Fixes:
1. Rewrote run_competitive_benchmark.py to test BOTH capabilities:
- Telepathic (JSON element extraction)
- Brain (Free-text action extraction with format_json=False)
2. Normalizes scores by averaging per-scenario (fixes score inflation
where models with more scenarios tested scored higher but were marked unsuitable).
3. Added 4 new brain_action scenarios to ensure the 'think=false' code path
is actively benchmarked going forward.
4. Added test_benchmark_integrity.py to lock in scenario format rules.
5. Cleared stale llm_benchmarks.json data to force clean re-evaluations.
ROOT CAUSE: qwen3.5 (reasoning model) returns response='' with thinking
block containing all reasoning. llm_provider.py line 352 silently
substituted the thinking block as the response via:
content = raw_response or raw_thinking or ''
The Brain then extracted random actions from the reasoning text.
FIXES:
1. llm_provider.py: Conditional thinking isolation
- format_json=True (SAE/perception): thinking fallback preserved
- format_json=False (Brain): thinking NEVER substituted
- Added think=false for Ollama free-text calls to force direct response
2. planner.py: No-Op Guard strips tab actions that navigate to
the current screen (e.g. 'tap profile tab' on OWN_PROFILE)
3. test_brain_live.py: Stochastic testing (5 runs, 60% min valid)
to handle non-deterministic LLM behavior reliably
4. tests/integration/test_llm_provider_pipeline.py: NEW test layer
mocking at HTTP level (requests.post) to exercise the FULL
llm_provider → Brain pipeline. This would have caught the
thinking substitution bug from day one.
Suite: 168 passed, 0 failed
- Add permanent mock ban guard in root conftest.py that fails any test
importing unittest.mock at COLLECTION TIME (before execution)
- Add 8 brain output contract tests reproducing the exact production bug:
LLM thinks 'press back' but parser extracts 'tap messages tab' from
the <think> block
- Add UI change noise threshold (MIN_UI_CHANGE_BYTES=50) to prevent
false-positive 'ui_changed' from 1-byte XML diffs (timestamps/whitespace)
- Verify planner correctly strips masked actions from Brain prompt
The central autonomous brain (GoalExecutor.achieve()) had ZERO E2E coverage.
The deleted lying test_e2e_autonomous_session.py never called it at all,
allowing the AttributeError and dead code bugs to survive undetected.
New tests exercise the REAL achieve() with production XML fixture sequences:
- Navigation: HOME_FEED → tap explore tab → EXPLORE_GRID (HD Map routing)
- Already-on-target recognition (0-step achievement)
- max_steps exhaustion → returns False (anti-infinite-loop)
- Return type contract enforcement (bool, not string)
All 4 tests use make_real_device_with_xml with real fixture sequences.
No mocks. No patches. No lies.
E2E: 60 passed, 5 skipped, 0 failures.
CRITICAL LIES FIXED:
- bot_flow.py:474 compared achieve() (returns bool) to 'GOAL_ACHIEVED'
(string). Success path was dead code — True never == string.
- TestBotFlowDMGating built its own local target_map dict and asserted
against it. bot_flow.py no longer has target_map (uses GoalExecutor).
Tests verified their own imagination, not production code.
- test_perception_mock_theater_purged was a skip+pass ghost creating
false 'skipped' coverage in reports.
- test_perceive_notification_shade silently passed on FileNotFoundError
instead of reporting the missing fixture.
- test_resolve_uses_visual_discovery_when_device_available only checked
hasattr — verifying method existence, not behavior.
PRODUCTION BUGS FIXED:
- GoalExecutor constructor called with wrong args (memory, telepathic,
config, session_state) — it only accepts (device, bot_username).
- achieve() result comparison was dead code: always hit warning branch.
E2E: 57 passed, 4 skipped (live_llm waivers), 0 failures.
- Added strict TDD coverage for all autonomous changes.
- Implemented GrowthBrain.get_current_goal to select high-level objectives.
- Replaced procedural orchestrator with GoalExecutor in bot_flow.
- Purged hardcoded resource-ids in dm_engine in favor of ScreenIdentity.
- Removed regex parsing in unfollow_engine in favor of telepathic semantic extraction.
Two root causes for 'scroll on story' bug:
1. ScreenIdentity had ZERO structural markers for story views.
reel_viewer_media_layout, reel_viewer_header, reel_viewer_progress_bar
and content-desc 'Like Story'/'Send story' now → STORY_VIEW.
2. Brain prompt was prescriptive ('you MUST scroll down'), overriding
the LLM's intelligence. Rewritten to give context about screen types
and let the AI reason autonomously about which action makes sense.
Philosophy: AI decides navigation, we provide correct perception data.
No hardcoded 'if story → press back' escape hatch.
4 new perception tests (all green), 0 regressions.
Bug evidence from run 2026-04-27_23-46-57:
- Bot started on a story (reel_viewer_media_layout, 'Like Story')
- ScreenIdentity classified it as UNKNOWN
- GOAP chose 'scroll down' 4 times (stories don't scroll)
- Bot was trapped in infinite scroll loop
Captured real XML fixture: story_view_full.xml
1 test FAILS (screen_identity → UNKNOWN instead of STORY_VIEW)
Kill-Switch: Refuse DM processing when dm_reply.enabled=false in config.
Root cause: checked nonexistent 'disable_ai_messaging' flag instead of
actual plugin config.
Context Guard: Skip threads with no extractable message text.
Root cause: LLM was fed 'No previous context' → produced garbage like
'the to the'.
Send Verification: Structurally verify Send button resource-id/desc.
Root cause: VLM returned reactions_pill_container, edit fields, etc.
and engine blindly clicked them, logging 'Successfully sent'.
Iteration Cap: MAX_REPLIES_PER_INBOX_VISIT=3 prevents spam.
Root cause: no loop guard → 8 DMs sent in 2 minutes in production.
Refactored: removed dead 'if True' guard, de-indented block,
moved dm_memory.log_sent_dm into success branch only.
All 6 E2E tests pass. No regressions (54/55 passed, 1 pre-existing).
Tests expose:
1. DM Engine ignores dm_reply.enabled config (checks nonexistent 'disable_ai_messaging')
2. Logs 'Successfully sent' without verifying actual Send button click
3. Generates garbage replies from 'No previous context' (story replies)
4. No max-iteration guard — sent 20 messages in test, 8 in production
All 4 tests FAIL. Ready for GREEN phase.
Implements the _intent_matches_node() guard — a shared SSOT function that
validates clicked elements against intent keywords before trusting any
verification result.
Fixes applied:
1. action_memory.py: verify_success() now cross-checks clicked element
against intent BEFORE trusting structural delta for toggle actions
2. action_memory.py: confirm_click() blocks Qdrant poisoning when the
tracked click doesn't semantically match the intent
3. q_nav_graph.py: 'follow' added to action_checks map (screen-sanity)
4. goap.py: Pre-click semantic guard prevents device.click() on elements
that don't match toggle intents (follow/like/save)
TOGGLE_INTENT_MARKERS dict is SSOT for intent→element validation keywords.
Supports DE locale (gefolgt, abonnieren, gefällt, speichern).
162 passed, 0 regressions. All 5 previously-RED tests now GREEN.
TDD RED Phase: These tests PROVE the gaps that allowed the production bug
where the bot logged 'Followed @missiongreenenergy ✓' after clicking a photo grid item.
5 RED tests expose:
1. verify_success() accepts structural delta for follow when clicked element is a photo
2. verify_success() accepts 500-char delta without semantic match check
3. QNavGraph.do() missing 'follow' in action_checks screen-sanity map
4. ActionMemory.confirm_click() poisons Qdrant with mismatched intent→element
5. GOAP._execute_action() clicks first without pre-click sanity check
All 5 tests FAIL (RED) as expected — proving the lies in the current test suite.
No production code was changed.
BREAKING: IntentResolver now resolves intents by SEEING the screenshot
instead of parsing XML text descriptions.
Architecture:
- PRIMARY: Visual Discovery (SoM) — annotates screenshot with numbered
bounding boxes, sends to VLM, VLM visually picks the right box
- FALLBACK: Text-based VLM resolution (only when no device available)
- Removed: _visual_critic (redundant — visual discovery IS visual)
- Removed: _humanize_desc regex (the VLM reads the actual screen now)
Key innovations:
- Spatial Deduplication: child nodes fully contained in parent bounds
are suppressed (83 → ~19 boxes), eliminating visual noise
- System UI filtering: statusbar, notifications excluded from candidates
- VLM prompt is pure visual: 'look at the numbered boxes and pick one'
Proven by live LLM test: VLM correctly identifies 'following' (not
'followers') by SEEING the screen content, with zero string matching.
