feat(competences): extract batch candidates

2026-05-29 11:25:00 +02:00
parent 4ba426c205
commit e8a0fb0e42
60 changed files with 18176 additions and 0 deletions
--- a/data/competences/observed/open_application_via_run.yaml
+++ b/data/competences/observed/open_application_via_run.yaml
@@ -0,0 +1,149 @@
+schema_version: 1
+id: open_application_via_run
+name: Ouvrir une application via Executer
+version: 1
+learning_state: observed
+
+intent:
+  fr: ouvrir une application Windows via la boite Executer
+
+parameters:
+  app_name: notepad
+  expected_process_name: Notepad.exe
+
+preconditions:
+  - id: windows_session_active
+    kind: heartbeat_present
+    max_age_ms: 3000
+  - id: no_blocking_system_dialog
+    kind: not_window_title_matches
+    pattern: "^(UAC|Windows Security|SmartScreen).*"
+
+methods_execution: sequence
+
+methods:
+  - id: step_1_open_run_dialog
+    kind: key_combo
+    primitive_ref: key_combo
+    parameters:
+      keys: ["win", "r"]
+    keys: ["win", "r"]
+    observed: true
+    trace_source: live_events.jsonl
+    trace_event_indices: [3]
+    description: "Ouvre la boite Executer avec Win+R"
+
+  - id: step_2_type_app_name
+    kind: text_input
+    primitive_ref: text_input_focused
+    parameters:
+      text: "notepad"
+      concat_rule: concat_in_order
+    observed: true
+    trace_source: live_events.jsonl
+    trace_event_indices: [6, 7, 9, 10, 11]
+    concat_rule: "join(text_input fragments in segment)"
+    reconstructed_text: "notepad"
+    description: "Saisit le nom de l'application dans la boite Executer"
+
+  - id: step_3_validate_with_enter
+    kind: key_combo
+    primitive_ref: key_combo
+    parameters:
+      keys: ["enter"]
+    keys: ["enter"]
+    observed: false
+    allowed_runtime_substitution: true
+    note: "Trace humaine #13 = mouse_click sur OK. Runtime = key_combo([enter]) equivalent semantique."
+    description: "Valide la boite Executer au runtime"
+
+success_marker:
+  mode: any_of
+  timeout_ms: 5000
+  markers:
+    - kind: active_process_name_is
+      value: Notepad.exe
+  supervised_requires:
+    - kind: active_process_name_is
+      value: Notepad.exe
+      evidence_state: observed_offline
+      required_for: replay_verified
+
+failure_message_template:
+  intention: ouvrir l'application demandee via la boite Executer
+  attendu: voir la fenetre principale de l'application attendue au premier plan
+  vu: "{observed_human_state}"
+  demande: confirmer que l'application est installee sur ce poste, ou m'indiquer un autre moyen de l'ouvrir
+
+chain_refs:
+  source_session: sess_20260324T165824_55b380
+  machine_id: DESKTOP-58D5CAC_windows
+  streaming_session_path: data/training/live_sessions/streaming_sessions/sess_20260324T165824_55b380.json
+  live_events_path: data/training/live_sessions/DESKTOP-58D5CAC_windows/sess_20260324T165824_55b380/live_events.jsonl
+  cleaned_segment:
+    status: documented_offline
+    keep_event_indices: [3, 4, 6, 7, 9, 10, 11, 16]
+    method_event_indices: [3, 6, 7, 9, 10, 11]
+    success_event_indices: [16]
+    excluded_event_indices: [5, 8, 12, 13, 14, 15]
+    stop_before_event_index: 17
+    stop_before:
+      - heartbeat_post_notepad_focus
+      - later_session_activity
+    ignored_between_method_and_success:
+      - action_result_open_run_dialog
+      - heartbeat_without_window_metadata
+      - human_mouse_click_ok_replaced_by_enter_runtime
+      - program_manager_transit_focus
+      - generic_action_result
+    notes:
+      - "Event #3 ouvre la boite Executer via Win+R."
+      - "Events #6/#7/#9/#10/#11 reconstruisent exactement 'notepad'."
+      - "Event #13 est un mouse_click humain sur OK sans anchor_ref; il est exclu de la methode runtime."
+      - "Au runtime, key_combo([enter]) remplace le mouse_click humain pour valider la boite Executer."
+      - "Event #16 prouve le succes par focus_change vers Notepad.exe."
+  workflow_pipeline_id: null
+  graph_node_id: null
+  faiss_state_signatures: []
+  target_memory_keys: []
+  dashboard_knowledge_visible: false
+
+promotion:
+  candidate_requires:
+    - cleaned_segment_validated
+    - method_trace_present
+    - success_marker_defined
+    - failure_message_template_valid
+    - primitive_ref_satisfied
+    - methods_sequence_valid
+  supervised_requires:
+    - replay_verified_once
+    - success_marker_matched_after_action
+    - human_validation
+  stable_requires:
+    min_successes: 3
+    distinct_contexts: 3
+    max_unexplained_failures: 0
+  t2_known_gaps:
+    - id: enter_action_not_in_trace
+      description: "Le mouse_click #13 valide la boite Executer; aucun key_combo([enter]) n'est dans la trace."
+      impact: "Au runtime, Lea emet key_combo([enter]) sans preuve directe dans cette trace humaine."
+      proposed_resolution: "Au replay supervise, utiliser active_process_name_is=Notepad.exe comme preuve de validation."
+      acted_by: Dom
+      acted_at: "2026-05-28T12:45:00+02:00"
+    - id: mouse_click_replaced_by_keyboard_at_runtime
+      description: "La methode runtime diverge de la trace humaine: mouse_click remplace par key_combo([enter])."
+      impact: "La validation T2 doit confirmer que key_combo([enter]) est equivalent fonctionnel dans la boite Executer."
+      proposed_resolution: "Verifier au replay supervise sur plusieurs applications Windows simples."
+      acted_by: Dom
+      acted_at: "2026-05-28T12:45:00+02:00"
+
+generalisation:
+  seen_contexts: []
+  method_success_rate: {}
+  variance_log: []
+
+failure_log: []
+
+created_at: "2026-05-28T12:45:00+02:00"
+last_updated_at: "2026-05-28T12:45:00+02:00"
--- a/data/competences/observed/saisir_requete_recherche.yaml
+++ b/data/competences/observed/saisir_requete_recherche.yaml
@@ -0,0 +1,118 @@
+schema_version: 1
+id: saisir_requete_recherche
+name: Saisir une requete dans la recherche Windows
+version: 1
+learning_state: observed
+
+intent:
+  fr: saisir du texte dans le champ de recherche Windows
+
+parameters:
+  query_text: "test lea apprentissage"
+
+preconditions:
+  - id: open_windows_search_satisfied
+    kind: competence_required
+    competence: open_windows_search
+    state: observed
+  - id: search_field_active
+    kind: active_window
+    any_of:
+      - title_in: ["Rechercher", "Search"]
+      - process_active: SearchHost.exe
+
+methods:
+  - id: text_input_concat
+    kind: text_input
+    primitive_ref: text_input_focused
+    parameters:
+      text: "test lea apprentissage"
+      concat_rule: concat_in_order
+    description: "Saisie texte par fragments dans le champ Rechercher"
+    observed: true
+    trace_source: live_events.jsonl
+    # Les text_input atomises sont concatenes pour former le texte complet
+    concat_rule: "join(all text_input events in segment)"
+    reconstructed_text: "test lea apprentissage"
+    # Note: event #12 "pprentissage" n'est PAS un mot complet
+    # Il complete event #10 "a" pour former "apprentissage"
+
+success_marker:
+  mode: all_of
+  timeout_ms: 5000
+  markers:
+    - kind: active_window_title_in
+      values: ["Rechercher", "Search"]
+    - kind: active_process_name_is
+      value: SearchHost.exe
+    - kind: text_input_reconstructed_equals
+      value: "test lea apprentissage"
+      evidence_source: trace_text_input_concat
+  supervised_requires:
+    - kind: ocr_contains
+      text: "test lea apprentissage"
+      region_hint: search_field
+      evidence_state: hypothesis_offline
+      required_for: supervised_or_replay_verified
+
+failure_message_template:
+  intention: saisir du texte dans la recherche Windows
+  attendu: voir le texte saisi apparaitre dans le champ Rechercher
+  vu: "{observed_human_state}"
+  demande: saisir le texte attendu dans le champ Rechercher puis me rendre la main
+
+chain_refs:
+  source_session: sess_20260527T185155_98ad9a
+  machine_id: DESKTOP-58D5CAC_windows
+  streaming_session_path: data/training/live_sessions/streaming_sessions/sess_20260527T185155_98ad9a.json
+  live_events_path: data/training/live_sessions/DESKTOP-58D5CAC_windows/sess_20260527T185155_98ad9a/live_events.jsonl
+  cleaned_segment:
+    status: documented_offline
+    keep_event_indices: [5, 6, 7, 8, 9, 10, 11, 12, 13]
+    method_event_indices: [5, 6, 8, 9, 10, 12]
+    success_event_indices: [7, 11, 13]
+    excluded_event_indices: []
+    stop_before_event_index: 14
+    stop_before:
+      - mouse_click_systray
+      - explorer_overflow_window
+      - pythonw_unknown_focus
+    ignored_after_success: []
+    notes:
+      - "Events #5/#6 sont exclus du P0 (open_windows_search) car ils appartiennent a la saisie P1 apres Win+S."
+      - "P1 commence a #5, la premiere saisie apres l'ouverture de la recherche"
+      - "Event #7 heartbeat post-action P0, confirme que SearchHost.exe est actif pendant la saisie"
+      - "Event #12 'pprentissage' complete #10 'a' pour former 'apprentissage'"
+      - "Texte reconstruit: 'test lea apprentissage' (22 chars)"
+  workflow_pipeline_id: null
+  graph_node_id: null
+  faiss_state_signatures: []
+  target_memory_keys: []
+  dashboard_knowledge_visible: false
+
+promotion:
+  candidate_requires:
+    - cleaned_segment_validated
+    - method_trace_present
+    - success_marker_defined
+    - failure_message_template_valid
+    - competence_dependency_satisfied
+  supervised_requires:
+    - replay_verified_once
+    - success_marker_matched_after_action
+    - ocr_or_replay_verified_text
+    - human_validation
+  stable_requires:
+    min_successes: 3
+    distinct_contexts: 3
+    max_unexplained_failures: 0
+
+generalisation:
+  seen_contexts: []
+  method_success_rate: {}
+  variance_log: []
+
+failure_log: []
+
+created_at: "2026-05-27T18:51:55+02:00"
+last_updated_at: "2026-05-28T08:13:52+02:00"
--- a/data/competences/observed/scroll_down_pdf_edge.yaml
+++ b/data/competences/observed/scroll_down_pdf_edge.yaml
@@ -0,0 +1,118 @@
+schema_version: 1
+id: scroll_down_pdf_edge
+name: Scroller vers le bas dans un PDF Edge
+version: 1
+learning_state: observed
+
+intent:
+  fr: faire defiler un document PDF vers le bas dans Microsoft Edge
+
+parameters: {}
+
+preconditions:
+  - id: edge_pdf_active
+    kind: active_window
+    any_of:
+      - process_active: msedge.exe
+
+methods:
+  - id: scroll_down_mouse
+    kind: scroll
+    primitive_ref: scroll_view
+    parameters:
+      direction: down
+      amount: 9
+      unit: lines
+    description: "Scroll vers le bas via molette souris dans un PDF Edge"
+    observed: true
+    trace_source: live_events.jsonl
+    trace_event_indices: [129, 130, 131, 133, 134, 135, 137, 138, 139]
+
+success_marker:
+  mode: all_of
+  timeout_ms: 5000
+  markers:
+    - kind: active_process_name_is
+      value: msedge.exe
+  supervised_requires:
+    - kind: ocr_contains
+      text: "contenu different apres scroll"
+      region_hint: document_body
+      evidence_state: hypothesis_offline
+      required_for: supervised_or_replay_verified
+
+failure_message_template:
+  intention: faire defiler le PDF vers le bas
+  attendu: le contenu visible doit changer apres le defilement
+  vu: "{observed_human_state}"
+  demande: indiquer si le document PDF actif peut defiler vers le bas
+
+chain_refs:
+  source_session: sess_20260318T010719_62a058
+  machine_id: DESKTOP-58D5CAC_windows
+  streaming_session_path: data/training/live_sessions/streaming_sessions/sess_20260318T010719_62a058.json
+  live_events_path: data/training/live_sessions/DESKTOP-58D5CAC_windows/sess_20260318T010719_62a058/live_events.jsonl
+  cleaned_segment:
+    status: documented_offline
+    keep_event_indices: [126, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140]
+    method_event_indices: [129, 130, 131, 133, 134, 135, 137, 138, 139]
+    success_event_indices: [140]
+    excluded_event_indices: [127, 128]
+    stop_before_event_index: 141
+    stop_before:
+      - subsequent_scroll_bursts
+      - heartbeat_without_window_metadata_later_in_session
+    ignored_between_method_and_success: []
+    notes:
+      - "Event #126 focus vers msedge.exe avec le PDF RapportS1 actif."
+      - "Events #129/#130/#131/#133/#134/#135/#137/#138/#139 sont des mouse_scroll dans msedge.exe."
+      - "Tous les events de methode ont delta [0, -1], ce qui prouve direction=down pour cette trace."
+      - "Events #132 et #136 sont des heartbeats sans metadonnees fenetre au milieu du burst."
+      - "Events #127/#128 sont un clic de positionnement et son action_result avant le burst scroll; ils sont exclus."
+      - "Event #140 est le premier mouse_scroll post-methode avec msedge.exe encore actif; il prouve la continuite active, pas le changement de contenu."
+  workflow_pipeline_id: null
+  graph_node_id: null
+  faiss_state_signatures: []
+  target_memory_keys: []
+  dashboard_knowledge_visible: false
+
+promotion:
+  candidate_requires:
+    - cleaned_segment_validated
+    - method_trace_present
+    - success_marker_defined
+    - failure_message_template_valid
+    - primitive_ref_satisfied
+    - scroll_trace_validated
+  supervised_requires:
+    - replay_verified_once
+    - success_marker_matched_after_action
+    - ocr_or_replay_verified_scroll_effect
+    - human_validation
+  stable_requires:
+    min_successes: 3
+    distinct_contexts: 3
+    max_unexplained_failures: 0
+  t2_known_gaps:
+    - id: scroll_effect_not_observed_offline
+      description: "La trace prouve les mouse_scroll et la fenetre active, mais pas le changement visuel du contenu PDF."
+      impact: "Le niveau T2 doit verifier que le viewport ou le texte visible change apres le scroll."
+      proposed_resolution: "Ajouter OCR runtime, screenshot diff ou marker visuel avant promotion supervised."
+      acted_by: Dom
+      acted_at: "2026-05-28T14:20:00+02:00"
+    - id: no_ocr_offline
+      description: "Aucune preuve OCR avant/apres scroll n'est disponible dans cette validation offline."
+      impact: "Le success_marker offline reste une preuve de continuite active, pas une preuve de contenu different."
+      proposed_resolution: "Verifier par OCR ou replay supervise avant promotion supervised."
+      acted_by: Dom
+      acted_at: "2026-05-28T14:20:00+02:00"
+
+generalisation:
+  seen_contexts: []
+  method_success_rate: {}
+  variance_log: []
+
+failure_log: []
+
+created_at: "2026-05-28T14:20:00+02:00"
+last_updated_at: "2026-05-28T14:20:00+02:00"