--- id: FAS-PFM-LLM-01 title: Metered AI Requests acceptance feature: FEAT-PFM-LLM-01 product_area: platform module: llm revision: 1 status: proposed product_intent: proposed source_sha: services-platform: ad3f6a6582843008ed0f5a17b131878d15829af1 react-console: baf6dae382d6f41dce976c7f848667ccc37e076f rhythm-v2-docs: 9b95399181c8ec6670ce3d12d700e59643668692 capabilities: - CAP-PFM-LLM-001 - CAP-PFM-LLM-002 - CAP-PFM-LLM-003 - CAP-PFM-LLM-004 scenarios: - id: FAS-PFM-LLM-01-01 title: One task accounts for its model conversations kind: behaviour exercises: - CAP-PFM-LLM-001 - CAP-PFM-LLM-002 - CAP-PFM-LLM-003 - CAP-PFM-LLM-004 relies_on: - ATS-PFM-LLM-001-03 - ATS-PFM-LLM-003-05 - ATS-PFM-LLM-004-01 preconditions: A declared task has three scripted requests and a configured capture store. steps: 'Complete the task and compare request charges, task totals, monthly accounting and capture states.' expected: >- The returned result, task totals and monthly charge agree with the settled requests, whose redacted evidence is retained. negative: false data_needs: >- An isolated workspace with stated usage, callers, clock instants and a separate workspace for scope checks. bound_to: [] provenance: proposed source: - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-001-make-every-ai-request-through-one-metered-path.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-002-keep-an-ai-conversation-within-its-limits.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-003-account-for-a-multi-step-ai-task-as-one-execution.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-004-keep-redacted-evidence-of-every-ai-request.md - id: FAS-PFM-LLM-01-02 title: A repaired answer remains one metered request kind: interaction exercises: - CAP-PFM-LLM-001 - CAP-PFM-LLM-002 - CAP-PFM-LLM-003 relies_on: - ATS-PFM-LLM-002-08 - ATS-PFM-LLM-003-05 preconditions: A task request first returns a rejected final answer and then a valid repair. steps: >- Reject the first answer through the consumer judge, accept the repair and inspect attempts and settlement. expected: >- Both turns belong to one request and one task; usage includes the repair and the rejected call consumes tool budget. negative: false data_needs: >- An isolated workspace with stated usage, callers, clock instants and a separate workspace for scope checks. bound_to: [] provenance: proposed source: - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-001-make-every-ai-request-through-one-metered-path.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-002-keep-an-ai-conversation-within-its-limits.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-003-account-for-a-multi-step-ai-task-as-one-execution.md - id: FAS-PFM-LLM-01-03 title: A failed task retains completed request accounts kind: lifecycle exercises: - CAP-PFM-LLM-001 - CAP-PFM-LLM-003 - CAP-PFM-LLM-004 relies_on: - ATS-PFM-LLM-003-01 - ATS-PFM-LLM-004-01 preconditions: A task completes two requests before its body throws. steps: 'Let the body fail and inspect the task, associated requests and evidence.' expected: >- The task closes failed with timing while earlier settled costs and capture references remain investigable. negative: false data_needs: >- An isolated workspace with stated usage, callers, clock instants and a separate workspace for scope checks. bound_to: [] provenance: proposed source: - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-001-make-every-ai-request-through-one-metered-path.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-003-account-for-a-multi-step-ai-task-as-one-execution.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-004-keep-redacted-evidence-of-every-ai-request.md - id: FAS-PFM-LLM-01-04 title: Workspace removal respects the diagnostic boundary kind: permission exercises: - CAP-PFM-LLM-001 - CAP-PFM-LLM-004 relies_on: - ATS-PFM-LLM-004-04 preconditions: Two workspaces have model records and captured content. steps: Remove the first workspace’s AI records and inspect both workspaces through scoped readers. expected: >- The removed workspace has no surviving dependent AI records or evidence versions; the other workspace remains unchanged. negative: true data_needs: >- An isolated workspace with stated usage, callers, clock instants and a separate workspace for scope checks. bound_to: [] provenance: proposed source: - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-001-make-every-ai-request-through-one-metered-path.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-004-keep-redacted-evidence-of-every-ai-request.md - id: FAS-PFM-LLM-01-05 title: Unknown usage keeps an explicit estimate kind: regression exercises: - CAP-PFM-LLM-001 - CAP-PFM-LLM-002 - CAP-PFM-LLM-003 relies_on: - ATS-PFM-LLM-001-03 - ATS-PFM-LLM-003-05 preconditions: A task’s scripted model completes successfully without usage figures. steps: 'Settle the request and compare request, task and period costs.' expected: >- The reservation estimate is charged with unknown cents and the hold is released; the task retains that accounting distinction. negative: true data_needs: >- An isolated workspace with stated usage, callers, clock instants and a separate workspace for scope checks. bound_to: [] provenance: proposed source: - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-001-make-every-ai-request-through-one-metered-path.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-002-keep-an-ai-conversation-within-its-limits.md - >- rhythm-v2-docs:specs/platform/llm/planning/capabilities/CAP-PFM-LLM-003-account-for-a-multi-step-ai-task-as-one-execution.md evidence: - artifact: CAP-PFM-LLM-001 provenance: documented source: >- specs/platform/llm/planning/capabilities/CAP-PFM-LLM-001-make-every-ai-request-through-one-metered-path.md - artifact: CAP-PFM-LLM-002 provenance: documented source: specs/platform/llm/planning/capabilities/CAP-PFM-LLM-002-keep-an-ai-conversation-within-its-limits.md - artifact: CAP-PFM-LLM-003 provenance: documented source: >- specs/platform/llm/planning/capabilities/CAP-PFM-LLM-003-account-for-a-multi-step-ai-task-as-one-execution.md - artifact: CAP-PFM-LLM-004 provenance: documented source: specs/platform/llm/planning/capabilities/CAP-PFM-LLM-004-keep-redacted-evidence-of-every-ai-request.md - provenance: documented source: 'rhythm-v2-docs:specs/CONVENTIONS.md#5.3 Feature Acceptance Suite' execution_gap: FND-PFM-LLM-081 --- # Metered AI Requests acceptance ## What the Feature Must Do as a Whole A declared task has three scripted requests and a configured capture store. Complete the task and compare request charges, task totals, monthly accounting and capture states. The suite compares the feature outcome with its capability prerequisites and keeps partial proof visible. ## Interactions Between Capabilities A task request first returns a rejected final answer and then a valid repair. Reject the first answer through the consumer judge, accept the repair and inspect attempts and settlement. The interaction scenario checks what changes across that shared boundary. Lifecycle, permission and regression scenarios preserve the wider feature obligations. ## Regression Record | Revision | Change | Scenario | Why | |---|---|---|---| | 1 | initial | all | Establish cumulative feature acceptance from the pinned capability and design baseline. |