prizmkit 1.1.125 → 1.1.126

package/bundled/dev-pipeline/templates/bugfix-bootstrap-prompt.md

@@ -180,24 +180,15 @@ After implement completes, verify:
 If `FAST_PATH=true` (≤ 2 tasks, obvious root cause), skip this phase entirely and mark checkpoint step `prizmkit-code-review` as `skipped` before continuing.
 
 Run `/prizmkit-code-review` with `artifact_dir=.prizmkit/bugfix/{{BUG_ID}}/`:
-- The skill chooses a platform-supported fresh read-only Reviewer, prefers direct access to the active checkout, and supplies changed content that a platform-forced worktree cannot see
-- The skill runs an internal review-fix loop (Reviewer → Main-Agent filter → Main-Agent fix, up to ten completed rounds) and writes `review-report.md`
-- `review-report.md` must contain `## Verdict`
+- The skill first runs up to ten Main-Agent self-review rounds, then semantic risk limits independent Reviewers to `low=0`, `medium=1`, `high=2`
+- High risk starts Reviewer 2 only after an accepted high finding was repaired, tested, and reconverged; Reviewer 3 is forbidden
+- The skill preserves its independent Reviewer prompt and appends progress to `review-report.md`
 
-**Gate Check — Review Report**:
-After `/prizmkit-code-review` returns, verify the review report:
+**Gate Check — Final Review Result**:
 ```bash
-{{RUNTIME_HELPER_CMD}} text contains .prizmkit/bugfix/{{BUG_ID}}/review-report.md --pattern "## Verdict"
+{{RUNTIME_HELPER_CMD}} text final-verdict .prizmkit/bugfix/{{BUG_ID}}/review-report.md
 ```
-If GATE:MISSING:
-- Do not re-run `/prizmkit-code-review` in an unbounded report-repair loop.
-- Perform one bounded status check; retry `/prizmkit-code-review` at most once only if the missing report appears caused by a transient team/config/lock error.
-- If the report is still missing after that single check/retry, write `failure-log.md` with the spawn/skill error and last observable state, then stop with a clear recovery failure. Do not replace the independent Reviewer with Main-Agent self-review.
-
-Read `review-report.md` and check the Verdict:
-- `PASS` → proceed
-- `NEEDS_FIXES` → the skill reached the ten-round safety fuse or left an accepted finding unresolved; log remaining findings and proceed according to the task failure policy
-- `BLOCKED` → log the infrastructure reason and stop so the review can be retried
+Handle `REVIEW_PASS`, `REVIEW_NEEDS_FIXES`, and `REVIEW_BLOCKED` according to task policy. `REVIEW_INCOMPLETE` or `REVIEW_INVALID` means review did not complete: inspect progress once, write `failure-log.md`, and stop. Do not accept stale verdict text or replace required independent review with Main-Agent self-review.
 
 {{IF_BROWSER_INTERACTION}}
 

package/bundled/dev-pipeline/templates/refactor-bootstrap-prompt.md

@@ -139,22 +139,21 @@ Apply the browser behavior preservation protocol where UI behavior can be affect
 
 ### Phase 4: Review — Code Review & Behavior Verification
 
-Run `/prizmkit-code-review` with `artifact_dir=.prizmkit/refactor/{{REFACTOR_ID}}/` directly from the main orchestrator. The skill loads its skill-owned `references/reviewer-agent-prompt.md`, chooses a platform-supported fresh read-only Reviewer, prefers direct access to the active checkout, and supplies changed content that a platform-forced worktree cannot see.
+Run `/prizmkit-code-review` with `artifact_dir=.prizmkit/refactor/{{REFACTOR_ID}}/` directly from the main orchestrator. The skill first runs up to ten Main-Agent self-review rounds, then semantic risk limits independent Reviewers to `low=0`, `medium=1`, `high=2`. High risk starts Reviewer 2 only after an accepted high finding was repaired, tested, and reconverged; Reviewer 3 is forbidden.
 
 Review requirements:
 - verify refactor quality against spec.md and plan.md;
 - verify behavior preservation evidence;
 - run or confirm relevant tests;
-- write `.prizmkit/refactor/{{REFACTOR_ID}}/review-report.md` with `## Verdict`.
+- preserve append-only progress in `.prizmkit/refactor/{{REFACTOR_ID}}/review-report.md`.
 
-If the report is missing after one bounded status check, write `failure-log.md` and stop with a clear recovery failure. Do not replace the independent Reviewer with Main-Agent self-review.
-
-Read `review-report.md` and check the Verdict:
-- `PASS` → proceed.
-- `NEEDS_FIXES` → log unresolved accepted findings and proceed only when they do not block behavior preservation.
-- `BLOCKED` → log the infrastructure reason and stop so the review can be retried.
+**Gate Check — Final Review Result**:
+```bash
+{{RUNTIME_HELPER_CMD}} text final-verdict .prizmkit/refactor/{{REFACTOR_ID}}/review-report.md
+```
+Handle `REVIEW_PASS`, `REVIEW_NEEDS_FIXES`, and `REVIEW_BLOCKED` according to behavior-preservation policy. `REVIEW_INCOMPLETE` or `REVIEW_INVALID` means review did not complete: inspect progress once, write `failure-log.md`, and stop. Do not accept stale verdict text or replace required independent review with Main-Agent self-review.
 
-**Checkpoint update**: set step `prizmkit-code-review` to `completed` only after `PASS`, or after an explicitly accepted non-blocking `NEEDS_FIXES` result. Do not complete the checkpoint for `BLOCKED`.
+**Checkpoint update**: set step `prizmkit-code-review` to `completed` only for an accepted valid Final Result. Do not complete it for blocked, incomplete, or invalid reports.
 
 ---
 

package/bundled/dev-pipeline/templates/sections/bugfix-phase-review.md

@@ -12,19 +12,19 @@ If `FAST_PATH=true`, skip this phase intentionally:
 
 Then continue to the next checkpoint step.
 
-Otherwise, run `/prizmkit-code-review` with `artifact_dir=.prizmkit/bugfix/{{BUG_ID}}/`. The skill chooses a platform-supported fresh read-only Reviewer, prefers direct access to the active checkout, and supplies changed content that a platform-forced worktree cannot see.
+Otherwise, run `/prizmkit-code-review` with `artifact_dir=.prizmkit/bugfix/{{BUG_ID}}/`. The skill first runs a Main-Agent self-review loop for up to ten completed rounds, then semantic risk permits independent Reviewers at limits `low=0`, `medium=1`, `high=2`. High risk starts Reviewer 2 only after Reviewer 1 produced an accepted high finding that was repaired, tested, and reconverged. Reviewer 3 is forbidden.
 
-The skill runs its internal fresh Reviewer → Main-Agent filter → Main-Agent fix loop for up to ten completed rounds and writes `review-report.md`. Do not prescribe Agent launch parameters from this pipeline template.
+The skill owns its unchanged independent Reviewer prompt, current-checkout preference, forced-worktree changed-content fallback, one bounded infrastructure retry per required Reviewer position, and append-only progress in `review-report.md`.
 
-**Gate Check — Review Report**:
+**Gate Check — Final Review Result**:
 
 ```bash
-{{RUNTIME_HELPER_CMD}} text contains .prizmkit/bugfix/{{BUG_ID}}/review-report.md --pattern "## Verdict"
+{{RUNTIME_HELPER_CMD}} text final-verdict .prizmkit/bugfix/{{BUG_ID}}/review-report.md
 ```
 
-- `PASS` → proceed.
-- `NEEDS_FIXES` → log unresolved accepted findings and proceed only when they do not block acceptance criteria.
-- `BLOCKED` → log the infrastructure reason and stop so the review can be retried.
-- Missing report → one bounded retry, then write `failure-log.md` and stop if independent review remains blocked; do not replace it with Main-Agent self-review.
+- `REVIEW_PASS` → proceed.
+- `REVIEW_NEEDS_FIXES` → log unresolved findings and proceed only when explicit task policy allows them.
+- `REVIEW_BLOCKED` → log infrastructure evidence and stop so recovery can retry.
+- `REVIEW_INCOMPLETE` or `REVIEW_INVALID` → inspect progress once, write `failure-log.md`, and stop. Do not accept old or incidental verdict text and do not replace required independent review with Main-Agent self-review.
 
-**Checkpoint update**: Set step `prizmkit-code-review` to `completed` when review is complete, or `skipped` for FAST_PATH.
+**Checkpoint update**: Set step `prizmkit-code-review` to `completed` only for an accepted valid Final Result, or `skipped` for FAST_PATH.

package/bundled/dev-pipeline/templates/sections/phase-review-agent.md

@@ -1,26 +1,23 @@
 ### Review — Code Review
 
-Review the completed implementation before the scoped feature test gate runs. Use `artifact_dir=.prizmkit/specs/{{FEATURE_SLUG}}/` so the code-review skill can read spec, plan, context snapshot, implementation log, and current diff artifacts. Do not require or read a prior `/prizmkit-test` report here; test generation and test-fix looping happen in the next phase.
+Review the completed implementation before the scoped feature test gate runs. Use `artifact_dir=.prizmkit/specs/{{FEATURE_SLUG}}/` so the code-review skill can read spec, plan, context snapshot, implementation log, and current diff artifacts. Do not require a prior `/prizmkit-test` report here; the next phase owns the formal test gate.
 
-Run `/prizmkit-code-review` with `artifact_dir=.prizmkit/specs/{{FEATURE_SLUG}}/`. The skill chooses a platform-supported fresh read-only Reviewer, prefers direct access to the active checkout, and supplies changed content that a platform-forced worktree cannot see. Do not prescribe Agent launch parameters from this pipeline template.
+Run `/prizmkit-code-review` with `artifact_dir=.prizmkit/specs/{{FEATURE_SLUG}}/`. The skill first runs a Main-Agent self-review loop for up to ten completed rounds, then semantic risk permits independent Reviewers at limits `low=0`, `medium=1`, `high=2`. High risk starts Reviewer 2 only after Reviewer 1 produced an accepted high finding that was repaired, tested, and reconverged. Reviewer 3 is forbidden.
 
-The skill runs an internal review-fix loop (fresh Reviewer → Main-Agent filter → Main-Agent fix, up to ten completed rounds) and writes `review-report.md` to the artifact directory.
+The skill owns Reviewer launch parameters, current-checkout preference, forced-worktree changed-content fallback, one bounded infrastructure retry per required Reviewer position, and append-only progress in `review-report.md`.
+
+**Gate Check — Final Review Result**:
 
-**Gate Check — Review Report**:
-After `/prizmkit-code-review` returns, verify the review report:
 ```bash
-{{RUNTIME_HELPER_CMD}} text contains .prizmkit/specs/{{FEATURE_SLUG}}/review-report.md --pattern "## Verdict"
+{{RUNTIME_HELPER_CMD}} text final-verdict .prizmkit/specs/{{FEATURE_SLUG}}/review-report.md
 ```
-If GATE:MISSING:
-- Do not re-run `/prizmkit-code-review` in an unbounded report-repair loop.
-- Perform one bounded status check; retry at most once: inspect the skill output, `review-report.md` path, and any Reviewer agent spawn messages.
-- If the missing report is caused by team/config/lock errors from the Reviewer agent spawn or review skill error, retry `/prizmkit-code-review` at most once only if it appears transient.
-- If the report is still missing after that single check/retry, write `failure-log.md` with the spawn/skill error and last observable state, then stop with a clear recovery failure. Do not synthesize a passing report through Main-Agent self-review.
-
-Read `review-report.md` and check the Verdict:
-- `PASS` → proceed to next phase
-- `NEEDS_FIXES` → log the unresolved accepted findings and proceed according to the task failure policy (do not retry externally)
-- `BLOCKED` → log the infrastructure reason and stop so the review can be retried
-
-**CP-3**: Review complete, report written.
+
+Handle the marker:
+
+- `REVIEW_PASS` → proceed to next phase.
+- `REVIEW_NEEDS_FIXES` → log unresolved findings and follow the task failure policy; do not externally restart the review loop.
+- `REVIEW_BLOCKED` → log infrastructure evidence and stop so recovery can retry.
+- `REVIEW_INCOMPLETE` or `REVIEW_INVALID` → inspect the append-only progress once, write `failure-log.md`, and stop. Do not treat an old or incidental verdict as completion and do not synthesize a passing result through Main-Agent self-review.
+
+**CP-3**: Review has a valid Final Result and `review-report.md` preserves this execution's progress.
 

package/bundled/dev-pipeline/templates/sections/phase-review-full.md

@@ -1,29 +1,23 @@
 ### Review — Code Review
 
-Review the completed implementation before the scoped feature test gate runs. Use `artifact_dir=.prizmkit/specs/{{FEATURE_SLUG}}/` so the code-review skill can read spec, plan, context snapshot, implementation log, and current diff artifacts. Do not require or read a prior `/prizmkit-test` report here; test generation and test-fix looping happen in the next phase.
+Review the completed implementation before the scoped feature test gate runs. Use `artifact_dir=.prizmkit/specs/{{FEATURE_SLUG}}/` so the code-review skill can read spec, plan, context snapshot, implementation log, and current diff artifacts. Do not require a prior `/prizmkit-test` report here; the next phase owns the formal test gate.
 
-Run `/prizmkit-code-review` with `artifact_dir=.prizmkit/specs/{{FEATURE_SLUG}}/`. The skill chooses a platform-supported fresh read-only Reviewer, prefers direct access to the active checkout, and supplies changed content that a platform-forced worktree cannot see. Do not prescribe Agent launch parameters from this bootstrap prompt.
+Run `/prizmkit-code-review` with `artifact_dir=.prizmkit/specs/{{FEATURE_SLUG}}/`. The skill first runs a Main-Agent self-review loop for up to ten completed rounds, then semantic risk permits independent Reviewers at limits `low=0`, `medium=1`, `high=2`. High risk starts Reviewer 2 only after Reviewer 1 produced an accepted high finding that was repaired, tested, and reconverged. Reviewer 3 is forbidden.
 
-If the independent Reviewer cannot start or return usable output because of a transient platform error, retry at most once. If the second attempt fails, do not keep spawning variants or enter report-polling loops; write `failure-log.md` with the observable error and let the skill report `BLOCKED`. Do not replace the independent Reviewer with Main-Agent self-review.
+The skill owns Reviewer launch parameters, current-checkout preference, forced-worktree changed-content fallback, one bounded infrastructure retry per required Reviewer position, and append-only progress in `review-report.md`.
 
-The skill runs an internal review-fix loop (fresh Reviewer → Main-Agent filter → Main-Agent fix, up to ten completed rounds) and writes `review-report.md` to the artifact directory. The review skill owns strategy selection and forced-worktree content fallback.
+**Gate Check — Final Review Result**:
 
-**Gate Check — Review Report**:
-After `/prizmkit-code-review` returns, verify the review report:
 ```bash
-{{RUNTIME_HELPER_CMD}} text contains .prizmkit/specs/{{FEATURE_SLUG}}/review-report.md --pattern "## Verdict"
+{{RUNTIME_HELPER_CMD}} text final-verdict .prizmkit/specs/{{FEATURE_SLUG}}/review-report.md
 ```
-If GATE:MISSING:
-- Do not re-run `/prizmkit-code-review` in an unbounded report-repair loop.
-- Perform one bounded status check; retry at most once: inspect the skill output, `review-report.md` path, and any Reviewer agent spawn messages.
-- If the missing report is caused by team/config/lock errors from the Reviewer agent spawn or review skill error, retry `/prizmkit-code-review` at most once only if it appears transient.
-- If the report is still missing after that single check/retry, write `failure-log.md` with the spawn/skill error and last observable state, then stop with a clear recovery failure. Do not synthesize a passing report through Main-Agent self-review.
 
-Read `review-report.md` and check the Verdict:
-- `PASS` → proceed to next phase
-- `NEEDS_FIXES` → log the unresolved accepted findings and proceed according to the task failure policy (do not retry externally)
-- `BLOCKED` → log the infrastructure reason and stop so the review can be retried
+Handle the marker:
 
-**CP-3**: Review complete, report written.
+- `REVIEW_PASS` → proceed to next phase.
+- `REVIEW_NEEDS_FIXES` → log unresolved findings and follow the task failure policy; do not externally restart the review loop.
+- `REVIEW_BLOCKED` → log infrastructure evidence and stop so recovery can retry.
+- `REVIEW_INCOMPLETE` or `REVIEW_INVALID` → inspect the append-only progress once, write `failure-log.md`, and stop. Do not treat an old or incidental verdict as completion and do not synthesize a passing result through Main-Agent self-review.
 
+**CP-3**: Review has a valid Final Result and `review-report.md` preserves this execution's progress.
 

package/bundled/dev-pipeline/templates/sections/refactor-phase-review.md

@@ -1,20 +1,20 @@
 ### Phase 3: Review — Code Review & Behavior Verification
 
-Run `/prizmkit-code-review` directly with `artifact_dir=.prizmkit/refactor/{{REFACTOR_ID}}/`. The skill loads its skill-owned `references/reviewer-agent-prompt.md`, chooses a platform-supported fresh read-only Reviewer, prefers direct access to the active checkout, and supplies changed content that a platform-forced worktree cannot see.
+Run `/prizmkit-code-review` directly with `artifact_dir=.prizmkit/refactor/{{REFACTOR_ID}}/`. The skill first runs a Main-Agent self-review loop for up to ten completed rounds, then semantic risk permits independent Reviewers at limits `low=0`, `medium=1`, `high=2`. High risk starts Reviewer 2 only after Reviewer 1 produced an accepted high finding that was repaired, tested, and reconverged. Reviewer 3 is forbidden.
 
-Do not spawn a top-level Reviewer subagent. The code-review skill owns its internal fresh Reviewer → Main-Agent filter → Main-Agent fix loop for up to ten completed rounds. Do not prescribe Agent launch parameters from this pipeline template.
+Do not spawn a top-level Reviewer subagent. The code-review skill owns its unchanged independent Reviewer prompt, current-checkout preference, forced-worktree changed-content fallback, one bounded infrastructure retry per required Reviewer position, and append-only report progress.
 
-**Gate Check — Review Report**:
+**Gate Check — Final Review Result**:
 
 ```bash
-{{RUNTIME_HELPER_CMD}} text contains .prizmkit/refactor/{{REFACTOR_ID}}/review-report.md --pattern "## Verdict"
+{{RUNTIME_HELPER_CMD}} text final-verdict .prizmkit/refactor/{{REFACTOR_ID}}/review-report.md
 ```
 
-- `PASS` → proceed.
-- `NEEDS_FIXES` → log unresolved accepted findings and proceed only when they do not block behavior preservation.
-- `BLOCKED` → log the infrastructure reason and stop so the review can be retried.
-- Missing report → one bounded retry, then write `failure-log.md` and stop if independent review remains blocked; do not replace it with Main-Agent self-review.
+- `REVIEW_PASS` → proceed.
+- `REVIEW_NEEDS_FIXES` → log unresolved findings and proceed only when they do not block behavior preservation under explicit task policy.
+- `REVIEW_BLOCKED` → log infrastructure evidence and stop so recovery can retry.
+- `REVIEW_INCOMPLETE` or `REVIEW_INVALID` → inspect progress once, write `failure-log.md`, and stop. Do not accept stale verdict text or replace required independent review with Main-Agent self-review.
 
 Verify the review explicitly checks behavior preservation, public API compatibility, imports, tests, and rollback assumptions.
 
-**Checkpoint update**: Set step `prizmkit-code-review` to `completed` only after `PASS`, or after an explicitly accepted non-blocking `NEEDS_FIXES` result. Do not complete the checkpoint for `BLOCKED`.
+**Checkpoint update**: Set step `prizmkit-code-review` to `completed` only for an accepted valid Final Result. Do not complete the checkpoint for blocked, incomplete, or invalid reports.

package/bundled/dev-pipeline/tests/test_generate_bootstrap_prompt.py

@@ -882,8 +882,9 @@ class TestDirectOrchestratorImplementationPrompts:
             assert FORBIDDEN_TEST_CMD_PLACEHOLDER not in rendered
             assert FORBIDDEN_THREE_STRIKE not in rendered
             assert FORBIDDEN_TWENTY_STEP not in rendered
-            assert "fresh Reviewer → Main-Agent filter → Main-Agent fix" in rendered
-            assert "up to ten completed rounds" in rendered
+            assert "Main-Agent self-review loop" in rendered
+            assert "low=0" in rendered and "medium=1" in rendered and "high=2" in rendered
+            assert "Reviewer 3 is forbidden" in rendered
 
     def test_retained_tier3_fallback_template_uses_direct_orchestrator_implementation(self):
         rendered = Path("dev-pipeline/templates/bootstrap-tier3.md").read_text(encoding="utf-8")
@@ -1442,7 +1443,8 @@ class TestHeadlessPromptCleanupF033:
         refactor_prompt = _render_refactor_prompt(tmp_path / "refactor-mixed", platform="claude")
 
         assert "{{" + "REVIEWER_" + "SUBAGENT_PATH" + "}}" not in replacements
-        assert "reviewer-agent-prompt.md" in refactor_prompt
+        assert "Main-Agent self-review" in refactor_prompt
+        assert "Reviewer 3 is forbidden" in refactor_prompt
 
     def test_bugfix_refactor_explicit_codex_override(self, tmp_path, monkeypatch):
         from generate_bugfix_prompt import build_replacements as bugfix_build_replacements
@@ -1463,7 +1465,8 @@ class TestHeadlessPromptCleanupF033:
         refactor_prompt = _render_refactor_prompt(tmp_path / "refactor-codex", platform="claude", explicit_platform="codex")
 
         assert "{{" + "REVIEWER_" + "SUBAGENT_PATH" + "}}" not in replacements
-        assert "reviewer-agent-prompt.md" in refactor_prompt
+        assert "Main-Agent self-review" in refactor_prompt
+        assert "Reviewer 3 is forbidden" in refactor_prompt
 
 
 class TestFeatureBootstrapShellExtraction:
@@ -1503,7 +1506,7 @@ class TestFeatureBootstrapShellExtraction:
         assert "python" in prompt.lower()
         assert ".prizmkit/dev-pipeline/scripts/prizmkit-runtime-helper.py" in prompt
         assert "artifact touch .prizmkit/specs/123-prompt-cleanup/.prizmkit-test-started" in prompt
-        assert "text contains .prizmkit/specs/123-prompt-cleanup/review-report.md --pattern \"## Verdict\"" in prompt
+        assert "text final-verdict .prizmkit/specs/123-prompt-cleanup/review-report.md" in prompt
 
     def test_rendered_section_feature_prompts_use_runtime_helper_not_extracted_bash(self, tmp_path):
         for mode in ("lite", "standard", "full"):

package/bundled/dev-pipeline/tests/test_generate_bugfix_prompt.py

@@ -436,7 +436,7 @@ class TestBugfixBootstrapShellExtraction:
         assert "python" in prompt.lower()
         assert ".prizmkit/dev-pipeline/scripts/prizmkit-runtime-helper.py" in prompt
         assert "artifact ensure-dir .prizmkit/bugfix/B-001" in prompt
-        assert "text contains .prizmkit/bugfix/B-001/review-report.md --pattern \"## Verdict\"" in prompt
+        assert "text final-verdict .prizmkit/bugfix/B-001/review-report.md" in prompt
 
     def test_bugfix_template_source_is_clean_of_helper_covered_snippets(self):
         content = Path("dev-pipeline/templates/bugfix-bootstrap-prompt.md").read_text(encoding="utf-8")

package/bundled/dev-pipeline/tests/test_generate_refactor_prompt.py

@@ -296,7 +296,7 @@ class TestRefactorBootstrapShellExtraction:
         assert "python" in prompt.lower()
         assert ".prizmkit/dev-pipeline/scripts/prizmkit-runtime-helper.py" in prompt
         assert "artifact ensure-dir .prizmkit/refactor/R-001" in prompt
-        assert "text contains .prizmkit/refactor/R-001/review-report.md --pattern \"## Verdict\"" in prompt
+        assert "text final-verdict .prizmkit/refactor/R-001/review-report.md" in prompt
 
     def test_refactor_template_source_is_clean_of_helper_covered_snippets(self):
         content = Path("dev-pipeline/templates/refactor-bootstrap-prompt.md").read_text(encoding="utf-8")

package/bundled/dev-pipeline/tests/test_python_runner_parity.py

@@ -1173,6 +1173,47 @@ def test_feature_and_bugfix_explicit_worktree_launches_from_worktree_and_renders
     assert observed.prompt_calls[0]["execution_root"] == expected_worktree.resolve()
 
 
+def test_worktree_checkout_mode_and_path_survive_environment_changes(monkeypatch, tmp_path):
+    from prizmkit_runtime import runners
+    from prizmkit_runtime.runner_models import family_for, parse_invocation
+    from prizmkit_runtime.task_checkout import CHECKOUT_MODE_WORKTREE, TaskCheckout, persist_task_checkout
+
+    paths = _make_paths(tmp_path)
+    family = family_for("feature", paths)
+    invocation = parse_invocation(family, "run", ())
+    invocation = invocation.__class__(**{**invocation.__dict__, "list_path": paths.feature_plan})
+    branch = "dev/F-001-worktree-resume"
+    worktree_path = (paths.state_dir / "worktrees" / "dev-F-001-worktree-resume").resolve()
+    persist_task_checkout(
+        family.state_dir,
+        TaskCheckout.active(
+            "feature",
+            "F-001",
+            branch,
+            "main",
+            checkout_mode=CHECKOUT_MODE_WORKTREE,
+            worktree_path=str(worktree_path),
+        ),
+    )
+    observed = _install_runner_session_fakes(monkeypatch, runners)
+    setup_calls = []
+
+    def fake_setup(runtime):
+        setup_calls.append(runtime)
+        runtime.worktree_path.mkdir(parents=True, exist_ok=True)
+        return SimpleNamespace(ok=True, runtime=runtime, reason="")
+
+    monkeypatch.delenv("USE_WORKTREE", raising=False)
+    monkeypatch.delenv("DEV_BRANCH", raising=False)
+    monkeypatch.setattr(runners, "ensure_linked_worktree", fake_setup)
+
+    status = runners._process_item(family, invocation, "F-001", paths, initial_metadata={})
+
+    assert status == "completed"
+    assert setup_calls[0].worktree_path == worktree_path
+    assert observed.launch_cwds == [worktree_path]
+
+
 def test_successful_worktree_session_merges_cleans_branch_and_honors_auto_push(monkeypatch, tmp_path):
     from prizmkit_runtime import runners
     from prizmkit_runtime.runner_models import family_for, parse_invocation
@@ -1192,6 +1233,9 @@ def test_successful_worktree_session_merges_cleans_branch_and_honors_auto_push(m
     assert status == "completed"
     assert observed.merge_calls and observed.merge_calls[0][1] is True
     assert observed.cleanup_calls and observed.cleanup_calls[0][1] is True
+    checkout = json.loads((paths.feature_state_dir / "F-001" / "checkout.json").read_text(encoding="utf-8"))
+    assert checkout["state"] == "completed"
+    assert checkout["active_dev_branch"] == "dev/F-001-test"
 
 
 def test_worktree_merge_conflict_writes_session_status_and_updates_merge_conflict(monkeypatch, tmp_path):
@@ -1212,7 +1256,10 @@ def test_worktree_merge_conflict_writes_session_status_and_updates_merge_conflic
 
     assert status == "completed"
     assert observed.updates[-1][0] == "merge_conflict"
-    assert observed.cleanup_calls[-1][1] is False
+    assert observed.cleanup_calls == []
+    checkout = json.loads((paths.feature_state_dir / "F-001" / "checkout.json").read_text(encoding="utf-8"))
+    assert checkout["state"] == "active"
+    assert checkout["checkout_mode"] == "worktree"
     session_status = json.loads((paths.feature_state_dir / "F-001" / "sessions" / "F-001-session-1" / "session-status.json").read_text(encoding="utf-8"))
     assert session_status["status"] == "merge_conflict"
     assert session_status["task_id"] == "F-001"
@@ -1222,7 +1269,7 @@ def test_worktree_merge_conflict_writes_session_status_and_updates_merge_conflic
     assert session_status["worktree_path"].endswith("dev-F-001-test")
 
 
-def test_terminal_worktree_failure_saves_wip_and_removes_worktree_without_deleting_branch(monkeypatch, tmp_path):
+def test_terminal_worktree_failure_saves_wip_and_preserves_checkout(monkeypatch, tmp_path):
     from prizmkit_runtime import runners
     from prizmkit_runtime.runner_models import family_for, parse_invocation
 
@@ -1239,7 +1286,10 @@ def test_terminal_worktree_failure_saves_wip_and_removes_worktree_without_deleti
 
     assert status == "failed"
     assert observed.wip_calls
-    assert observed.cleanup_calls[-1][1] is False
+    assert observed.cleanup_calls == []
+    checkout = json.loads((paths.feature_state_dir / "F-001" / "checkout.json").read_text(encoding="utf-8"))
+    assert checkout["state"] == "active"
+    assert checkout["checkout_mode"] == "worktree"
     assert observed.updates[-1][0] == "crashed"
 
 
@@ -1708,9 +1758,54 @@ def test_foreground_runner_emits_structured_feature_progress(monkeypatch, tmp_pa
     assert "Session result: success" in captured.err
 
 
+def test_durable_checkout_reservation_is_first_writer_wins_and_conflicts_fail(tmp_path):
+    from prizmkit_runtime.task_checkout import TaskCheckout, TaskCheckoutError, reserve_task_checkout
+
+    state_dir = tmp_path / "state"
+    first = TaskCheckout.active("feature", "F-001", "dev/F-001-first", "main")
+    conflicting = TaskCheckout.active("feature", "F-001", "dev/F-001-second", "main")
+
+    assert reserve_task_checkout(state_dir, first) == first
+    assert reserve_task_checkout(state_dir, first) == first
+    with pytest.raises(TaskCheckoutError, match="already owns checkout"):
+        reserve_task_checkout(state_dir, conflicting)
+
+
+def test_task_checkout_guard_rejects_concurrent_start_and_reset(tmp_path):
+    from prizmkit_runtime.task_checkout import TaskCheckoutError, task_checkout_guard
+
+    state_dir = tmp_path / "state"
+    with task_checkout_guard(state_dir, "F-001"):
+        with pytest.raises(TaskCheckoutError, match="already running or being reset"):
+            with task_checkout_guard(state_dir, "F-001"):
+                pass
+
+
+def test_completed_checkout_blocks_rerun_until_explicit_reset(monkeypatch, tmp_path):
+    from prizmkit_runtime import runners
+    from prizmkit_runtime.runner_models import family_for, parse_invocation
+    from prizmkit_runtime.task_checkout import TaskCheckout, persist_task_checkout
+
+    paths = _make_paths(tmp_path)
+    family = family_for("feature", paths)
+    invocation = parse_invocation(family, "run", ())
+    invocation = invocation.__class__(**{**invocation.__dict__, "list_path": paths.feature_plan})
+    completed = TaskCheckout.active("feature", "F-001", "dev/F-001-completed", "main").completed()
+    persist_task_checkout(family.state_dir, completed)
+    monkeypatch.setattr(
+        runners,
+        "load_runtime_config",
+        lambda paths: SimpleNamespace(ai_client=SimpleNamespace(command="fake-ai", platform="codebuddy"), model="", effort=None),
+    )
+
+    with pytest.raises(RuntimeError, match="completed; run explicit reset or clean"):
+        runners._process_item(family, invocation, "F-001", paths, initial_metadata={})
+
+
 def test_no_worktree_interrupt_saves_wip_and_returns_to_base(monkeypatch, tmp_path):
     from prizmkit_runtime import runners
     from prizmkit_runtime.runner_models import family_for, parse_invocation
+    from prizmkit_runtime.task_checkout import load_task_checkout
 
     monkeypatch.setenv("DEV_BRANCH", "dev/F-001-interrupt")
     paths = _make_paths(tmp_path)
@@ -1734,9 +1829,50 @@ def test_no_worktree_interrupt_saves_wip_and_returns_to_base(monkeypatch, tmp_pa
     with pytest.raises(KeyboardInterrupt):
         runners._process_item(family, invocation, "F-001", paths, initial_metadata={})
 
+    checkout = load_task_checkout(family.state_dir, family.task_type, "F-001")
+    assert checkout is not None
+    assert checkout.is_active
+    assert checkout.active_dev_branch == "dev/F-001-interrupt"
+    assert checkout.base_branch == "main"
     assert ensure_calls == [("main", "dev/F-001-interrupt")]
 
 
+@pytest.mark.parametrize(
+    ("kind", "item_id", "branch"),
+    [
+        ("feature", "F-001", "dev/F-001-interrupt"),
+        ("bugfix", "B-001", "bugfix/B-001-interrupt"),
+        ("refactor", "R-001", "refactor/R-001-interrupt"),
+    ],
+)
+def test_interrupted_task_reuses_durable_checkout_on_next_process(monkeypatch, tmp_path, kind, item_id, branch):
+    from prizmkit_runtime import runners
+    from prizmkit_runtime.runner_models import family_for, parse_invocation
+    from prizmkit_runtime.task_checkout import TaskCheckout, persist_task_checkout
+
+    paths = _make_paths(tmp_path / kind)
+    family = family_for(kind, paths)
+    plan_path = {"feature": paths.feature_plan, "bugfix": paths.bugfix_plan, "refactor": paths.refactor_plan}[kind]
+    invocation = parse_invocation(family, "run", ())
+    invocation = invocation.__class__(**{**invocation.__dict__, "list_path": plan_path})
+    persist_task_checkout(family.state_dir, TaskCheckout.active(kind, item_id, branch, "main"))
+    observed = _install_runner_session_fakes(monkeypatch, runners)
+    branch_contexts = []
+
+    def fake_git_plan(project_root, plan):
+        branch_contexts.append((plan.commands[1].args[-2], plan.commands[1].args[-1]))
+        return SimpleNamespace(ok=True, plan=plan, results=())
+
+    monkeypatch.delenv("DEV_BRANCH", raising=False)
+    monkeypatch.setattr(runners, "run_git_plan", fake_git_plan)
+
+    status = runners._process_item(family, invocation, item_id, paths, initial_metadata={})
+
+    assert status == "completed"
+    assert branch_contexts[0] == (branch, "main")
+    assert observed.launch_cwds == [paths.project_root]
+
+
 def _init_bookkeeping_repo(project_root, *, configure_identity=True):
     subprocess.run(["git", "init", "-b", "main"], cwd=project_root, check=True, stdout=subprocess.DEVNULL)
     if configure_identity:
@@ -2641,6 +2777,8 @@ def _write_family_case(paths, kind, item_id=""):
 
 def test_python_reset_warns_before_mutation_and_matches_snapshot_cleanup(tmp_path):
     from prizmkit_runtime.reset import run_reset_command
+    from prizmkit_runtime.runner_models import family_for
+    from prizmkit_runtime.task_checkout import load_task_checkout
 
     snapshot = json.loads((REPO_ROOT / "tests" / "fixtures" / "dev-pipeline-behavior-snapshots" / "unix-runtime.json").read_text(encoding="utf-8"))
     for kind, item_id in [("feature", "F-001"), ("bugfix", "B-001"), ("refactor", "R-001")]:
@@ -2659,6 +2797,71 @@ def test_python_reset_warns_before_mutation_and_matches_snapshot_cleanup(tmp_pat
         for key, value in expected_runtime.items():
             assert runtime.get(key) == value
         assert "continuation_pending" not in runtime
+        family = family_for(kind, paths)
+        checkout = load_task_checkout(family.state_dir, family.task_type, item_id)
+        assert checkout is not None
+        assert checkout.is_reset
+        assert checkout.active_dev_branch == ""
+        assert checkout.base_branch == ""
+
+
+def test_python_reset_cleans_authoritative_custom_branch_before_reset_marker(monkeypatch, tmp_path):
+    from prizmkit_runtime import reset as reset_runtime
+    from prizmkit_runtime.reset import run_reset_command
+    from prizmkit_runtime.runner_models import family_for
+    from prizmkit_runtime.task_checkout import TaskCheckout, load_task_checkout, persist_task_checkout
+
+    paths = _make_paths(tmp_path)
+    _write_family_case(paths, "feature", "F-001")
+    family = family_for("feature", paths)
+    custom_branch = "custom/feature-checkout"
+    persist_task_checkout(family.state_dir, TaskCheckout.active("feature", "F-001", custom_branch, "main"))
+    cleanup_calls = []
+
+    monkeypatch.setattr(reset_runtime, "list_reset_branches", lambda *args: ())
+    monkeypatch.setattr(
+        reset_runtime,
+        "reset_cleanup_branch",
+        lambda project_root, branch, *, return_branch, delete: cleanup_calls.append((branch, delete))
+        or SimpleNamespace(branch=branch, cleaned=True, deleted=False, returned_to=return_branch, warnings=()),
+    )
+
+    result = run_reset_command("feature", ("F-001",), paths)
+
+    assert result.exit_code == 0
+    assert cleanup_calls == [(custom_branch, False)]
+    checkout = load_task_checkout(family.state_dir, family.task_type, "F-001")
+    assert checkout is not None and checkout.is_reset
+
+
+def test_python_reset_failure_keeps_authoritative_checkout_active(monkeypatch, tmp_path):
+    from prizmkit_runtime import reset as reset_runtime
+    from prizmkit_runtime.reset import run_reset_command
+    from prizmkit_runtime.runner_models import family_for
+    from prizmkit_runtime.task_checkout import TaskCheckout, load_task_checkout, persist_task_checkout
+
+    paths = _make_paths(tmp_path)
+    _write_family_case(paths, "feature", "F-001")
+    family = family_for("feature", paths)
+    checkout = TaskCheckout.active("feature", "F-001", "custom/feature-checkout", "main")
+    persist_task_checkout(family.state_dir, checkout)
+    monkeypatch.setattr(reset_runtime, "list_reset_branches", lambda *args: ())
+    monkeypatch.setattr(
+        reset_runtime,
+        "reset_cleanup_branch",
+        lambda *args, **kwargs: SimpleNamespace(
+            branch="custom/feature-checkout",
+            cleaned=False,
+            deleted=False,
+            returned_to="main",
+            warnings=("could not checkout custom/feature-checkout",),
+        ),
+    )
+
+    result = run_reset_command("feature", ("F-001",), paths)
+
+    assert result.exit_code == 1
+    assert load_task_checkout(family.state_dir, family.task_type, "F-001") == checkout
 
 
 def test_python_clean_reset_deletes_continuation_artifacts_for_all_families(tmp_path):

package/bundled/dev-pipeline/tests/test_runtime_helper.py

@@ -75,6 +75,42 @@ class TestArtifactAndTextHelpers:
         missing = run_helper("text", "contains", str(target), "--pattern", "gamma")
         assert first_line(missing) == "GATE:MISSING"
 
+    def test_final_verdict_requires_last_final_result_section(self, tmp_path):
+        report = tmp_path / "review-report.md"
+        report.write_text(
+            "# Review Report\n\n## Status: IN_PROGRESS\n\n"
+            "## Main Review Round 1\n\n- Note: incidental PASS text\n",
+            encoding="utf-8",
+        )
+        incomplete = run_helper("text", "final-verdict", str(report))
+        assert incomplete.returncode != 0
+        assert first_line(incomplete) == "REVIEW_INCOMPLETE"
+
+        report.write_text(
+            report.read_text(encoding="utf-8")
+            + "\n## Final Result\n\n- Verdict: PASS\n- Summary: done\n",
+            encoding="utf-8",
+        )
+        complete = run_helper("text", "final-verdict", str(report))
+        assert complete.returncode == 0
+        assert first_line(complete) == "REVIEW_PASS"
+        assert "verdict: PASS" in complete.stdout
+
+    def test_final_verdict_uses_last_final_result_and_rejects_unknown_value(self, tmp_path):
+        report = tmp_path / "review-report.md"
+        report.write_text(
+            "## Final Result\n- Verdict: PASS\n\n"
+            "## Final Result\n- Verdict: BLOCKED\n",
+            encoding="utf-8",
+        )
+        blocked = run_helper("text", "final-verdict", str(report))
+        assert first_line(blocked) == "REVIEW_BLOCKED"
+
+        report.write_text("## Final Result\n- Verdict: UNKNOWN\n", encoding="utf-8")
+        invalid = run_helper("text", "final-verdict", str(report))
+        assert invalid.returncode != 0
+        assert first_line(invalid) == "REVIEW_INVALID"
+
 
 class TestPlatformExecutableAndPortHelpers:
     def test_platform_detection_and_skill_lookup_follow_config(self, tmp_path, monkeypatch):