| name | friday-state-schema |
| description | State file JSON schema, status values, and generic delegation rules for the FRIDAY orchestrator. Load this skill whenever FRIDAY needs to read, write, or validate .temper/friday-state.json.
|
FRIDAY State Schema & Delegation Rules
State File JSON Schema
.temper/friday-state.json is FRIDAY's only persistent orchestration memory.
{
"last_updated": "ISO timestamp",
"status": "fresh | in-progress | awaiting-approval | awaiting-task-approval | awaiting-agent-cycle | complete | blocked",
"request_summary": "one line description",
"context": {
"project": "project name or description",
"architecture": "Clean Architecture | Hexagonal Architecture | Vertical Slice Architecture | Onion Architecture | unknown",
"stack": "EF Core, Blazor, etc.",
"notes": "any other relevant context"
},
"approved_plan": [
{
"step": 1,
"agent": "temper-analyst",
"phase": "phase-1-prd | phase-2-specs | null",
"description": "one line description",
"status": "complete | pending | in-cycle",
"output": "file path or null"
}
],
"current_step": 2,
"total_steps": 4,
"current_agent": "temper-backend",
"current_execution_mode": "task-driven | direct-action | direct-bugfix | recovery | null",
"current_task": "T001",
"current_work_item_type": "user-story | bug | refactor | setup | future-type | null",
"current_work_item_id": "US-001 | BUG-001 | REF-001 | SETUP | null",
"current_task_category": "Backend | Frontend | Testing | DevOps | null",
"current_task_location": "Plan/.../T001-[slug].md | null",
"task_title": "one line description",
"total_tasks": 6,
"completed_tasks": [
{ "task_id": "T002", "work_item_type": "user-story", "work_item_id": "US-001", "category": "Backend", "agent": "temper-backend", "title": "description", "status": "complete", "location": "Plan/User-Stories/US-001-[slug]/Backend/T002-[slug].md" }
],
"pending_tasks": [
{ "task_id": "T001", "work_item_type": "user-story", "work_item_id": "US-001", "category": "Backend", "agent": "temper-backend", "title": "description", "status": "pending", "dependencies": ["T000"], "location": "Plan/User-Stories/US-001-[slug]/Backend/T001-[slug].md" }
],
"active_cycle": {
"agent": "temper-analyst | temper-architect",
"phase": "phase-1-prd | phase-2-specs | null",
"cycle_type": "gap-resolution | spec-ambiguity-resolution | architect-loop",
"mode": "architectural-design | problem-solving | null",
"waiting_for": "gap-answer | gap-batch-send | ambiguity-answer | ambiguity-batch-send | mode-clarification | context-clarification | preference-clarification | problem-clarification | proposal-confirmation | document-selection | manual-review | none",
"last_report_type": "gap-report | resolution-status | ambiguity-stop | ambiguity-resolution-status | mode-report | clarification-request | domain-analysis | problem-analysis | architectural-proposal | architectural-plan | updated-proposal | document-offer | completion-report",
"question_origin": "analyst | architect | null",
"pending_interaction": {
"surface_via": "question | plain-text",
"interaction_type": "analyst-gap | analyst-ambiguity | architect-clarification | architect-decision | parse-fallback",
"prompt_text": "single actionable prompt or fallback instruction",
"expected_reply": "single gap answer | single ambiguity answer | mode details | clarification answer | preference bucket answer | proposal confirmation/change request | document selection | manual reply",
"source_ref": {
"report_type": "gap-report | resolution-status | ambiguity-stop | ambiguity-resolution-status | clarification-request | architectural-proposal | architectural-plan | updated-proposal | document-offer",
"item_id": "GAP-001 | AMB-001 | null",
"sequence": 1,
"total": 3
},
"resume_hint": "one-line reminder of what the user still needs to answer or decide",
"fallback_reason": "null or short explanation"
},
"gap_queue": [
{
"gap_id": "GAP-001",
"severity": "BLOCKING | IMPORTANT | CLARIFYING",
"question_text": "single actionable analyst question"
}
],
"collected_gap_answers": {
"GAP-001": "user answer captured exactly as provided"
},
"ambiguity_queue": [
{
"ambiguity_id": "AMB-001",
"severity": "BLOCKING",
"question_text": "single actionable analyst question"
}
],
"collected_ambiguity_answers": {
"AMB-001": "user answer captured exactly as provided"
},
"current_gap_index": 0,
"current_ambiguity_index": 0,
"unresolved_blocking_gaps": 3,
"cycle_count": 1
},
"session_metrics": {
"specialist_runs_total": 2,
"specialist_runs_current_thread": 1,
"distinct_specialists": ["temper-analyst", "temper-architect"],
"distinct_specialist_count": 2,
"cycle_rounds_total": 3,
"recovery_attempts": 1,
"prompt_failure_retries": 1,
"long_outputs_count": 2,
"large_summaries_count": 1,
"last_output_class": "completion-report | proposal | task-result | gap-report | ambiguity-report | failure | partial-output | unclear-output | recovery-report | none",
"next_action_type": "none | direct-answer | output-approval | approved-next-specialist | approved-exact-continuation | analyst-cycle-input | architect-cycle-input | recovery-approval | repair-or-reset | blocked-input | change-direction",
"scope_confusion_signals": 1,
"mixed_scope_signal": false,
"risk_score": 3
},
"block_reason": null,
"next_action": "what the next session should do"
}
Status Values
| Status | Meaning |
|---|
fresh | No active approved plan is in progress; FRIDAY should wait for a new request or a reset flow just completed. |
in-progress | FRIDAY is actively working on an approved step. |
awaiting-approval | A plan or next specialist step is proposed and waiting for user approval. |
awaiting-task-approval | A task-oriented agent completed and FRIDAY is waiting for the user to confirm the output or next action. |
awaiting-agent-cycle | An analyst or architect loop is active and waiting for the next loop action or user input. |
complete | All approved steps are complete. |
blocked | FRIDAY cannot proceed without user intervention. |
Startup State Health Check
At startup, FRIDAY must produce a concise operational note containing:
- State file:
found or not found.
- Status: current
status, or invalid with the shortest reason.
- Active plan: current step and total steps, or
none.
- Next action: the next safe orchestration action.
Do not use ceremonial startup phrases. If state is missing and the user is not asking to continue prior work, report fresh state and proceed with normal classification.
Required-Field Validation
Before resuming, delegating, or writing state, validate these fields:
- Root object must be valid JSON. Invalid JSON blocks delegation until repaired, reset, or replaced with explicit approval.
last_updated, status, request_summary, and next_action are required for any persisted state.status must be one of the defined status values.approved_plan is required when status is in-progress, awaiting-approval, awaiting-task-approval, or complete for a multi-step plan.current_step and total_steps are required when approved_plan exists.current_step must be an integer from 1 through total_steps; if all steps are complete, it may equal total_steps with status: "complete".- Each
approved_plan item must include step, agent, description, status, and output.
approved_plan[].step must be unique, sequential, and match its order.approved_plan[].status must be complete, pending, or in-cycle.current_agent is required when status is in-progress or a pending approval targets a specific next agent.current_execution_mode is allowed for implementation routing and should be set when FRIDAY is delegating or awaiting approval for an implementation agent.session_metrics is optional but allowed at the root and should be persisted once FRIDAY has enough checkpoint data to maintain session-mode recommendations safely.- When
session_metrics exists, it must be an object with non-negative integer counters, a boolean mixed_scope_signal, and a non-negative integer risk_score.
block_reason is required and non-empty when status is blocked.
If required fields are missing or contradictory, stop delegation, warn that state cannot be resumed safely, and ask for repair/reset approval or one concise clarification.
Session Metrics
session_metrics is the persisted summary FRIDAY uses for clean-session versus continue-here recommendation behavior.
Allowed fields:
specialist_runs_total: Total completed specialist runs captured in the active workflow state.specialist_runs_current_thread: Completed specialist runs in the current conversation thread since the latest clean restart boundary.distinct_specialists: Distinct specialist agent names already used in the current thread.distinct_specialist_count: Count of distinct_specialists.cycle_rounds_total: Total analyst/architect cycle rounds accumulated in the active workflow.recovery_attempts: Recovery or re-delegation attempts after unusable output, failure, or blocked continuation.prompt_failure_retries: Retries specifically caused by FRIDAY prompt-quality failure.long_outputs_count: Count of long specialist outputs or reports that materially increase live context load.large_summaries_count: Count of broad FRIDAY summaries or synthesis turns that materially increase live context load.last_output_class: Class of the most recent completed specialist output.next_action_type: Normalized next-step class used by the recommendation policy.scope_confusion_signals: Count of recent drift, ambiguity, duplication, or scope-blur indicators.mixed_scope_signal: true when the conversation mixed materially different scopes or specialist concerns in a way that raises continuation risk.risk_score: Computed non-negative integer summarizing continuation risk for the session-mode recommendation policy.
Persistence rules:
- Older state files may omit
session_metrics; FRIDAY may initialize it lazily from zero/default values when the next completed specialist checkpoint is written.
- Keep
session_metrics compact and derived. It is a summary layer, not a transcript or second workflow log.
- Refresh
last_output_class, next_action_type, and risk_score whenever FRIDAY writes post-execution state after a specialist completion.
- Keep
distinct_specialists and distinct_specialist_count consistent.
- Do not store bulky excerpts, raw reports, or free-form narrative inside
session_metrics.
Pending Step And Output Validation
When approved_plan exists:
- The pending step is the first item with
status: "pending" or status: "in-cycle"; it must match current_step unless state is complete.
- Do not skip a pending step unless the user explicitly approves a change-direction plan.
- A completed step must have an
output value when that specialist is expected to produce a durable artifact or completion report.
- If expected output is missing, mark state as blocked or recovery-needed instead of advancing.
- If the next step depends on a previous output, verify the previous step is
complete and has the expected output reference before delegation.
current_task, task_title, current_work_item_type, current_work_item_id, current_task_category, current_task_location, pending_tasks, and task counters must agree before delegating task-driven implementation.- For
current_execution_mode: "direct-action", task metadata may be null and task counters may be absent if no task artifacts exist.
- For
current_execution_mode: "direct-action", FRIDAY must still preserve an approved plan or explicit approval state before delegation.
FRIDAY does not need to read full artifact contents for this check; it validates that required references and workflow statuses exist and are coherent.
Prerequisite Checks
Before routing between workflow phases, enforce these prerequisites:
- Analyst Phase 2 may start only after analyst Phase 1 is complete and the PRD output has been approved or the user explicitly approves a change-direction exception.
- Architect routing for normal project architecture may start only after Phase 2 specs are complete or the user explicitly changes direction with approval.
- Task generation may start only after required requirements/spec and architecture outputs are complete and approved.
- Plan generation may start only after task generation is complete and approved.
- Implementation agents may start only from an approved plan/task, an explicitly approved direct-action path, or an explicitly approved direct bugfix/recovery path.
If a prerequisite is missing, do not delegate. Report the missing prerequisite and propose the smallest valid next step.
Active-Cycle Validation
When status is awaiting-agent-cycle, active_cycle is required and must include:
agent with temper-analyst or temper-architect.cycle_type matching the agent.waiting_for with a valid value for that cycle.last_report_type.question_origin matching the agent.pending_interaction unless waiting_for is none.cycle_count as a positive integer.
pending_interaction must include:
surface_via.interaction_type.prompt_text or a resume_hint for manual review.expected_reply.source_ref.report_type.source_ref.sequence and source_ref.total when presenting queued items.
For analyst cycles:
phase is required and must be phase-1-prd or phase-2-specs.- Phase 1 requires
cycle_type: "gap-resolution", gap_queue, collected_gap_answers, and current_gap_index.
- Phase 2 requires
cycle_type: "spec-ambiguity-resolution", ambiguity_queue, collected_ambiguity_answers, and current_ambiguity_index.
- Queue index must be within the queue bounds unless the next action is batch send.
For architect cycles:
cycle_type must be architect-loop.mode must be architectural-design or problem-solving unless the current wait is mode clarification.waiting_for must match the stage represented by last_report_type.
If active-cycle state is incomplete, ask only for the missing resume decision or propose safe parse fallback. Do not infer answers or fabricate queues.
Blocked Handling And Recovery Warnings
When status is blocked:
- Do not delegate.
- Show
block_reason in one sentence.
- Ask for the exact missing input, approval, repair decision, or reset decision.
- Preserve prior approvals unless the recovery changes their scope.
Warn before recovery when any of these are true:
- State JSON was invalid and had to be ignored.
- Required fields are missing from an active plan or active cycle.
- A pending step points to an agent that does not match prerequisites.
- An expected output reference is missing for a completed step.
- The user's request conflicts with saved next action.
- Recovery would supersede approved plan, approved requirements, approved architecture, tasks, or implementation output.
Delegation Rules - Domain Language Only
FRIDAY never tells an agent how to build something. FRIDAY tells specialists what outcome is needed and lets the specialist load its own context and skills.
For prompt construction, context control, recovery prompts, and domain-language reformulation, load workflow/friday/prompt-excellence.
For agent-specific contracts, load:
workflow/friday/implementation-delegationworkflow/friday/analyst-communicationworkflow/friday/architect-communication
Absolute Delegation Prompt Prohibitions
These prohibitions apply to every delegated agent, including temper-analyst and temper-architect.
- Never mention file paths,
.temper/, Markdown files, source files, or folder locations.
- Never mention skill names in a delegation prompt.
- Never copy acceptance criteria into implementation prompts.
- Never mention class names, DTO names, interface names, or method names.
- Never say
Read, Load, Check, or See file.
- Never describe implementation layers or tell the specialist where code belongs.
- Never summarize away user answers that must be passed back exactly.
Pre-Send Rule
Before sending any prompt, load the workflow contract that matches the target agent and validate the prompt against that contract.
workflow/friday/implementation-delegation for task-driven execution agents, bugfix turns, and recovery turns.workflow/friday/analyst-communication for analyst Phase 1 and Phase 2 loops.workflow/friday/architect-communication for architect loop turns.
Implementation agents resolve their task from Plan/INDEX.md, read the task file at its Location, and then read the parent work item source file. Conversation-loop agents may receive raw user request and minimum plain-language context when their contract requires it.
Agent-Specific Cycle Contracts
active_cycle carries enough shared structure for analyst and architect loops. The authoritative interaction rules live in dedicated FRIDAY workflow skills:
workflow/friday/analyst-communication for Phase 1 gap-resolution, Phase 2 ambiguity-resolution, batching, parse fallback, and resume rules.workflow/friday/architect-communication for clarification, preference, confirmation, document-selection, parse fallback, and resume rules.
Session Mode Recommendation
When FRIDAY reaches post-completion continuation handling, load workflow/friday/session-mode-recommendation.
That skill is authoritative for:
- Whether a session-mode recommendation should be offered.
- The default recommendation bias.
- Thresholds, hard exceptions, and context-load signals.
- The required recommendation output format.
Analyst-Specific State Notes
temper-analyst is modeled as two explicit orchestrator phases:
phase: "phase-1-prd" - requirements elicitation and PRD generation.phase: "phase-2-specs" - spec generation from an approved PRD.
When active_cycle.agent is temper-analyst, FRIDAY must persist both phase and cycle_type so resume logic can distinguish:
- Phase 1 loop:
cycle_type: "gap-resolution".
- Phase 2 loop:
cycle_type: "spec-ambiguity-resolution".
FRIDAY must never infer analyst phase from the agent name alone.
