Spec Kitty

Quickstart — ATDD / red-first strategy & entry-point map

Every defect carries an issue-pinned scenario demonstrably RED before the fix, GREEN after (NFR-002). Entry points below are pre-existing unless marked NEW.

Per-SC red-first map

SC	Defect	WP	Test entry point	Red-first shape
SC-001	write-side record + populated location + resolvable ref, parity with →planned on a shared fixture	WP01	tests/specify_cli/cli/commands/agent/test_tasks_move_task_seam.py; parity in tests/characterization/test_trio_pure_cores.py	reject onto in_review→in_progress with feedback F; assert committed cycle, populated body == F, review_ref resolvable; compare field-for-field by value to the →planned record from the same F
SC-002	no-rationale rejection on re-implement edge refused	WP01	test_tasks_move_task_seam.py	no-rationale in_review→in_progress → observable non-zero refusal; paired with an accepted with-rationale rejection on the same fixture
SC-003	fix-mode generation failure → visible warning, no silent feedback-less substitution	WP02	tests/agent/test_workflow_review_cycle_pointer.py; NEW tests/agent/test_build_implement_prompt_lines.py	force generate_fix_prompt to raise → assert visible (console) warning; positive control: success path renders feedback
SC-004	feedback renders on single-branch AND coordination topologies	WP02	test_workflow_review_cycle_pointer.py (add coord fixture)	coord half is red-first (F4 independent-red): with a record already present, coord render finds no review_ref until the STATUS_STATE re-route lands; single-branch is the parity control
SC-005	end-to-end, no pre-seeding: reviewer text → reject re-implement edge → same text in regenerated prompt, both topologies	WP01+WP02 (integration)	test_workflow_review_cycle_pointer.py (end-to-end flow)	one continuous flow, NO fixture pre-seeding of the record; fails if any of the four write gates OR the render path is unfixed (anti-mask)
FR-008 (conditional #3451)	review-cycle counter double-increment on re-implement edge	WP01	test_tasks_move_task_seam.py	ONLY if it reproduces: assert cycle_number increments exactly once (not twice) on a re-implement rejection

NFR-003 non-regression guard (WP01)

Keep existing arbiter-override coverage green (the in_review → {approved,done} forward-emit path, is_arbiter_override / arb_review_ref). The predicate excludes those edges by construction; the existing tests are the guard that proves no regression.

Pure-unit coverage (Sonar new-code, both WPs)

False) in tests/characterization/test_trio_pure_cores.py; a focused test for the extracted _mt_persist_rejection_cycle helper.

• WP01: is_review_rejection_edge truth-table unit test (all lane pairs incl. aliases + arbiter-forward =
• WP02: direct branch tests for build_implement_prompt_lines and the re-routed read helpers.

Blast radius to run (per CLAUDE.md test policy)

make test-fast + tests/specify_cli/cli/commands/agent/ + tests/agent/ + tests/characterization/ (the module + owning-subsystem dirs for the touched files). Not a whole-repo suite.

Reference commands

# reproduce the write-side loss (red) on the re-implement edge
.venv/bin/spec-kitty agent tasks move-task WP01 --to in_progress --review-feedback-file feedback.md
# inspect the emitted event's review_ref (should be resolvable review-cycle://, not synthetic review:WP01)
# regenerate the fix-mode prompt and confirm the feedback text is present
.venv/bin/spec-kitty implement WP01