Pipeline Execution

Architectural Layer: L5/L6 Pipeline
Source Directory: src/pipeline
File Count: 18 source files

End-to-end task pipelines, staging, execution harnesses, and lifecycle transitions.

Module Files & Detailed Reference

Source File	Lines	Key Types	Responsibility
capture.rs	118	OutputCaptureManager	Budget-aware tool output capture and artifact externalization (TL-03, ...
dedup.rs	472	MutationDedupFence, DedupError, FenceVerdict	Durable mutating-tool deduplication fence. Crash-between-action-and-p...
dispatcher.rs	585	ProductionActionDispatcher	Production action dispatcher wiring WorkerRunner to the 11-stage pipel...
error.rs	267	ToolError, ToolErrorCategory	Strongly typed model-facing tool error taxonomy with bounded correctio...
mod.rs	21	Functions/Traits	Non-bypassable 11-stage tool execution pipeline, output budgeting, and...
runner.rs	344	ToolPipelineRunner	Non-bypassable 11-stage tool execution pipeline runner (TL-03, per D-0...
approval_resolution.rs	379	ApprovalResolutionStage, ExecutionAuthorizedState	Subsystem component
capability_check.rs	138	CapabilityAuthorizedState, CapabilityCheckStage	Stage 5: Capability Check Stage (TL-03, per D-09).
capture_normalization.rs	300	AuditDigest, CaptureNormalizationStage, DiagnosticOutput	Stage 10: Capture & Normalization Stage (TL-03, per D-09, D-11).
decoding.rs	80	ArgDecodingStage, DecodedArgsState	Stage 2: Argument Decoding Stage (TL-03, per D-09).
execution.rs	68	RawExecutionState, ToolExecutionStage	Stage 9: Tool Execution Stage (TL-03, per D-09). Non-negotiable secur...
mod.rs	42	Functions/Traits	The 11 non-bypassable typed state-machine stages of the tool execution...
policy_gate.rs	145	PolicyEvaluatedState, PolicyGateStage	Stage 7: PolicyGate Evaluation Stage (TL-03, per D-09).
resolution.rs	41	ResolvedToolState, ToolResolutionStage	Stage 1: Tool Resolution Stage (TL-03, per D-09).
resource_scope.rs	475	ResourceBinding, ResourceScopeStage, ResourceScopedState	Stage 6: Resource Scope & Lease Validation Stage (TL-03, per D-09)...
schema_validation.rs	207	SchemaValidatedState, SchemaValidationStage	Stage 3: Schema Validation Stage (TL-03, per D-09).
semantic_validation.rs	92	SemanticValidationStage, SemanticallyValidatedState	Stage 4: Semantic Validation Stage (TL-03, per D-09).
telemetry_record.rs	60	TelemetryDurableRecordStage	Stage 11: Telemetry & Durable Record Stage (TL-03, per D-09, D-12)...

---

src/pipeline/capture.rs

• Source Link: src/pipeline/capture.rs on GitHub
• Lines of Code: 118 (4397 bytes)

Primary Responsibility

Budget-aware tool output capture and artifact externalization (TL-03, per D-11).

Types & Trait Declarations

• Structs (1): OutputCaptureManager

Key Functions & Signatures (5 shown)

// private
fn default() -> Self;
// pub
fn new(max_inline_bytes: usize, max_inline_lines: usize) -> Self;
// pub
fn max_inline_bytes(&self) -> usize;
// pub
fn max_inline_lines(&self) -> usize;
// pub
async fn process_output(&self,
        raw_output: &str,
        artifact_store: Option<&Arc<dyn ArtifactStore>>,) -> Re;

---

src/pipeline/dedup.rs

• Source Link: src/pipeline/dedup.rs on GitHub
• Lines of Code: 472 (18024 bytes)
• Runtime Properties: 💾 Interacts with SQLite Persistence

Primary Responsibility

Durable mutating-tool deduplication fence. Crash-between-action-and-persistence replays an already-committed mutation as a brand-new side effect (at-least-once). This module provides the durable identity that closes the gap where safe: text same committed mutation (task + tool + canonical input hash) → recognized from the fence → no duplicate side effect Identity is semantic, never timestamp-only. Outcomes distinguish committed success (suppress duplicates) from failed attempts (legitimate retry still proceeds). Recovery is never made impossible: only a recorded success suppresses re-execution. Suppression applies exclusively to the idempotent file-mutation allowlist ([is_fenced_mutating_tool]). Non-idempotent tools (process execution, test runs, network) always execute; their outcomes are still recorded for audit. Suppressing a repeated cargo test would break legitimate verify-after-fix flows, so commands are fenced by record, not by skip.

Types & Trait Declarations

• Structs (1): MutationDedupFence
• Enums (2): DedupError, FenceVerdict

Key Functions & Signatures (11 shown)

// pub
fn is_fenced_mutating_tool(tool_name: &str) -> bool;
// pub
fn canonical_json(value: &serde_json::Value) -> String;
// pub
fn fingerprint_mutation(task_id: &TaskId,
    tool_name: &str,
    params: &serde_json::Value,) -> String;
// pub
fn fingerprint_parts(parts: &[&[u8]]) -> String;
// pub
fn sha_hex(data: &[u8]) -> String;
// pub
fn new(pool: SqlitePool) -> Self;
// pub
async fn check(&self,
        task_id: &TaskId,
        tool_name: &str,
        fingerprint: &str,) -> FenceVerdict;
// pub
async fn acquire_claim(&self,
        task_id: &TaskId,
        tool_name: &str,
        fingerprint: &str,
        stal;
// pub
async fn commit_outcome(&self,
        task_id: &TaskId,
        tool_name: &str,
        fingerprint: &str,
        suc;
// pub
async fn record(&self,
        task_id: &TaskId,
        tool_name: &str,
        fingerprint: &str,
        success: bo;
// pub
async fn recorded_outcome(&self,
        task_id: &TaskId,
        tool_name: &str,
        fingerprint: &str,) -> Optio;

---

src/pipeline/dispatcher.rs

• Source Link: src/pipeline/dispatcher.rs on GitHub
• Lines of Code: 585 (26707 bytes)
• Runtime Properties: 💾 Interacts with SQLite Persistence | 🛡️ Enforces Policy Gate | 🛑 Supports Cancellation Tokens

Primary Responsibility

Production action dispatcher wiring WorkerRunner to the 11-stage pipeline (TL-03, D-12).

Types & Trait Declarations

• Structs (1): ProductionActionDispatcher

Key Functions & Signatures (12 shown)

// pub
fn new(runner: Arc<ToolPipelineRunner>,
        context: ToolExecutionContext,
        policy_gate: Arc<dyn PolicyGate>,;
// pub
fn with_session_repo(mut self,
        repo: crate::interaction::session::SqliteSessionRepository,) -> Self;
// pub
fn with_session_id(mut self, id: crate::ids::SessionId) -> Self;
// pub
fn with_event_bus(mut self, bus: Arc<dyn crate::events::EventBus>) -> Self;
// pub
fn with_replan_authority(mut self,
        replan_authority: Arc<crate::recovery::ReplanAuthority>,) -> Self;
// pub
fn replan_authority(&self) -> Option<&Arc<crate::recovery::ReplanAuthority>>;
// pub
fn runner(&self) -> &Arc<ToolPipelineRunner>;
// pub
fn context(&self) -> &ToolExecutionContext;
// pub
fn policy_gate(&self) -> &Arc<dyn PolicyGate>;
// pub
fn autonomy_mode(&self) -> AutonomyMode;
// private
async fn dispatch(&self, action: &ActionRequest) -> Result<ActionResult, String>;
// private
async fn dispatch_action(&self,
        action: &crate::agent::action::AgentAction,) -> Result<crate::agent::action::Age;

---

src/pipeline/error.rs

• Source Link: src/pipeline/error.rs on GitHub
• Lines of Code: 267 (8296 bytes)

Primary Responsibility

Strongly typed model-facing tool error taxonomy with bounded correction hints (TL-04, per D-10).

Types & Trait Declarations

• Structs (1): ToolError
• Enums (1): ToolErrorCategory

Key Functions & Signatures (14 shown)

// pub
fn as_str(&self) -> &'static str;
// pub
fn new(category: ToolErrorCategory,
        code: impl Into<String>,
        message: impl Into<String>,
        provena;
// pub
fn validation(code: impl Into<String>,
        message: impl Into<String>,
        hint: Option<String>,) -> Self;
// pub
fn not_found(code: impl Into<String>,
        message: impl Into<String>,
        hint: Option<String>,) -> Self;
// pub
fn precondition_failed(code: impl Into<String>,
        message: impl Into<String>,
        hint: Option<String>,) -> Se;
// pub
fn permission_denied(code: impl Into<String>,
        message: impl Into<String>,
        hint: Option<String>,) -> Self;
// pub
fn execution_failed(code: impl Into<String>,
        message: impl Into<String>,
        hint: Option<String>,) -> Self;
// pub
fn resource_exhausted(code: impl Into<String>,
        message: impl Into<String>,
        hint: Option<String>,) -> Sel;
// pub
fn with_provenance(mut self, provenance: impl Into<String>) -> Self;
// pub
fn with_retryable(mut self, retryable: bool) -> Self;
// pub
fn to_model_diagnostic(&self) -> String;
// private
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result;
// pub
fn sanitize_hint(hint: &str) -> String;
// private
fn from(err: crate::tools::error::ToolError) -> Self;

---

src/pipeline/mod.rs

• Source Link: src/pipeline/mod.rs on GitHub
• Lines of Code: 21 (842 bytes)

Primary Responsibility

Non-bypassable 11-stage tool execution pipeline, output budgeting, and typed error taxonomy (TL-03, TL-04). Enforces compile-time and runtime guarantees: - 11 linear state-machine stages (stages) - Physical side effects occur strictly in Stage 9 (execution) - Budget-aware output capture and artifact externalization (capture) - Strongly typed model-facing error taxonomy with bounded hints (error) - Production dispatcher connecting WorkerRunner to the pipeline (dispatcher)

---

src/pipeline/runner.rs

• Source Link: src/pipeline/runner.rs on GitHub
• Lines of Code: 344 (13936 bytes)
• Runtime Properties: 💾 Interacts with SQLite Persistence | 🛡️ Enforces Policy Gate

Primary Responsibility

Non-bypassable 11-stage tool execution pipeline runner (TL-03, per D-09, D-12).

Types & Trait Declarations

• Structs (1): ToolPipelineRunner

Key Functions & Signatures (11 shown)

// pub
fn new(tool_registry: Arc<ToolRegistry>) -> Self;
// pub
fn with_db_pool(mut self, pool: SqlitePool) -> Self;
// pub
fn with_approval_coordinator(mut self, coordinator: Arc<ApprovalCoordinator>) -> Self;
// pub
fn with_artifact_store(mut self, store: Arc<dyn ArtifactStore>) -> Self;
// pub
fn with_capture_manager(mut self, capture_manager: Arc<OutputCaptureManager>) -> Self;
// pub
fn tool_registry(&self) -> &Arc<ToolRegistry>;
// pub
fn capture_manager(&self) -> &Arc<OutputCaptureManager>;
// pub
fn artifact_store(&self) -> Option<&Arc<dyn ArtifactStore>>;
// pub
fn approval_coordinator(&self) -> Option<&Arc<ApprovalCoordinator>>;
// pub
fn db_pool(&self) -> Option<&SqlitePool>;
// pub
async fn execute_action(&self,
        action: &ActionRequest,
        context: &ToolExecutionContext,
        policy_ga;

---

src/pipeline/stages/approval_resolution.rs

• Source Link: src/pipeline/stages/approval_resolution.rs on GitHub
• Lines of Code: 379 (18956 bytes)
• Runtime Properties: 💾 Interacts with SQLite Persistence | 🛡️ Enforces Policy Gate

Primary Responsibility

Internal implementation file.

Types & Trait Declarations

• Structs (2): ApprovalResolutionStage, ExecutionAuthorizedState

Key Functions & Signatures (3 shown)

// pub
fn execute(state: PolicyEvaluatedState,
        autonomy_mode: AutonomyMode,) -> Result<ExecutionAuthorizedState, ToolEr;
// pub
fn execute_sync(state: PolicyEvaluatedState,
        autonomy_mode: AutonomyMode,) -> Result<ExecutionAuthorizedState, T;
// pub
async fn execute_async(state: PolicyEvaluatedState,
        context: &ToolExecutionContext,
        autonomy_mode: Auton;

---

src/pipeline/stages/capability_check.rs

• Source Link: src/pipeline/stages/capability_check.rs on GitHub
• Lines of Code: 138 (6361 bytes)
• Runtime Properties: 🛡️ Enforces Policy Gate

Primary Responsibility

Stage 5: Capability Check Stage (TL-03, per D-09).

Types & Trait Declarations

• Structs (2): CapabilityAuthorizedState, CapabilityCheckStage

Key Functions & Signatures (1 shown)

// pub
fn execute(state: SemanticallyValidatedState,
        context: &ToolExecutionContext,) -> Result<CapabilityAuthorizedSta;

---

src/pipeline/stages/capture_normalization.rs

• Source Link: src/pipeline/stages/capture_normalization.rs on GitHub
• Lines of Code: 300 (10521 bytes)

Primary Responsibility

Stage 10: Capture & Normalization Stage (TL-03, per D-09, D-11).

Types & Trait Declarations

• Structs (7): AuditDigest, CaptureNormalizationStage, DiagnosticOutput, ModelVisibleOutput, NormalizedResultState, PipelineOutputEvidence, RawExecutionEvidence

Key Functions & Signatures (15 shown)

// pub
fn new(raw: impl Into<String>) -> Self;
// pub
fn classify(self,
        model_visible_hint: String,) -> (ModelVisibleOutput, DiagnosticOutput, AuditDigest);
// pub
fn expose_privileged(&self) -> &str;
// pub
fn as_str(&self) -> &str;
// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// pub
fn as_str(&self) -> &str;
// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// pub
fn as_str(&self) -> &str;
// pub
fn empty_digest() -> Self;
// private
fn hex_digest(raw: &str) -> String;
// pub
fn new(raw: String, model_visible: String) -> Self;
// pub
fn empty() -> Self;
// pub
fn raw_evidence(&self) -> RawExecutionEvidence;
// pub
fn model_visible(&self) -> ModelVisibleOutput;
// pub
fn diagnostic(&self) -> DiagnosticOutput;

Unit Verification (2 inline tests)

Tests defined within module: test_pipeline_output_evidence_classifications, test_pipeline_output_evidence_empty

---

src/pipeline/stages/decoding.rs

• Source Link: src/pipeline/stages/decoding.rs on GitHub
• Lines of Code: 80 (2935 bytes)

Primary Responsibility

Stage 2: Argument Decoding Stage (TL-03, per D-09).

Types & Trait Declarations

• Structs (2): ArgDecodingStage, DecodedArgsState

Key Functions & Signatures (1 shown)

// pub
fn execute(state: ResolvedToolState) -> Result<DecodedArgsState, ToolError>;

---

src/pipeline/stages/execution.rs

• Source Link: src/pipeline/stages/execution.rs on GitHub
• Lines of Code: 68 (2295 bytes)
• Runtime Properties: 🛡️ Enforces Policy Gate

Primary Responsibility

Stage 9: Tool Execution Stage (TL-03, per D-09). Non-negotiable security invariant: Side-effects occur STRICTLY in this stage. Stages 1–8 enforce structural, semantic, capability, resource, and policy authorization without mutations.

Types & Trait Declarations

• Structs (2): RawExecutionState, ToolExecutionStage

Key Functions & Signatures (1 shown)

// pub
async fn execute(state: ExecutionAuthorizedState,
        context: &ToolExecutionContext,) -> RawExecutionState;

---

src/pipeline/stages/mod.rs

• Source Link: src/pipeline/stages/mod.rs on GitHub
• Lines of Code: 42 (2069 bytes)
• Runtime Properties: 🛡️ Enforces Policy Gate

Primary Responsibility

The 11 non-bypassable typed state-machine stages of the tool execution pipeline (TL-03, D-09). State transitions: ActionRequest -> Stage 1: ToolResolutionStage (ResolvedToolState) -> Stage 2: ArgDecodingStage (DecodedArgsState) -> Stage 3: SchemaValidationStage (SchemaValidatedState) -> Stage 4: SemanticValidationStage (SemanticallyValidatedState) -> Stage 5: CapabilityCheckStage (CapabilityAuthorizedState) -> Stage 6: ResourceScopeStage (ResourceScopedState) -> Stage 7: PolicyGateStage (PolicyEvaluatedState) -> Stage 8: ApprovalResolutionStage (ExecutionAuthorizedState) -> Stage 9: ToolExecutionStage (RawExecutionState) [MUTATING SIDE-EFFECTS HERE ONLY] -> Stage 10: CaptureNormalizationStage (NormalizedResultState) -> Stage 11: TelemetryDurableRecordStage (ActionResult)

---

src/pipeline/stages/policy_gate.rs

• Source Link: src/pipeline/stages/policy_gate.rs on GitHub
• Lines of Code: 145 (5524 bytes)
• Runtime Properties: 💾 Interacts with SQLite Persistence | 🛡️ Enforces Policy Gate

Primary Responsibility

Stage 7: PolicyGate Evaluation Stage (TL-03, per D-09).

Types & Trait Declarations

• Structs (2): PolicyEvaluatedState, PolicyGateStage

Key Functions & Signatures (2 shown)

// pub
async fn execute(state: ResourceScopedState,
        context: &ToolExecutionContext,
        policy_gate: &dyn PolicyGat;
// pub
async fn execute_with_pool(state: ResourceScopedState,
        context: &ToolExecutionContext,
        policy_gate: &dyn;

---

src/pipeline/stages/resolution.rs

• Source Link: src/pipeline/stages/resolution.rs on GitHub
• Lines of Code: 41 (1237 bytes)

Primary Responsibility

Stage 1: Tool Resolution Stage (TL-03, per D-09).

Types & Trait Declarations

• Structs (2): ResolvedToolState, ToolResolutionStage

Key Functions & Signatures (1 shown)

// pub
fn execute(action: &ActionRequest,
        registry: &Arc<ToolRegistry>,) -> Result<ResolvedToolState, ToolError>;

---

src/pipeline/stages/resource_scope.rs

• Source Link: src/pipeline/stages/resource_scope.rs on GitHub
• Lines of Code: 475 (21741 bytes)
• Runtime Properties: 🛡️ Enforces Policy Gate

Primary Responsibility

Stage 6: Resource Scope & Lease Validation Stage (TL-03, per D-09).

Types & Trait Declarations

• Structs (3): ResourceBinding, ResourceScopeStage, ResourceScopedState

Key Functions & Signatures (3 shown)

// pub
fn execute(state: CapabilityAuthorizedState,
        context: &ToolExecutionContext,) -> Result<ResourceScopedState, Too;
// pub
fn verify_bindings_live(bound_resources: &[ResourceBinding],
        workspace_root: &Path,) -> Result<(), ToolError>;
// private
fn normalize_path(path: &Path) -> PathBuf;

---

src/pipeline/stages/schema_validation.rs

• Source Link: src/pipeline/stages/schema_validation.rs on GitHub
• Lines of Code: 207 (7907 bytes)

Primary Responsibility

Stage 3: Schema Validation Stage (TL-03, per D-09).

Types & Trait Declarations

• Structs (2): SchemaValidatedState, SchemaValidationStage

Key Functions & Signatures (4 shown)

// pub
fn execute(state: DecodedArgsState) -> Result<SchemaValidatedState, ToolError>;
// private
fn coerce_value_to_schema_type(val: &mut serde_json::Value, schema: &serde_json::Value);
// private
fn matches_schema_type(val: &serde_json::Value, schema: &serde_json::Value) -> bool;
// private
fn check_single_type(val: &serde_json::Value, expected: &str) -> bool;

---

src/pipeline/stages/semantic_validation.rs

• Source Link: src/pipeline/stages/semantic_validation.rs on GitHub
• Lines of Code: 92 (3699 bytes)

Primary Responsibility

Stage 4: Semantic Validation Stage (TL-03, per D-09).

Types & Trait Declarations

• Structs (2): SemanticValidationStage, SemanticallyValidatedState

Key Functions & Signatures (1 shown)

// pub
fn execute(state: SchemaValidatedState) -> Result<SemanticallyValidatedState, ToolError>;

---

src/pipeline/stages/telemetry_record.rs

• Source Link: src/pipeline/stages/telemetry_record.rs on GitHub
• Lines of Code: 60 (2273 bytes)

Primary Responsibility

Stage 11: Telemetry & Durable Record Stage (TL-03, per D-09, D-12).

Types & Trait Declarations

• Structs (1): TelemetryDurableRecordStage

Key Functions & Signatures (2 shown)

// pub
fn execute(state: NormalizedResultState) -> ActionResult;
// pub
fn record_failure(action_id: String, tool_id: String, error: ToolError) -> ActionResult;

---