v1.2: Framework Owns Flow Control (FOC-01..FOC-06)

.gitignore

@@ -54,6 +54,8 @@ docs/
 REVIEW_*.md
 review_*.md
 .planning/
+# Dev integration test driver (out-of-repo tool, runs against live UI).
+scripts/integration_scenarios.py
 
 # Coverage / CI artefacts
 coverage.xml

config/code_review.runtime.yaml

@@ -41,6 +41,20 @@ paths:
 # When no rule fires the session falls through to ``unreviewed``
 # (the v1.0 framework-default failure mode).
 orchestrator:
+  # Phase 11 (FOC-04): declarative HITL gating policy. Framework
+  # default threshold (0.7) -- code review is less prod-blast-radius
+  # than incident remediation so the stricter incident threshold
+  # (0.8) is unwarranted here.
+  gate_policy:
+    confidence_threshold: 0.7
+    gated_environments: [production]
+    gated_risk_actions: [approve]
+  # Phase 12 (FOC-05): declarative retry policy. Framework default --
+  # max_retries=2, transient retries on, confidence floor 0.4.
+  retry_policy:
+    max_retries: 2
+    retry_on_transient: true
+    retry_low_confidence_threshold: 0.4
   entry_agent: intake
   default_terminal_status: unreviewed
   statuses:
@@ -85,6 +99,16 @@ orchestrator:
   # state_overrides; orchestrator validates start_session's
   # state_overrides kwarg against this class.
   state_overrides_schema: examples.code_review.state.CodeReviewStateOverrides
+  # Phase 9 (D-09-02 / FOC-01 / FOC-02): session-derived tool-arg
+  # injection map. code_review's pr_url / repo live under
+  # ``Session.extra_fields`` (the framework-default Session has no
+  # typed fields for them) so the dotted paths reach into the dict.
+  # The framework's ``_resolve_dotted`` walks dict-valued attrs
+  # transparently.
+  injected_args:
+    session_id: session.id
+    pr_url: session.extra_fields.pr_url
+    repo: session.extra_fields.repo
 # Cross-cutting framework knobs read directly off AppConfig.framework.
 framework:
   # Per-app session-id prefix. Threaded through SessionStore into

config/config.yaml

@@ -21,10 +21,14 @@ llm:
       endpoint: ${AZURE_ENDPOINT}
       api_version: 2024-08-01-preview
       api_key: ${AZURE_OPENAI_KEY}
+    openrouter:
+      kind: openai_compat
+      base_url: https://openrouter.ai/api/v1
+      api_key: ${OPENROUTER_API_KEY}
   models:
     workhorse:
-      provider: ollama_cloud
-      model: gpt-oss:120b
+      provider: openrouter
+      model: openai/gpt-4o-mini
       temperature: 0.0
     cheap:
       provider: ollama_cloud
@@ -135,6 +139,19 @@ dedup:
 # ``incident_management.yaml`` since this is the bundled deployment
 # config for the example app.
 orchestrator:
+  # Phase 11 (FOC-04): declarative HITL gating policy. Framework
+  # default (threshold 0.7) -- mirrors incident_management v1.1
+  # behaviour with the production-class environment gate.
+  gate_policy:
+    confidence_threshold: 0.7
+    gated_environments: [production]
+    gated_risk_actions: [approve]
+  # Phase 12 (FOC-05): declarative retry policy. Framework default --
+  # max_retries=2, transient retries on, confidence floor 0.4.
+  retry_policy:
+    max_retries: 2
+    retry_on_transient: true
+    retry_low_confidence_threshold: 0.4
   entry_agent: intake
   default_terminal_status: needs_review
   statuses:
@@ -186,6 +203,15 @@ orchestrator:
   # state_overrides; orchestrator validates the start_session
   # kwarg against this class.
   state_overrides_schema: examples.incident_management.state.IncidentStateOverrides
+  # Phase 9 (D-09-02 / FOC-01 / FOC-02): session-derived tool-arg
+  # injection map. Strips the named args from each tool's LLM-visible
+  # signature and re-supplies them from the live Session at invocation
+  # time. Mirrors incident_management.yaml since this file is the
+  # bundled deployment config for the example app.
+  injected_args:
+    environment: session.extra_fields.environment
+    incident_id: session.id
+    session_id: session.id
 runtime:
   # Wires the orchestrator and storage layer to the incident-management
   # domain state class (see examples/incident_management/state.py).

config/incident_management.yaml

@@ -16,6 +16,24 @@ similarity_method: keyword
 # ``_TERMINAL_TOOL_RULES`` table in ``orchestrator.py`` (Phase 6 /
 # DECOUPLE-02 / DECOUPLE-03 / D-06-01..06).
 orchestrator:
+  # Phase 11 (FOC-04): declarative HITL gating policy. Tighter
+  # threshold than the framework default -- incident remediation
+  # pauses on production-class medium-risk tools and on any tool
+  # call below 80% turn confidence.
+  gate_policy:
+    confidence_threshold: 0.8
+    gated_environments: [production]
+    gated_risk_actions: [approve]
+  # Phase 12 (FOC-05): declarative retry policy. Default
+  # max_retries=2 mirrors the v1.2 ROADMAP. retry_on_transient=true
+  # keeps current auto-retry-on-network-blip behaviour.
+  # retry_low_confidence_threshold=0.4 sits below the gate_policy
+  # confidence_threshold (0.8) so the gate fires HITL approval
+  # before the retry path even considers a low-confidence give-up.
+  retry_policy:
+    max_retries: 2
+    retry_on_transient: true
+    retry_low_confidence_threshold: 0.4
   entry_agent: intake
   default_terminal_status: needs_review
   statuses:
@@ -74,6 +92,15 @@ orchestrator:
   # state_overrides; orchestrator validates the start_session
   # kwarg against this class.
   state_overrides_schema: examples.incident_management.state.IncidentStateOverrides
+  # Phase 9 (D-09-02 / FOC-01 / FOC-02): session-derived tool-arg
+  # injection map. Each entry strips the named arg from every tool's
+  # LLM-visible signature and re-supplies the value from the live
+  # Session at invocation time. The LLM cannot hallucinate values
+  # for args it cannot see.
+  injected_args:
+    environment: session.environment
+    incident_id: session.id
+    session_id: session.id
 
 # Cross-cutting framework knobs the runtime consumes directly.
 framework: