daily-loop-runner（每日循环跑步者）

Daily Loop Runner

Advance one project by one meaningful step per loop. State read before execution, structured writeback after.

Input

Required:

• project_card — full project card with current state, goals, blockers
• latest_weekly_review — most recent weekly review notes
• recent_daily_logs — list of recent daily loop logs (last 3-5)
• open_questions — unresolved questions from previous loops

Optional:

• forced_bottleneck — override automatic bottleneck selection

Output Schema

today_objective: string               # one clear goal for today
selected_agent: string | null         # which agent or tool will execute
task_input: object                    # structured input for the selected agent
expected_output: string                # what success looks like
execution_summary: string             # what actually happened
findings: string[]                    # key discoveries
decisions: string[]                   # decisions made based on findings
next_action: string | null            # what to do tomorrow
project_card_updates: object          # fields to update on project card
writeback_payload: object             # structured record for project memory
safe_to_proceed: boolean              # false if state was incomplete

Hard Rules

1. One project per run. Do not split focus.
2. One bottleneck per run. Pick the most critical blocker.
3. One main action per run. One meaningful step, not a sprint.
4. No execution without state read. Always read project_card and recent logs first.
5. No successful completion without next_action. Every loop must feed into the next.

Loop Phases

Phase 1: State Read

• Read project_card fully
• Read recent_daily_logs
• Note open_questions from previous runs
• Identify current project phase and milestone

Phase 2: Bottleneck Selection

• Pick the single most critical bottleneck
• If forced_bottleneck provided, use it
• If nothing is blocking, advance the primary goal

Phase 3: Task Input Construction

• Build a focused task_input for the selected agent
• Include: what to do, why it matters, what success looks like
• Exclude: everything else

Phase 4: Execution

• Dispatch task to selected agent
• Wait for execution_summary and findings

Phase 5: Writeback

• Update project_card with project_card_updates
• Write execution_summary + findings to daily log
• Populate next_action for tomorrow's loop
• Set safe_to_proceed = true

Failure Handling

If state is incomplete (missing project_card, no recent logs, unclear objective):

• Stop the run immediately
• Set safe_to_proceed = false
• Request cleanup or missing-state repair
• Do not attempt execution with partial state

State Machine

IDLE → READY → RUNNING → WRITING → DONE
                    ↓
              BLOCKED (if state incomplete)

Respect the state machine. Never skip from IDLE to RUNNING.