Fix missing recipe digests and checkpoint carrier delivery

docs/decisions/0004-recipe-redelivery.md

@@ -60,7 +60,8 @@ snapshot, while the public `open_kitchen` response contains only the segment ove
 initial bodies. Each mapped checkpoint tool verifies READY against that exact durable
 generation before its effect and then returns `recipe_segment` with either authenticated
 future bodies or an ordered manual `recipe_pull` closure. No raw YAML, LRU lookup, runtime
-skip carrier, or separately persisted projection participates in recovery.
+skip carrier, or separately persisted projection participates in recovery. Carriers are
+attached by the shared response wrapper from each tool's registry outcome declaration.
 
 Acknowledged `run_skill` receipts are replayable only for the same kitchen, request
 session, and receipt identity. Their tracker effect remains one-shot and cached, so a
@@ -127,8 +128,10 @@ Four invariants are non-negotiable:
   for envelope responses, `get_recipe_section`).
 - Terminal-page delivery enforces the four invariants above while preserving the
   separate completion call as a compatibility fallback for non-receipt responses.
-- Segmented recipes recover from the latest `recipe_segment`; the startup credential is
-  intentionally scoped to delivered bodies rather than the full future recipe.
+- Segmented recipes recover from the latest `recipe_segment`; the startup carrier's
+  credential is scoped to the first segment's bodies by the carrier byte cap, while the
+  completion receipt (invariant 4) always carries the complete credential and is
+  re-requestable while the recipe is READY.
 - The delivery-reachability suite verifies `load_recipe` / `open_kitchen` /
   `get_recipe_section` re-delivery restores pipeline execution capability, including
   complete-envelope and bounded-envelope per-step pulls.

src/autoskillit/cli/prompts/_prompts.py

@@ -178,12 +178,19 @@ def render(self) -> str:
                 "recipe_delivery.payload_metadata when the body is attested); the same "
                 "forwarding rule applies. If any response contains recipe_segment, "
                 "consume it as mandatory authority: verify and install ordered bodies "
-                "for startup/success carriers, use the carrier-scoped invocation "
-                "digests, and reconstruct every recovery pull_closure entry through "
+                "for startup/success carriers and reconstruct every recovery "
+                "pull_closure entry through "
                 "the unchanged recipe_pull. The overview is only a table of contents. "
                 "If subtype=recipe_segment_post_effect_delivery_failure, the operation "
-                "already ran; do not repeat it. Do not assume startup contains "
-                "full-horizon invocation_template_digests. For structured child inputs, "
+                "already ran; do not repeat it. "
+                "The complete_recipe_initialization(initialization_id) receipt carries every "
+                "named step's recipe_execution credential. Use the delivered completion receipt "
+                "when one is already present, including a terminal-page embedded receipt; "
+                "otherwise call complete_recipe_initialization(initialization_id) before the "
+                "first run_skill. The receipt can be re-requested at any time while the recipe "
+                "is READY, and a digest rejection names that exact call. Absence of a top-level "
+                "recipe_execution block alone does not identify segmented delivery. "
+                "For structured child inputs, "
                 "select "
                 "recipe_execution.skill_input_shapes[step_name], initialize skill_inputs "
                 "with exactly its ordered keys, and replace available values in place. "

src/autoskillit/core/__init__.pyi

@@ -1205,6 +1205,7 @@ from .types import RecipeDeliveryBudgetError as RecipeDeliveryBudgetError
 from .types import RecipeDeliveryDecision as RecipeDeliveryDecision
 from .types import RecipeDeliveryEvidenceDef as RecipeDeliveryEvidenceDef
 from .types import RecipeDeliveryMode as RecipeDeliveryMode
+from .types import RecipeDeliveryOutcomeDef as RecipeDeliveryOutcomeDef
 from .types import RecipeDeliveryRequest as RecipeDeliveryRequest
 from .types import RecipeDeliverySurfaceDef as RecipeDeliverySurfaceDef
 from .types import RecipeExecutionCredential as RecipeExecutionCredential

src/autoskillit/core/_tool_registry_builders.py

@@ -6,6 +6,7 @@
 from types import MappingProxyType
 
 from .types.recipe._type_recipe_binding import (
+    RecipeDeliveryOutcomeDef,
     ToolDef,
     ToolInitializationOperation,
     ToolParamDef,
@@ -139,6 +140,38 @@ def _initialization_operation(name: str) -> ToolInitializationOperation:
     )
 
 
+_RECIPE_DELIVERY_OUTCOMES: Mapping[str, RecipeDeliveryOutcomeDef] = MappingProxyType(
+    {
+        "check_repo_merge_state": RecipeDeliveryOutcomeDef(),
+        "enqueue_pr": RecipeDeliveryOutcomeDef(success_field="success"),
+        "wait_for_merge_queue": RecipeDeliveryOutcomeDef(success_field="success"),
+        "wait_for_ci": RecipeDeliveryOutcomeDef(
+            success_field="conclusion",
+            success_value="success",
+        ),
+        "push_to_remote": RecipeDeliveryOutcomeDef(success_field="success"),
+        "bootstrap_clone": RecipeDeliveryOutcomeDef(),
+        "run_python": RecipeDeliveryOutcomeDef(success_field="success"),
+        "complete_run_skill_result": RecipeDeliveryOutcomeDef(
+            success_field="run_skill_success",
+            step_name_field="step_name",
+        ),
+        "merge_worktree": RecipeDeliveryOutcomeDef(success_field="merge_succeeded"),
+        "check_pr_mergeable": RecipeDeliveryOutcomeDef(),
+        "test_check": RecipeDeliveryOutcomeDef(success_field="passed"),
+        "bind_plan_set": RecipeDeliveryOutcomeDef(success_field="success"),
+        "run_cmd": RecipeDeliveryOutcomeDef(success_field="success"),
+        "create_and_publish_branch": RecipeDeliveryOutcomeDef(),
+        "claim_and_resolve_issue": RecipeDeliveryOutcomeDef(success_field="claimed"),
+        "release_issue": RecipeDeliveryOutcomeDef(success_field="success"),
+        "verify_review_receipt": RecipeDeliveryOutcomeDef(
+            success_field="reviews_posted",
+            success_value="true",
+        ),
+    }
+)
+
+
 def _tool(
     name: str,
     params: tuple[str, ...] = (),
@@ -171,6 +204,7 @@ def _tool(
         initialization_operation=_initialization_operation(name),
         automatic_recipe_delivery=automatic_recipe_delivery,
         recovery_recipe_delivery=recovery_recipe_delivery,
+        recipe_delivery_outcome=_RECIPE_DELIVERY_OUTCOMES.get(name),
     )
 
 

src/autoskillit/core/types/constants/_type_constants.py

@@ -398,8 +398,8 @@ def _render_stop_step_evidence_doctrine(indent: str = "") -> str:
 RECIPE_EXECUTION_ATTESTATION_MISSING_MESSAGE: str = (
     "an active recipe requires recipe_execution_id and invocation_template_digest; "
     "take both from the recipe_execution block of the complete_recipe_initialization "
-    "receipt (bounded delivery) or of the open_kitchen response (inline delivery), "
-    "using invocation_template_digests[step_name] for this step; structured calls must "
+    "receipt, which carries every named step's credential, using "
+    "invocation_template_digests[step_name] for this step; structured calls must "
     "initialize skill_inputs from skill_input_shapes[step_name] ordered keys, replace "
     "available values in place, copy only advertised absence_values by key presence "
     'so "", 0, and False remain verbatim, and never delete or invent a key'

src/autoskillit/core/types/recipe/_type_recipe_binding.py

@@ -35,6 +35,7 @@
     "RecipeFlowEdge",
     "RUNTIME_ADMISSION_BY_ROLE",
     "RuntimeAdmission",
+    "RecipeDeliveryOutcomeDef",
     "ToolDef",
     "ToolInitializationOperation",
     "ToolParamDef",
@@ -161,6 +162,23 @@ class ToolParamDef:
     role: ToolParamRole = field(kw_only=True)
 
 
+@dataclass(frozen=True, slots=True)
+class RecipeDeliveryOutcomeDef:
+    """How the response wrapper reads a delivery tool's outcome from its response."""
+
+    success_field: str | None = None
+    success_value: bool | str = True
+    step_name_field: str | None = None
+
+    def succeeded(self, payload: Mapping[str, object]) -> bool:
+        if self.success_field is None:
+            return "error" not in payload and payload.get("success") is not False
+        actual = payload.get(self.success_field)
+        if isinstance(self.success_value, bool):
+            return actual is self.success_value
+        return actual == self.success_value
+
+
 @dataclass(frozen=True, slots=True)
 class ToolDef:
     """Static definition of one registered MCP tool."""
@@ -172,6 +190,7 @@ class ToolDef:
     automatic_recipe_delivery: bool = False
     # Failure edges may carry a recovery segment; this can be enabled independently.
     recovery_recipe_delivery: bool = False
+    recipe_delivery_outcome: RecipeDeliveryOutcomeDef | None = None
 
     def __post_init__(self) -> None:
         if not isinstance(self.initialization_operation, ToolInitializationOperation):
@@ -195,6 +214,29 @@ def __post_init__(self) -> None:
             )
         if structured and self.name != "run_skill":
             raise ValueError("Only run_skill may declare structured skill_inputs")
+        self._validate_recipe_delivery_outcome()
+
+    def _validate_recipe_delivery_outcome(self) -> None:
+        delivery_enabled = self.automatic_recipe_delivery or self.recovery_recipe_delivery
+        if not delivery_enabled:
+            if self.recipe_delivery_outcome is not None:
+                raise ValueError(
+                    f"ToolDef {self.name!r} declares a recipe delivery outcome "
+                    "without a delivery flag"
+                )
+            return
+        if self.recipe_delivery_outcome is None:
+            raise ValueError(
+                f"ToolDef {self.name!r} enables recipe delivery without an outcome declaration"
+            )
+        if (
+            self.recipe_delivery_outcome.step_name_field is None
+            and "step_name" not in self.param_set
+        ):
+            raise ValueError(
+                f"ToolDef {self.name!r} uses argument-sourced recipe delivery "
+                "without a step_name parameter"
+            )
 
     @property
     def param_names(self) -> tuple[str, ...]:

src/autoskillit/recipes/diagrams/research.md

@@ -1,4 +1,4 @@
-<!-- autoskillit-recipe-hash: sha256:7e7c9adec5fd90bd4086a03207b09949466e883cf04b24183b00f5484324fe85 -->
+<!-- autoskillit-recipe-hash: sha256:7e595fc0971b6e4b1348beaee6fecadfc41a9c251dd74991e2f78660e7418727 -->
 <!-- autoskillit-diagram-format: v7 -->
 
 ## research

src/autoskillit/recipes/research.json

@@ -527,7 +527,7 @@
           "source_doc": "${{ context.experiment_plan }}"
         },
         "cwd": "${{ context.worktree_path }}",
-        "step_name": "decompose",
+        "step_name": "decompose_phases",
         "output_dir": "{{AUTOSKILLIT_TEMP}}/make-groups"
       },
       "capture": {
@@ -1084,7 +1084,7 @@
           "adjust": true
         },
         "cwd": "${{ context.worktree_path }}",
-        "step_name": "adjust",
+        "step_name": "adjust_experiment",
         "output_dir": "."
       },
       "capture": {
@@ -1244,7 +1244,7 @@
           "group_manifest": "${{ context.group_manifest }}"
         },
         "cwd": "${{ context.worktree_path }}",
-        "step_name": "generate_report",
+        "step_name": "generate_report_inconclusive",
         "output_dir": "{{AUTOSKILLIT_TEMP}}/generate-report"
       },
       "capture": {
@@ -1276,7 +1276,7 @@
           "worktree_path": "${{ context.worktree_path }}"
         },
         "cwd": "${{ inputs.source_dir }}",
-        "step_name": "fix",
+        "step_name": "fix_tests",
         "output_dir": "."
       },
       "capture": {

src/autoskillit/recipes/research.yaml

@@ -492,7 +492,7 @@ steps:
       skill_inputs:
         source_doc: ${{ context.experiment_plan }}
       cwd: ${{ context.worktree_path }}
-      step_name: decompose
+      step_name: decompose_phases
       output_dir: '{{AUTOSKILLIT_TEMP}}/make-groups'
     capture:
       group_files: ${{ result.group_files }}
@@ -930,7 +930,7 @@ steps:
         worktree_path: ${{ context.worktree_path }}
         adjust: true
       cwd: ${{ context.worktree_path }}
-      step_name: adjust
+      step_name: adjust_experiment
       output_dir: .
     capture:
       experiment_results: ${{ result.results_path }}
@@ -1060,7 +1060,7 @@ steps:
         classification_timestamp: ${{ context.classification_timestamp }}
         group_manifest: ${{ context.group_manifest }}
       cwd: ${{ context.worktree_path }}
-      step_name: generate_report
+      step_name: generate_report_inconclusive
       output_dir: '{{AUTOSKILLIT_TEMP}}/generate-report'
     capture:
       report_path: ${{ result.report_path }}
@@ -1089,7 +1089,7 @@ steps:
         base_branch: ${{ inputs.base_branch }}
         worktree_path: ${{ context.worktree_path }}
       cwd: ${{ inputs.source_dir }}
-      step_name: fix
+      step_name: fix_tests
       output_dir: .
     capture:
       verdict: ${{ result.verdict }}

src/autoskillit/server/_notify.py

@@ -39,6 +39,10 @@
     complete_section_response,
 )
 from autoskillit.server.recipe._recipe_section_pagination import resolve_recipe_section_bound_bytes
+from autoskillit.server.recipe._recipe_segment_delivery import (
+    attach_checkpoint,
+    prepare_checkpoint,
+)
 from autoskillit.server.response._response_budget import (
     bounded_response_budget_failure,
     enforce_response_budget,
@@ -64,6 +68,31 @@ def _get_ctx_or_none() -> ToolContext | None:
     return _ctx_none_fn()
 
 
+async def _invoke_handler(
+    fn: Callable[..., Awaitable[Any]],
+    args: tuple[Any, ...],
+    kwargs: dict[str, Any],
+    tool_name: str,
+) -> Any:
+    try:
+        return await fn(*args, **kwargs)
+    except Exception as exc:
+        result = json.dumps(
+            {
+                "success": False,
+                "error": f"{type(exc).__name__}: {exc}",
+                "exit_code": -1,
+                "subtype": "tool_exception",
+                "user_visible_message": (
+                    f"An internal error occurred in {tool_name}: "
+                    f"{type(exc).__name__}. Run 'autoskillit doctor' or reinstall."
+                ),
+            }
+        )
+        logger.exception("track_response_size_handler_failed", tool_name=tool_name)
+        return result
+
+
 def _resolve_backend_capabilities(ctx: ToolContext | None) -> BackendCapabilities | None:
     """Extract the resolved backend's capabilities from ``ctx``, if any."""
     backend = getattr(ctx, "backend", None) if ctx is not None else None
@@ -126,22 +155,12 @@ async def get_token_summary(...) -> str:
     def decorator(fn: Callable[..., Awaitable[Any]]) -> Callable[..., Awaitable[Any]]:
         @functools.wraps(fn)
         async def wrapper(*args: Any, **kwargs: Any) -> Any:
-            try:
-                result = await fn(*args, **kwargs)
-            except Exception as exc:
-                result = json.dumps(
-                    {
-                        "success": False,
-                        "error": f"{type(exc).__name__}: {exc}",
-                        "exit_code": -1,
-                        "subtype": "tool_exception",
-                        "user_visible_message": (
-                            f"An internal error occurred in {tool_name}: "
-                            f"{type(exc).__name__}. Run 'autoskillit doctor' or reinstall."
-                        ),
-                    }
-                )
-                logger.exception("track_response_size_handler_failed", tool_name=tool_name)
+            ctx = _get_ctx_or_none()
+            prepared, early = prepare_checkpoint(ctx, fn, args, kwargs)
+            result = (
+                early if early is not None else await _invoke_handler(fn, args, kwargs, tool_name)
+            )
+            result = attach_checkpoint(ctx, fn, result, prepared)
             finalized = (
                 result
                 if isinstance(
@@ -159,7 +178,6 @@ async def wrapper(*args: Any, **kwargs: Any) -> Any:
                 response_value = finalized.rendered
             else:
                 response_value = result
-            ctx = _get_ctx_or_none()
             try:
                 response_str = (
                     response_value

src/autoskillit/server/recipe/AGENTS.md

@@ -9,7 +9,8 @@ Compiled-recipe generation and its persisted artifacts are owned by
 `_recipe_generation` and `_recipe_artifact` respectively. Transactional
 finalization lives in the `_recipe_delivery/` sub-package, with its own
 decision-support helpers in `_recipe_delivery_helpers` and its segmented
-startup/checkpoint carriers in `_recipe_segment_delivery`. Server-owned
+startup/checkpoint carriers in `_recipe_segment_delivery`, which also owns checkpoint
+attachment through `track_response_size`. Server-owned
 execution and initialization commits are split across `_recipe_execution`
 and `_recipe_initialization`. Section pagination is split between the
 page-fitting engine (`_recipe_section_pagination`, `_recipe_section_planning`)