event log (`review_ref`)	`STATUS_STATE`	COORD	must route `read_dir(STATUS_STATE)` — the render fix

Single-branch topology collapses PRIMARY and COORD to repo_root, so the re-route is a no-op there and cannot regress single-branch rendering.