T5-P1-A3-WP1 — Prevent Orchestrator From Improvising Recovery on Degraded Tool Responses (#4067)

## Summary

Add an explicit catch-all halt rule to both the cook-session
orchestrator prompt (`_prompts_orchestrator.py`) and the fleet
food-truck prompt (`_prompts.py`) that declares any tool response with
`success:false` AND absent/null/empty `content` as a degraded state. The
orchestrator must print `user_visible_message`, refuse to improvise
recovery, and halt. The fleet prompt must emit the sentinel with
`reason="degraded_tool_response"` and stop. The canonical sentinel
contract in `_type_dispatch_identity.py` must also include the new
reason value. Both prompt builders get corresponding contract tests.

## Implementation Plan

Plan file:
`/home/talon/projects/autoskillit-runs/impl-20260611-081716-639257/.autoskillit/temp/make-plan/t5_p1_a3_wp1_degraded_tool_response_halt_plan_2026-06-11_081900.md`

Closes #4000

🤖 Generated with [Claude Code](https://claude.com/claude-code) via
AutoSkillit
<!-- autoskillit:pipeline-signature
steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr
-->

## Token Usage Summary

| Step | Model | count | uncached | output | cache_read | peak_ctx |
turns | cache_write | time |

|------|-------|-------|----------|--------|------------|----------|-------|-------------|------|
| plan* | opus[1m] | 1 | 1.8k | 13.5k | 1.1M | 87.2k | 41 | 70.3k | 9m
9s |
| verify* | sonnet | 1 | 108 | 8.8k | 849.3k | 84.4k | 42 | 63.6k | 2m
59s |
| implement* | MiniMax-M3 | 1 | 2.1M | 7.6k | 0 | 0 | 59 | 0 | 4m 29s |
| audit_impl* | sonnet | 1 | 882 | 7.5k | 207.1k | 43.4k | 16 | 31.2k |
3m 36s |
| prepare_pr* | MiniMax-M3 | 1 | 200.4k | 1.8k | 0 | 0 | 13 | 0 | 49s |
| compose_pr* | MiniMax-M3 | 1 | 211.8k | 1.3k | 0 | 0 | 13 | 0 | 40s |
| **Total** | | | 2.5M | 40.5k | 2.1M | 87.2k | | 165.1k | 21m 43s |

\* *Step used a non-Anthropic provider; caching behavior may differ.*

## Token Efficiency

| Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC |
|------|-------------|----------------|-----------------|------------|
| plan | 0 | — | — | — |
| verify | 0 | — | — | — |
| implement | 116 | 0.0 | 0.0 | 65.8 |
| audit_impl | 0 | — | — | — |
| prepare_pr | 0 | — | — | — |
| compose_pr | 0 | — | — | — |
| **Total** | **116** | 18525.6 | 1423.5 | 349.3 |

## Model Usage Breakdown

| Model | steps | uncached | output | cache_read | cache_write | time |
|-------|-------|----------|--------|------------|-------------|------|
| opus[1m] | 1 | 1.8k | 13.5k | 1.1M | 70.3k | 9m 9s |
| sonnet | 2 | 990 | 16.4k | 1.1M | 94.8k | 6m 35s |
| MiniMax-M3 | 3 | 2.5M | 10.7k | 0 | 0 | 5m 58s |

Author: Trecek

src/autoskillit/cli/_prompts_orchestrator.py

@@ -163,6 +163,16 @@ def _build_orchestrator_prompt(
     2. DO NOT call AskUserQuestion.
     3. End the session with a final text response.
 
+FAILURE PREDICATE — DEGRADED TOOL RESPONSE:
+  If ANY tool response contains `"success": false` AND the `content` field is
+  absent, null, or an empty string — this is a degraded state.
+    1. Extract and print the value of "user_visible_message" from the
+       response (fall back to the raw response text if the field is missing).
+    2. DO NOT call AskUserQuestion.
+    3. Do not improvise a recovery path. Do not route to on_failure.
+       Do not retry the tool without explicit instructions from the recipe step.
+    4. End the session with a final text response.
+
 CONTEXT LIMIT ROUTING — run_skill only (check BEFORE on_failure):
 - When run_skill returns "success: False" AND "needs_retry: true" AND "retry_reason: resume":
   - Check "subtype" to discriminate the termination cause:

src/autoskillit/core/types/_type_dispatch_identity.py

@@ -32,7 +32,7 @@ def _build_sentinel_contract(dispatch_id: str, short: str) -> str:
 Fields:
 - success: true if all mandatory steps completed without unresolved failures
 - reason: "completed", "failed", "fleet_quota_exhausted", "timeout",
-  "open_kitchen_failed", "missing_on_failure"
+  "open_kitchen_failed", "missing_on_failure", "degraded_tool_response"
 - summary: One-line description of what happened
 
 The sentinel markers ---l3-result::{dispatch_id}--- and ---end-l3-result::{dispatch_id}---

src/autoskillit/fleet/_prompts.py

@@ -206,6 +206,13 @@ def _build_food_truck_prompt(
     1. Emit the sentinel block with success=false and reason="open_kitchen_failed".
     2. End the session.
 
+FAILURE PREDICATE — DEGRADED RESPONSE:
+  If ANY tool response contains success=false AND the content field is
+  absent, null, or an empty string — this is a degraded state.
+    1. Emit the sentinel block with success=false and reason="degraded_tool_response".
+    2. Stop. Do not improvise a recovery path, do not route to on_failure,
+       and do not retry the tool.
+
 TWO FAILURE TIERS FOR PREDICATE-FORMAT STEPS:
 - Tool-level failure (run_skill returns "success: False"): Follow on_failure. This fires
   BEFORE any result object exists. on_result conditions are NOT evaluated.
@@ -340,7 +347,7 @@ def _build_food_truck_prompt(
 Fields:
 - success: true if all mandatory steps completed without unresolved failures
 - reason: "completed", "failed", "quota_exhausted", "timeout",
-  "open_kitchen_failed", "missing_on_failure"
+  "open_kitchen_failed", "missing_on_failure", "degraded_tool_response"
 - summary: One-line description of what happened{extra_fields_docs}
 
 The sentinel markers ---l3-result::{dispatch_id}--- and ---end-l3-result::{dispatch_id}---

tests/cli/test_orchestrator_prompt_contract.py

@@ -51,6 +51,47 @@ def test_open_kitchen_predicate_uses_json_field_not_substring(self):
         assert "ingredients_table" in section
 
 
+class TestDegradedToolResponsePredicate:
+    """Guards for the FAILURE PREDICATE — DEGRADED TOOL RESPONSE block."""
+
+    def _get_section(self) -> str:
+        prompt = _get_prompt()
+        start = prompt.index("FAILURE PREDICATE — DEGRADED TOOL RESPONSE")
+        end = prompt.index("CONTEXT LIMIT ROUTING", start)
+        return prompt[start:end]
+
+    def test_prompt_contains_degraded_tool_response_section(self):
+        prompt = _get_prompt()
+        ok_idx = prompt.index("FAILURE PREDICATE — open_kitchen")
+        deg_idx = prompt.index("FAILURE PREDICATE — DEGRADED TOOL RESPONSE")
+        ctx_idx = prompt.index("CONTEXT LIMIT ROUTING")
+        assert ok_idx < deg_idx < ctx_idx, (
+            "Degraded predicate must appear after open_kitchen and before CONTEXT LIMIT ROUTING"
+        )
+
+    def test_degraded_predicate_covers_success_false_empty_content(self):
+        section = self._get_section()
+        assert "success" in section
+        assert "false" in section
+        assert any(word in section for word in ("absent", "null", "empty")), (
+            "Section must reference absent, null, or empty content"
+        )
+
+    def test_degraded_predicate_prohibits_improvisation(self):
+        section = self._get_section()
+        lower = section.lower()
+        assert any(
+            phrase in lower
+            for phrase in (
+                "do not improvise",
+                "do not retry",
+                "do not attempt",
+                "never",
+                "halt",
+            )
+        ), "Section must prohibit improvisation or retry"
+
+
 class TestStep0ToolPredicateCoverage:
     """Every tool referenced in STEP 0 must have a failure predicate or shared rule."""
 

tests/fleet/test_food_truck_prompt.py

@@ -174,3 +174,57 @@ def test_food_truck_open_kitchen_predicate_uses_json_field_not_substring():
     section = prompt[idx : idx + 500]
     assert "--- INGREDIENTS TABLE ---" not in section
     assert "ingredients_table" in section
+
+
+def test_food_truck_prompt_has_degraded_response_halt_rule():
+    prompt = _build_food_truck_prompt(
+        recipe="test-recipe",
+        task="Test task",
+        ingredients={},
+        mcp_prefix=DIRECT_PREFIX,
+        dispatch_id="test-dispatch",
+        campaign_id="test-campaign",
+        l3_timeout_sec=300,
+    )
+    assert "FAILURE PREDICATE — DEGRADED RESPONSE" in prompt
+    assert 'reason="degraded_tool_response"' in prompt
+
+
+def test_food_truck_prompt_degraded_rule_triggers_sentinel_on_empty_content():
+    prompt = _build_food_truck_prompt(
+        recipe="test-recipe",
+        task="Test task",
+        ingredients={},
+        mcp_prefix=DIRECT_PREFIX,
+        dispatch_id="test-dispatch",
+        campaign_id="test-campaign",
+        l3_timeout_sec=300,
+    )
+    start = prompt.index("FAILURE PREDICATE — DEGRADED RESPONSE")
+    end = prompt.index("TWO FAILURE TIERS", start)
+    section = prompt[start:end]
+    assert "success" in section
+    assert "false" in section
+    assert any(word in section for word in ("absent", "null", "empty")), (
+        "Rule must reference absent, null, or empty content"
+    )
+
+
+def test_food_truck_prompt_degraded_reason_in_reason_enum():
+    prompt = _build_food_truck_prompt(
+        recipe="test-recipe",
+        task="Test task",
+        ingredients={},
+        mcp_prefix=DIRECT_PREFIX,
+        dispatch_id="test-dispatch",
+        campaign_id="test-campaign",
+        l3_timeout_sec=300,
+    )
+    fields_idx = prompt.index("Fields:", prompt.index("--- SECTION 8:"))
+    reason_region = prompt[fields_idx : fields_idx + 300]
+    assert "degraded_tool_response" in reason_region, (
+        "degraded_tool_response must appear in Section 8 reason enum"
+    )
+    ok_idx = prompt.index("FAILURE PREDICATE — open_kitchen")
+    deg_idx = prompt.index("FAILURE PREDICATE — DEGRADED RESPONSE")
+    assert ok_idx != deg_idx, "Degraded and open_kitchen predicates must be distinct blocks"