feat: Product v3.6 completion — schema extraction, repository pattern, obligation test coverage

.claude/agents/executor.md

@@ -113,6 +113,23 @@ Record the results. This confirms the clean slate. If anything fails here
 Follow the **mol-execute** skill (`.claude/skills/mol-execute/SKILL.md`).
 Read that file for the full protocol: cook the molecule, walk atoms, close.
 
+### Formula Selection
+
+If your spawn prompt specifies a formula, use that. Otherwise, auto-select:
+
+```bash
+bd show <task-id>   # Check type and description
+```
+
+| Condition | Formula | Cook command |
+|-----------|---------|-------------|
+| Type is `bug` | `bug-triage.yaml` | `--formula .claude/formulas/bug-triage.yaml --var "BUG_IDS=<ids>"` |
+| Task is well-defined (test rewrite, allowlist fix, mechanical change) | `task-single.yaml` | `--formula .claude/formulas/task-single.yaml --var "TASK_IDS=<ids>"` |
+| Task needs research or TDD (new feature, refactor with unknowns) | `task-execute.yaml` | `--formula .claude/formulas/task-execute.yaml --var "TASK_IDS=<ids>"` |
+
+**Default to `task-single`** unless the task explicitly needs research or
+architect review. Most beads tasks have sufficient context in their description.
+
 ## Quality Gates (HARD REQUIREMENTS)
 
 You have a Postgres container. You MUST use it. The quality gates below are
@@ -132,9 +149,24 @@ This starts Docker, runs all 5 suites (unit, integration, integration_v2,
 e2e, ui) via tox, and tears down. Do NOT substitute with individual pytest
 commands — `./run_all_tests.sh` is the single source of truth.
 
+JSON results are saved to `test-results/<ddmmyy_HHmm>/`. Use those to
+review results — background processes may crash and lose terminal output.
+
 **If any test fails, your implementation has a bug. Fix it before
 committing.** You started from a clean slate — every failure is yours.
 
+### Test Integrity Policy — ZERO TOLERANCE
+
+**Read and follow CLAUDE.md "Test Integrity Policy" section. Summary:**
+
+- **NEVER** use `--ignore`, `-k "not ..."`, `--deselect`, `pytest.mark.skip`,
+  or `pytest.mark.xfail` to work around failures.
+- **NEVER** rationalize failures as "pre-existing", "infrastructure issue",
+  "misplaced test", "needs a running server", or "was deselected in full run".
+- If a test needs Docker → start Docker (`./run_all_tests.sh` handles this).
+- If infrastructure is broken → STOP and report. Do NOT skip tests and report success.
+- A failing test is a failing test. Fix it or report it as a blocker.
+
 ### What tests to write
 
 **For repository migrations, data access changes, and session management:**

.claude/commands/team.md

@@ -39,23 +39,44 @@ TeamCreate: team_name="<descriptive-name>", description="<what the team does>"
 Analyze the user's prompt and break it into parallel work items. For beads
 task IDs, run `bd show <id>` to read descriptions and verify they're unblocked.
 
+**Grouping strategy:**
+- Group tasks that touch the **same file** into one executor (avoids merge conflicts)
+- Each executor gets one file or a small set of non-overlapping files
+- The allowlist (`obligation_test_quality_allowlist.json`) is a shared resource —
+  coordinate updates or do a reconciliation pass after all executors finish
+
+**Formula selection:**
+- If the user specifies a formula, use it
+- Otherwise, check each task with `bd show <id>`:
+  - Well-defined tasks (test rewrites, allowlist fixes, mechanical changes) → `task-single.yaml`
+  - Tasks needing research or TDD (new features, refactors) → `task-execute.yaml`
+  - Bugs → `bug-triage.yaml`
+
 Each work item becomes:
 - A task in the team's task list (via TaskCreate)
 - An executor teammate spawned to handle it (via Task with team_name)
 
 ### Step 4: Spawn executor teammates
-For each work item, spawn an executor:
+For each work item, spawn an executor with an **explicit formula and cook command**:
 ```
 Task:
   team_name: "<team-name>"
-  name: "executor-<short-id>"
+  name: "executor-<short-label>"
   subagent_type: "executor"
   prompt: |
-    Execute beads task salesagent-<id>.
+    Execute these beads tasks using <formula> formula:
+    salesagent-<id1> salesagent-<id2> salesagent-<id3>
+
+    Cook:
+    python3 .claude/scripts/cook_formula.py \
+      --formula .claude/formulas/<formula>.yaml \
+      --var "TASK_IDS=salesagent-<id1> salesagent-<id2> salesagent-<id3>" \
+      --epic-title "<descriptive title>"
+
+    Then walk atoms: bd ready → bd show <id> → execute → bd close <id> → repeat
 
-    Run `bd show <id>` for full description and acceptance criteria.
-    Follow the executor protocol: setup DB, cook molecule, walk atoms,
-    quality gates, commit.
+    Your files: <list of files this executor owns>
+    Shared resource: <any shared files like allowlists>
 ```
 
 **NOTE: `isolation: "worktree"` is a no-op for team agents.** All executors
@@ -70,9 +91,15 @@ non-overlapping files to avoid conflicts.
 
 ### Step 6: Verify and commit
 After all executors complete:
-1. Run `make quality` on the combined result
-2. Squash or organize commits if needed
-3. Push if the user requests it
+1. Run `./run_all_tests.sh` on the combined result (NOT just `make quality` —
+   the full suite including e2e and ui is mandatory)
+2. Review JSON results in `test-results/<ddmmyy_HHmm>/` if terminal output is lost
+3. Squash or organize commits if needed
+4. Push if the user requests it
+
+**Test Integrity — ZERO TOLERANCE**: If any test fails in the combined result,
+do NOT skip it or rationalize it. See CLAUDE.md "Test Integrity Policy".
+Every failure must be fixed or reported as a blocker.
 
 ## User's Request
 

.claude/formulas/bug-triage.yaml

@@ -605,6 +605,10 @@ finalize:
          e2e, ui) via tox, and tears down. Do NOT substitute with individual pytest
          commands — `./run_all_tests.sh` is the single source of truth.
 
+         JSON results are saved to `test-results/<ddmmyy_HHmm>/`. If terminal
+         output is lost (background process crash, timeout), read the JSON files
+         to review results.
+
       3. **Compare counts**:
          - Passes must be >= baseline (new tests are expected)
          - Failures must be exactly 0
@@ -616,15 +620,24 @@ finalize:
          - make quality: [N] passed, 0 failed (baseline: [N])
          - integration: [N] passed, 0 failed (baseline: [N])
          - integration_v2: [N] passed, 0 failed (baseline: [N])
+         - e2e: [N] passed, [N] skipped, 0 failed
+         - ui: [N] passed, [N] skipped, 0 failed
          - Delta: +[N] new tests, 0 regressions"
          ```
 
-      ## HARD STOP
+      ## HARD STOP — Test Integrity (ZERO TOLERANCE)
+
+      **Read CLAUDE.md "Test Integrity Policy". These rules apply here:**
 
-      If ANY gate has failures that were not in the baseline:
+      If ANY test fails in `./run_all_tests.sh`:
       - Do NOT close this atom.
       - Do NOT commit.
-      - Fix the regressions first, then re-run this atom.
+      - Do NOT use `--ignore`, `-k "not ..."`, or `--deselect` to skip failures.
+      - Do NOT rationalize failures as "pre-existing", "infrastructure issue",
+        "misplaced test", "needs a running server", or "was deselected".
+      - Fix the regressions first, then re-run `./run_all_tests.sh`.
+      - If infrastructure is broken and you cannot fix it, report it as a
+        BLOCKER to the user. Do NOT skip tests and report success.
 
     acceptance: "All three gates pass. Zero regressions vs baseline. Final counts recorded in epic notes."
     depends_on_all_barriers: true

.claude/formulas/task-execute.yaml

@@ -604,6 +604,10 @@ finalize:
          e2e, ui) via tox, and tears down. Do NOT substitute with individual pytest
          commands — `./run_all_tests.sh` is the single source of truth.
 
+         JSON results are saved to `test-results/<ddmmyy_HHmm>/`. If terminal
+         output is lost (background process crash, timeout), read the JSON files
+         to review results.
+
       3. **Compare counts**:
          - Passes must be >= baseline (new tests are expected)
          - Failures must be exactly 0
@@ -615,15 +619,24 @@ finalize:
          - make quality: [N] passed, 0 failed (baseline: [N])
          - integration: [N] passed, 0 failed (baseline: [N])
          - integration_v2: [N] passed, 0 failed (baseline: [N])
+         - e2e: [N] passed, [N] skipped, 0 failed
+         - ui: [N] passed, [N] skipped, 0 failed
          - Delta: +[N] new tests, 0 regressions"
          ```
 
-      ## HARD STOP
+      ## HARD STOP — Test Integrity (ZERO TOLERANCE)
+
+      **Read CLAUDE.md "Test Integrity Policy". These rules apply here:**
 
-      If ANY gate has failures that were not in the baseline:
+      If ANY test fails in `./run_all_tests.sh`:
       - Do NOT close this atom.
       - Do NOT commit.
-      - Fix the regressions first, then re-run this atom.
+      - Do NOT use `--ignore`, `-k "not ..."`, or `--deselect` to skip failures.
+      - Do NOT rationalize failures as "pre-existing", "infrastructure issue",
+        "misplaced test", "needs a running server", or "was deselected".
+      - Fix the regressions first, then re-run `./run_all_tests.sh`.
+      - If infrastructure is broken and you cannot fix it, report it as a
+        BLOCKER to the user. Do NOT skip tests and report success.
 
     acceptance: "All three gates pass. Zero regressions vs baseline. Final counts recorded in epic notes."
     depends_on_all_barriers: true