Add Simurgh turn intent routing frame

Author: alireza787b

CHANGELOG.md

@@ -14,6 +14,9 @@ and this project uses simple two-part versioning: `X.Y` (Major.Minor).
   polish, and hover-only copy controls for a cleaner operator experience.
 - Simurgh documentation/context coverage for PX4-first MDS support boundaries,
   MAVLink routing, and read-only fleet/telemetry guidance.
+- Simurgh turn-level semantic intent frame for one coherent route decision per
+  operator message before confirmation, action drafting, read-only tools, or
+  provider fallback, with sanitized trace metadata for PM/test inspection.
 - Simurgh guarded action monitoring for long-running flight commands and SITL
   lifecycle operations, including same-session implicit target inference and
   conditional post-action cleanup after terminal success.
@@ -32,6 +35,10 @@ and this project uses simple two-part versioning: `X.Y` (Major.Minor).
   fields for dashboard progress rendering instead of requiring text scraping.
 
 ### Fixed
+- Simurgh no longer lets bare approval wording such as "go ahead" confirm an
+  old pending action when the same message is actually a new read/status task.
+- Simurgh no longer drafts flight actions from advisory motion-status questions
+  such as asking whether a drone should land or can RTL safely.
 - Simurgh no longer treats fresh action requests that contain approval-like
   wording such as "send it to test flight" as bare confirmations; PM-style
   takeoff-wait-move-return sequences now reach the guarded action planner.

docs/agent-context/generated/simurgh-docs-index.json

@@ -28,10 +28,43 @@ read-only operator slice plus selected guarded actions:
 - generated public docs/chunk index with MCP search and bounded chunk retrieval
 - read-only drone log and onboard PX4 ULog metadata summaries through approved
   GCS-side log endpoints, without raw ULog download/erase exposure
+- a provider-neutral turn-level semantic intent frame that produces one coherent
+  routing decision before confirmation, action drafting, read-only tools, or
+  provider fallback
 - guarded SITL lifecycle actions through canonical GCS SITL Control routes
 - guarded curated flight-command drafts through canonical GCS command routes
 - no direct drone API exposure
 
+## Semantic Understanding Boundary
+
+Simurgh should not rely on scattered keyword checks as the user-facing
+intelligence layer. Each assistant turn is first interpreted as one structured
+intent frame that records the adapted routing text, read-only plan, possible
+action draft, confirmation/rejection signal, confidence, and reasons. The route
+then consumes that single frame instead of independently asking several regex
+classifiers whether the message is a confirmation, action, read-only query, or
+provider turn.
+
+The sanitized response trace exposes this frame under `trace.intent` so PM
+tests, dashboard UI, and future agent tooling can explain why a turn routed as a
+read-only check, action draft, confirmation, rejection, or provider fallback
+without scraping answer prose.
+
+This frame is an interpretation layer, not an authority layer:
+
+- semantic interpretation may decide that "go ahead and check SITL instances" is
+  a read/status task instead of an approval;
+- semantic interpretation may preserve a multi-step action plan such as
+  takeoff, wait, move north, then RTL;
+- deterministic enforcement still owns registry schema validation, runtime mode,
+  approval, circuit breaker, auth, command submission, monitoring, and audit.
+
+Future provider-backed structured-output classifiers should plug into this same
+frame contract. They may improve language, tone, paraphrase, and target-memory
+understanding, but they must still emit typed plans that the existing policy and
+executor layers validate. Provider prose must never be treated as approval or as
+proof that an action executed.
+
 ## Dashboard Chat UX Contract
 
 The `/simurgh` dashboard surface should stay chat-first and low-noise:

docs/guides/simurgh-operator.md

@@ -39,19 +39,22 @@ Implemented:
 
 - deterministic language/tone profile;
 - config-driven query adaptation in `config/agent_query_adaptation.yaml`;
+- provider-neutral turn-level semantic intent frame in
+  `gcs-server/agent_runtime/turn_intent.py`, consumed by the dashboard
+  assistant route before confirmation/action/read-only/provider branching;
 - local read-only tool routing and selected guarded actions;
 - provider composition for safe text turns;
 - optional public web-search lane for public/current facts;
 - action confirmation, circuit breaker, and audit enforcement outside provider
   prose;
 - dashboard prompt evals for PM-style conversations.
 
-Gap:
+Remaining gap:
 
-- route selection still leans too heavily on deterministic wording rules;
-- multi-step action intent can collapse to the last command when the planner
-  misses sequence semantics;
-- target memory is not yet treated as a first-class structured input/output;
+- provider-backed structured-output semantic classification is not yet enabled
+  in the action path; the current frame is provider-neutral and testable;
+- target memory is partially structured through last action/result context, but
+  broader live-fleet target inference still needs careful safety review;
 - some failures still answer from docs instead of running the most relevant
   local evidence tool;
 - answer localization and tone adaptation are not yet uniformly available for
@@ -188,12 +191,26 @@ understanding, target memory, planner, executor, answer composition, or UI.
 - New failures add evals first; aliases are added only when they meet the rule
   boundary above.
 
-## Current Handoff Note
+## Current Implementation Checkpoint
 
-This document does not change runtime behavior. The deployed `v5.5.96` fix for
-SITL-created-drone readiness follow-ups remains the current PM-test baseline.
-The next Simurgh slice should implement semantic-understanding dry-run and
-evals before more alias expansion.
+The first implementation slice adds a structured, provider-neutral
+`TurnIntentFrame`, routes dashboard assistant turns through it, and exposes the
+sanitized interpretation under `trace.intent`. The frame prevents two PM-visible
+failures:
+
+- approval-like wording with a new read/status task, such as "go ahead and
+  check SITL instances now", no longer confirms an old pending action;
+- advisory motion questions, such as "tell me if drone 1 should land", no
+  longer draft guarded flight commands.
+
+Regression coverage now includes frame-level tests and dashboard-route tests for
+PM-style compound action plans, exact draft confirmations, read/status task
+arbitration, and advisory-vs-command flight wording.
+
+The next Simurgh slice should use the same frame contract for optional
+provider-backed structured semantic classification, broader target memory, and
+multilingual/tone-sensitive evals. Do not add broad alias lists as a substitute
+for this frame.
 
 ## References Reviewed
 

docs/plans/2026-06-20-simurgh-language-intent-modernization.md

@@ -657,6 +657,19 @@ def _extract_sitl_instance_names(normalized: str) -> list[str]:
 
 
 def _looks_conceptual(normalized: str) -> bool:
+    flight_action_terms = r"land|landing|rtl|return|take\s*off|takeoff"
+    advisory_terms = r"status|ready|safe|safely|should|whether|if"
+    if re.search(
+        rf"\b(tell\s+me|show\s+me|check|report)\b.{{0,80}}\b(?:{flight_action_terms})\b.{{0,80}}\b(?:{advisory_terms})\b",
+        normalized,
+    ):
+        return True
+    if re.search(rf"\b(?:{advisory_terms})\b.{{0,80}}\b(?:{flight_action_terms})\b", normalized):
+        return True
+    if re.search(rf"\b(?:{flight_action_terms})\b.{{0,80}}\b(?:{advisory_terms})\b", normalized):
+        return True
+    if re.search(r"\bcan\s+(?:drone|vehicle|aircraft)\s+\d+\b.{0,80}\b(land|rtl|return)\b", normalized):
+        return True
     instructional = (
         "how to",
         "how do i",

gcs-server/agent_runtime/action_planner.py

@@ -2855,6 +2855,11 @@ def create_assistant_turn(
             "provider_composed_from_previous_evidence": provider_composed_from_previous_evidence,
             "evidence_followup_kind": evidence_followup_kind if tool_intent == "evidence_followup" else None,
             "provider_composition_error": provider_composition_error,
+            "turn_intent": (
+                dict(metadata.get("turn_intent") or {})
+                if isinstance(metadata, Mapping) and isinstance(metadata.get("turn_intent"), Mapping)
+                else {}
+            ),
             "query_adaptation": query_adaptation.public_metadata(),
             "routing_strategy": query_adaptation.strategy,
             "routing_language": query_adaptation.routing_language,