jessekemp1 · jessekemp1 · Sep 7, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -10,7 +10,7 @@ name: ci
 #     ruff is a pre-existing 347-error backlog — see step comment).
 #   - macos: proves the install path on macOS-14. Cannot be validated from the
 #     Linux dev box; needs a real macOS runner (see NOTE in the job).
-#   - wheel-canary: NON-BLOCKING honest red marker — builds + installs the wheel
+#   - wheel-canary: blocking install check — builds + installs the wheel
 #     and imports the console entrypoints.
 #   - integration: runs `pytest -m integration` so integration-marked tests
 #     can't silently rot. Most skip gracefully when their external dependency
@@ -46,13 +46,15 @@ jobs:
           # Isolate all state under a throwaway dir so tests never touch a real
           # ~/.cortex or the checkout's own state.
           CORTEX_ROOT_DIR: ${{ runner.temp }}/cortex-home
+          CORTEX_STATE_DIR: ${{ runner.temp }}/cortex-state
+          CORTEX_HOME: ${{ runner.temp }}/cortex-state
         run: |
           mkdir -p "$CORTEX_ROOT_DIR"
           # synthetic/tests/ depends on [analytics] (xgboost, shap) which we
           # don't install in smoke — excluded here (matches --collect-only
           # policy). This runs the FULL test BODIES, not just collection.
           set +e
-          python -m pytest -p no:cacheprovider --ignore=synthetic/tests -q \
+          python -m pytest -p tests.network_guard -p no:cacheprovider --ignore=synthetic/tests -q --junitxml=/tmp/pytest.xml \
             | tee /tmp/pytest.out
           code=${PIPESTATUS[0]}
           set -e
@@ -62,25 +64,34 @@ jobs:
           #    branch. A floor of 1960 catches accidental mass-skips or a
           #    collection regression that silently drops tests while still
           #    "passing". Bump the floor deliberately as the suite grows.
-          passed="$(grep -oE '[0-9]+ passed' /tmp/pytest.out | tail -1 | grep -oE '[0-9]+')"
-          echo "pytest reported ${passed} passed (floor: 1960)"
-          test -n "$passed"
-          test "$passed" -ge 1960
+          python - <<'PYTEST_COUNTS'
+          import xml.etree.ElementTree as ET
+          root = ET.parse("/tmp/pytest.xml").getroot()
+          suites = [root] if root.tag == "testsuite" else list(root.findall("testsuite"))
+          passed = sum(int(s.get("tests", 0)) - int(s.get("failures", 0))
+                       - int(s.get("errors", 0)) - int(s.get("skipped", 0)) for s in suites)
+          print(f"JUnit reported {passed} passed (floor: 1960)")
+          assert passed >= 1960, f"Unexpected test count: {passed}"
+          PYTEST_COUNTS
 
       - name: cortex demo must print the FK trail
         env:
           CORTEX_ROOT_DIR: ${{ runner.temp }}/cortex-home
+          CORTEX_STATE_DIR: ${{ runner.temp }}/cortex-state
+          CORTEX_HOME: ${{ runner.temp }}/cortex-state
         run: |
           output="$(cortex demo)"
           echo "$output"
           # Three synthesized prompts must each appear as a linked entry
           # with score 0.80. If this assertion ever fails, the headline
-          # 'compounding intelligence' claim is broken on this build.
+          # synthetic linkage fixture is broken; this does not prove useful learning.
           echo "$output" | grep -c 'score 0.80' | grep -q '^3$'
 
       - name: cortex briefing must exit 2 without ANTHROPIC_API_KEY
         env:
           CORTEX_ROOT_DIR: ${{ runner.temp }}/cortex-home
+          CORTEX_STATE_DIR: ${{ runner.temp }}/cortex-state
+          CORTEX_HOME: ${{ runner.temp }}/cortex-state
         run: |
           # Pre-flight contract: missing key → exit 2 with actionable msg.
           # Anything else (silent hang, exit 0, exit 1) regresses Phase 1.
@@ -94,6 +105,8 @@ jobs:
       - name: MCP smoke drill — bridge DOWN (crash-proof headline)
         env:
           CORTEX_ROOT_DIR: ${{ runner.temp }}/cortex-home
+          CORTEX_STATE_DIR: ${{ runner.temp }}/cortex-state
+          CORTEX_HOME: ${{ runner.temp }}/cortex-state
         run: |
           # No bridge is started. The core-8 in-process MCP tools must still be
           # green (crash-proof contract). Passthrough tools that need the bridge
@@ -133,11 +146,15 @@ jobs:
 
       - name: ruff check (release-touched paths)
         run: |
-          # Scoped to the Python files this release (Workstream F) touched.
-          # Both are ruff-clean; widen this list as backlog dirs reach zero.
+          # Core release paths plus the execution-integrity regression surface.
+          # Widen this list as other backlog directories reach zero.
           python -m ruff check \
             bridge_intelligence.py \
-            tests/contract/test_phase5_invariants.py
+            tests/contract/test_phase5_invariants.py \
+            supervisor/dispatch.py supervisor/collector.py supervisor/router.py \
+            store_registry.py tests/test_dispatch_integrity.py \
+            tests/test_orchestration_pipeline.py tests/test_model_router.py \
+            tests/test_doctor_json_reset.py tests/test_mcp_research_tools.py tests/network_guard.py
 
   macos:
     # macOS-14 install path. NOTE: this job CANNOT be validated from the Linux
@@ -166,6 +183,8 @@ jobs:
       - name: cortex doctor must run without crashing (non-zero exit allowed)
         env:
           CORTEX_ROOT_DIR: ${{ runner.temp }}/cortex-home
+          CORTEX_STATE_DIR: ${{ runner.temp }}/cortex-state
+          CORTEX_HOME: ${{ runner.temp }}/cortex-state
         run: |
           mkdir -p "$CORTEX_ROOT_DIR"
           # doctor reports env facts (API key, launchd, bridge) — a red env is
@@ -184,6 +203,8 @@ jobs:
       - name: cortex demo must print the FK trail
         env:
           CORTEX_ROOT_DIR: ${{ runner.temp }}/cortex-home
+          CORTEX_STATE_DIR: ${{ runner.temp }}/cortex-state
+          CORTEX_HOME: ${{ runner.temp }}/cortex-state
         run: |
           output="$(cortex demo)"
           echo "$output"
@@ -192,17 +213,14 @@ jobs:
       - name: pytest (full suite, hermetic) — Python 3.12
         env:
           CORTEX_ROOT_DIR: ${{ runner.temp }}/cortex-home
+          CORTEX_STATE_DIR: ${{ runner.temp }}/cortex-state
+          CORTEX_HOME: ${{ runner.temp }}/cortex-state
         run: |
-          python -m pytest -p no:cacheprovider --ignore=synthetic/tests -q
+          python -m pytest -p tests.network_guard -p no:cacheprovider --ignore=synthetic/tests -q
 
   wheel-canary:
-    # NON-BLOCKING honest red marker for the git-clone-only beta. Builds the
-    # wheel, installs it into a clean venv, and imports the console entrypoints.
-    # Kept continue-on-error per the release plan even though the packaging was
-    # fixed on this branch and it is expected to pass locally — a green here is
-    # a bonus, a red here is a signal, neither blocks the release gate.
+    # The deployed artifact must contain the diagnostics as well as entrypoints.
     runs-on: ubuntu-latest
-    continue-on-error: true
     steps:
       - uses: actions/checkout@v4
 
@@ -226,7 +244,7 @@ jobs:
           # The two console-script entrypoints must import from the installed
           # wheel (not the checkout) — run from /tmp so cwd isn't the repo root.
           cd /tmp
-          /tmp/wheelvenv/bin/python -c "import cli, mcp_server; print('wheel imports OK')"
+          /tmp/wheelvenv/bin/python -c "import cli, mcp_server, store_registry, recall_confidence; print('wheel imports OK')"
           /tmp/wheelvenv/bin/cortex --help
 
   integration:

diff --git a/docs/design/truthful-execution.md b/docs/design/truthful-execution.md
@@ -0,0 +1,76 @@
+# Execution receipts and unverified outcomes
+
+This change is a containment step, not a completed autonomous learning system.
+
+## Contract
+
+- `DispatchResult.success` and summary `succeeded` describe execution only.
+- `execution_status` distinguishes completed, failed, incomplete, timed_out,
+  unsupported, and dry_run. Dry runs are neither successful nor failed execution.
+- Truncation, refusal, missing final text, and tool-turn exhaustion cannot
+  produce successful Anthropic execution. Partial output and known token usage
+  survive incomplete responses. A batch API `succeeded` envelope is insufficient.
+- Direct and message-batch adapters reject explicit shell commands; they do not
+  have a command-execution capability. Free-text capability inference is not solved.
+- Persisted results and routing receipts say `verification_status=unverified`
+  and `learning_eligible=false`. A successful response without independent
+  verification is counted as unverified and flagged for review.
+- `VerificationEvidence` is a legacy heuristic assessment. Even its maximum
+  score cannot verify a task or automatically requeue it. It is serialized under
+  `assessment`, not `verification`.
+- Each dispatch has an attempt ID; batch IDs plus custom IDs provide stable
+  attempt identities on retrieval. Assessments attach to attempts. Ambiguous
+  work-item-only attachments fail rather than attach to the wrong retry.
+- Repeated collection of the same receipt in one collector is idempotent;
+  conflicting receipts fail. Run directories cannot overwrite another run in
+  the same second. Cross-process journal deduplication is still pending.
+
+## Routing behavior change
+
+ModelRouter retains existing journals and appends execution observations, but
+does not ingest them into adaptive outcome statistics or update those statistics
+from new execution-only observations. Static task/complexity routing continues.
+Batch completion no longer fabricates quality 1.0. Existing quality heuristics
+remain advisory telemetry under `heuristic_quality_score`; measured `quality_score`
+remains null, not a learning signal.
+
+There is deliberately no switch to re-enable unverified history. Adaptive
+outcome routing needs a separate ingestion contract that binds task, attempt,
+artifact revision, check version, coverage and result, and supports corrections.
+This pause concerns this supervisor router; it does not claim to disable every
+other learning or retrieval heuristic in Cortex.
+
+## Compatibility and limits
+
+Consumers reading `verification` must migrate to the explicitly advisory
+`assessment` field. Summary success counts are retained for compatibility but
+must be labeled execution counts in downstream interfaces. Evidence attachment
+before collection, or ambiguous attachment after retry, now raises ValueError.
+
+Local providers currently return only text/token tuples, so their truncation
+status cannot be established here. Nonempty local output remains unverified.
+Timeout usage may be unknown, and cancelling an asyncio thread wrapper does not
+prove an underlying network request stopped. No exactly-once external execution
+guarantee is made.
+
+Path validation resolves components instead of comparing string prefixes. It
+rejects sibling-prefix and symlink escapes under a stable filesystem. This is
+not an OS sandbox against a hostile concurrent filesystem mutation.
+
+## Release and next gates
+
+The focused regression suite is `tests/test_dispatch_integrity.py`. It exercises
+real dispatcher/collector/router code with a synthetic provider and actual
+temporary filesystem persistence. No live-model or installed-machine claim is
+made by that suite. CI also gates touched files and installed-wheel imports.
+
+Next: transactional event identity and correction, unified memory reads/writes,
+atomic supersession, capture privacy, dependency/lease/cancellation enforcement,
+and an independent exact-artifact verifier. Then prove a real interrupted work
+loop and useful lesson reuse. Machine cutover requires a verified source/state
+inventory, a consistent backup, and an explicit rollback path.
+
+The offline pytest plugin also checks intended HTTP destinations before proxy
+transport, permits in-memory test transports, and blocks external Python sockets.
+Use a credential-free environment for broad test runs. This is not an OS-level
+sandbox and does not constrain arbitrary subprocess network access.
diff --git a/pyproject.toml b/pyproject.toml
@@ -92,11 +92,12 @@ Repository = "https://github.com/jessekemp1/cortex"
 # wheel (mcp_handlers, then state_paths, then config… ModuleNotFoundError one
 # import deeper each time), so we package every root module. All 42 are
 # import-safe (no top-level side effects), so unused ones are inert.
-py-modules = ["agent_factory", "ai_intelligence", "alert_monitor", "bridge", "bridge_intelligence", "bridge_system", "briefing_resilient", "config", "context_intelligence", "data_migration", "deep_assessment", "env_loader", "feedback", "formatter", "goal_commands", "goal_parser", "goal_velocity", "heartbeat", "learning", "learning_config", "mcp_handlers", "mcp_server", "metrics_tracker", "orchestrator", "portfolio_analyzer", "portfolio_memory", "recommendation_engine", "recommendations", "reflection", "scheduler", "security", "self_audit", "session_cache", "session_delta", "session_manager", "session_snapshot", "session_watcher", "setup_ops", "spec_knowledge_base", "state_paths", "task_discovery", "validation", "watches", "weekly_planner"]
+py-modules = ["agent_factory", "ai_intelligence", "alert_monitor", "bridge", "bridge_intelligence", "bridge_system", "briefing_resilient", "config", "context_intelligence", "data_migration", "deep_assessment", "env_loader", "feedback", "formatter", "goal_commands", "goal_parser", "goal_velocity", "heartbeat", "learning", "learning_config", "mcp_handlers", "mcp_server", "metrics_tracker", "orchestrator", "portfolio_analyzer", "portfolio_memory", "recommendation_engine", "recommendations", "reflection", "scheduler", "security", "self_audit", "session_cache", "session_delta", "session_manager", "session_snapshot", "session_watcher", "setup_ops", "spec_knowledge_base", "state_paths", "store_registry", "recall_confidence", "task_discovery", "validation", "watches", "weekly_planner"]
 
 [tool.setuptools.packages.find]
 exclude = [
     "tests*",
+    "build*",
     "docs*",
     "examples*",
     "reports*",

diff --git a/store_registry.py b/store_registry.py
@@ -28,11 +28,13 @@
 from pathlib import Path
 from typing import List, Optional
 
+from state_paths import get_cortex_dir
+
 WRITER_NONE = "NONE"
 
 
 def _cortex_dir() -> Path:
-    return Path.home() / ".cortex"
+    return get_cortex_dir()
 
 
 def _portfolio_dir() -> Path:
@@ -134,8 +136,8 @@ class StoreStatus:
     store: Store
     exists: bool
     age_days: Optional[float]
-    problems: List[str] = field(default_factory=list)   # hard failures
-    warnings: List[str] = field(default_factory=list)   # advisory, not a failure
+    problems: List[str] = field(default_factory=list)  # hard failures
+    warnings: List[str] = field(default_factory=list)  # advisory, not a failure
 
     @property
     def ok(self) -> bool:
@@ -166,9 +168,7 @@ def evaluate(stores: Optional[List[Store]] = None) -> List[StoreStatus]:
         if not exists:
             problems.append(f"missing at {s.path}")
         elif not s.event_driven:
-            past_sla = (
-                s.max_age_days is not None and age is not None and age > s.max_age_days
-            )
+            past_sla = s.max_age_days is not None and age is not None and age > s.max_age_days
             if past_sla and s.scheduled:
                 # A store its own scheduler is supposed to keep fresh is stale:
                 # the scheduler is broken. Hard fail — someone must fix it.
@@ -187,7 +187,11 @@ def evaluate(stores: Optional[List[Store]] = None) -> List[StoreStatus]:
 
         out.append(
             StoreStatus(
-                store=s, exists=exists, age_days=age, problems=problems, warnings=warnings
+                store=s,
+                exists=exists,
+                age_days=age,
+                problems=problems,
+                warnings=warnings,
             )
         )
     return out