Dynamic Re-Decomposition Prompt for Failed Sub-Tasks

The original sub-task contract that failed, including its input schema, expected output schema, assigned agent role, and completion criteria

{"task_id": "extract-invoice-line-items", "agent_role": "document-extractor", "input_schema": {"document_id": "string"}, "output_schema": {"line_items": "array"}, "completion_criteria": "all line items extracted with confidence >= 0.9"}

Must be a valid JSON object with task_id, agent_role, input_schema, output_schema, and completion_criteria fields. Reject if any field is null or missing

Structured description of why the sub-task failed, including error type, failure stage, and any partial output or error messages produced

{"error_type": "tool_timeout", "failure_stage": "extraction", "error_message": "PDF parsing timed out after 30s on page 4", "partial_output": {"line_items": [{"item": "Widget A", "confidence": 0.95}]}}

Must contain error_type, failure_stage, and error_message. partial_output may be null. Validate error_type against known failure taxonomy: tool_timeout, schema_violation, confidence_below_threshold, agent_unavailable, dependency_failure, context_overflow

Array of all successfully completed sub-tasks with their outputs, preserving work that must not be re-executed or invalidated by the new plan

[{"task_id": "classify-document-type", "output": {"document_type": "invoice", "confidence": 0.97}, "completed_at": "2025-01-15T10:30:00Z"}, {"task_id": "extract-header-fields", "output": {"vendor": "Acme Corp", "invoice_date": "2025-01-10"}, "completed_at": "2025-01-15T10:31:00Z"}]

Must be an array. Each entry requires task_id, output, and completed_at. Validate that no task_id duplicates exist. Check that completed_at timestamps are monotonically increasing. Reject if any output field is missing

Map of sub-tasks that are waiting on the failed task's output, including what specific fields they require and their current blocking status

{"validate-invoice-totals": {"required_fields": ["line_items"], "blocking_status": "blocked", "timeout_at": "2025-01-15T10:35:00Z"}, "generate-report": {"required_fields": ["line_items", "vendor"], "blocking_status": "partially_blocked"}}

Must be a JSON object keyed by dependent task_id. Each value requires required_fields array and blocking_status. Validate blocking_status against enum: blocked, partially_blocked, at_risk. Check that required_fields references actual fields from the failed task's output_schema

Current state of available agents including their capabilities, current load, health status, and cost tier for re-assignment decisions

{"available_agents": [{"agent_id": "doc-extractor-v2", "capabilities": ["pdf-extraction", "table-parsing"], "status": "healthy", "current_load": 2, "cost_tier": "standard"}, {"agent_id": "doc-extractor-v3", "capabilities": ["pdf-extraction", "table-parsing", "ocr"], "status": "healthy", "current_load": 0, "cost_tier": "premium"}]}

Must include at least one available agent. Each agent requires agent_id, capabilities array, status, current_load, and cost_tier. Validate status against enum: healthy, degraded, overloaded. Reject if no agent has capabilities matching the failed task's requirements

Remaining token budget for the overall orchestration, constraining how much new decomposition overhead is acceptable

{"total_budget": 128000, "consumed": 89000, "remaining": 39000, "per_task_limit": 8000, "overhead_reserved": 5000}

Must include total_budget, consumed, remaining, per_task_limit, and overhead_reserved. Validate that remaining >= overhead_reserved + (per_task_limit * number_of_replacement_tasks). Reject if remaining < overhead_reserved

Rules defining when re-decomposition should be abandoned in favor of human escalation, including max retry count, time budget, and risk thresholds

{"max_recomposition_attempts": 2, "current_attempt": 1, "max_total_latency_ms": 60000, "elapsed_ms": 35000, "risk_threshold": "medium", "escalation_target": "human-ops-queue"}

Must include max_recomposition_attempts, current_attempt, max_total_latency_ms, elapsed_ms, risk_threshold, and escalation_target. Validate that current_attempt < max_recomposition_attempts. Reject if elapsed_ms >= max_total_latency_ms. Validate risk_threshold against enum: low, medium, high, critical

The original user request that initiated the full task decomposition, preserved for intent alignment in the re-decomposition

"Extract all line items from this invoice PDF and validate the totals against the header amounts. Flag any discrepancies over $10."

Must be a non-empty string. Validate that the original intent is still achievable given the failure diagnostic. Reject if the original request has been invalidated by upstream state changes

string (UUID v4)

Must be a new UUID distinct from the original plan_id. Parse check: valid UUID v4 format.

string (UUID v4)

Must match the plan_id of the failed decomposition. Parse check: exact match against orchestrator state.

string (UUID v4)

Must reference a sub-task that exists in the original plan and is marked as FAILED. Parse check: membership in original task list with status=FAILED.

string (enum)

Must be one of: CONTRACT_VIOLATION, TIMEOUT, AGENT_UNAVAILABLE, TOOL_FAILURE, INVALID_OUTPUT, DEPENDENCY_FAILURE, POLICY_REFUSAL. Enum check: exact match.

array of UUID strings

Every UUID must reference a sub-task in the original plan with status=COMPLETED. No UUID may appear in the replacement_tasks array. Schema check: array of strings, no duplicates, all valid UUIDs.

array of objects

Must contain at least 1 task object. Each object must include task_id (new UUID), agent_role, input_contract, output_contract, and depends_on fields. Schema check: valid JSON Schema against replacement_task_schema.

object

Keys are original downstream task UUIDs that depended on the failed task. Values are arrays of replacement task UUIDs that now satisfy the dependency. Schema check: all keys and values must be valid UUIDs; no orphaned dependencies allowed.

string (max 500 chars)

Must describe what intermediate state from completed tasks is preserved and how replacement tasks consume it. Null not allowed. Length check: 1-500 characters.

number (0.0-1.0)

Orchestrator's confidence that the revised plan will succeed. Must be a float between 0.0 and 1.0 inclusive. Threshold check: if below 0.6, human approval is required before dispatch.

boolean

Must be true if confidence_score < 0.6 or if any replacement task involves a regulated action. Approval check: if true, plan must not be dispatched until approval is logged.

All previously completed sub-tasks are explicitly marked as DONE and excluded from the new execution plan

Completed sub-task appears in the new task list or its output is overwritten

Diff the input [COMPLETED_TASKS] list against the output task list; assert zero overlap

Every replacement sub-task lists only valid upstream dependencies that exist in the revised plan or are already completed

Replacement sub-task references a non-existent, failed, or circular dependency

Parse the dependency graph from the output; assert all referenced task IDs exist and no cycles are detected

Replacement sub-task contracts include all required inputs from preserved outputs and exclude inputs that were only produced by the failed task

Replacement contract requires an output field that no upstream task produces

Schema-diff each replacement contract's required inputs against available outputs from upstream tasks

The revised plan does not introduce conflicting writes to shared state or duplicate side effects already performed by completed tasks

Two tasks in the revised plan write to the same resource without a defined merge rule

Scan output for shared resource identifiers; assert each resource has at most one writer or an explicit merge task

The re-decomposition explicitly addresses the reported failure reason and adjusts the approach, not just retries the same task with the same contract

Replacement sub-task has identical description, inputs, and agent assignment as the failed task

Semantic similarity check between failed task definition and replacement task definition; assert cosine similarity below 0.9

If the failure reason matches an escalation rule in [ESCALATION_POLICY], the output routes to human review instead of re-decomposing

Re-decomposition is produced for a failure that policy marks as non-retryable

Check failure reason against escalation policy regex patterns; assert no re-decomposition when escalation is required

The revised plan conforms exactly to [OUTPUT_SCHEMA] with all required fields present and correctly typed

Missing required field, wrong type, or extra undeclared fields in the output

Validate output against the JSON Schema provided in [OUTPUT_SCHEMA]; assert zero validation errors

The revised plan includes a new plan version or idempotency key distinct from the original plan to prevent duplicate execution

Revised plan reuses the same plan ID as the original or omits a version marker

Assert output.plan_id != input.original_plan_id and output.plan_version > input.original_plan_version