Loop Ledger
A five-section persistent artifact that lets a task survive tool-call seams, context switches, compression, and restarts without re-deriving state.
When to use
- Loop-tier work: multi-file, multi-tool, multi-session, goal-run, or subagent-coordinated tasks.
- Not for fast single-step work — overhead exceeds value there.
Scaffold
Create threads/<thread_id>/artifacts/specs/ledger.md:
# Ledger: <one-line task name>
## Goal
<One sentence: the deliverable, not the activity.>
## Core
<Anchors established exactly once, broadcast to every branch/child:>
- Paths: <repo dirs, key files>
- Names/IDs: <goal_run_id, thread_id, session IDs, task IDs>
- Constraints: <hard requirements, acceptance criteria>
- Style/conventions: <language, format, coding conventions in force>
## Verified
<Append-only. One line per claim:>
- <claim> | verifier: <cmd/test/review + exit code> | covers: <scope> | not: <gap>
## Open
<Unresolved questions/blockers with owner or next probe.>
## Next
<Exactly ONE next action, executable cold without further context.>
Maintenance rules
- Core is broadcast-once. Subagent descriptions, child task briefs, and resume turns reference Core entries verbatim instead of paraphrasing. New anchors get added to Core and re-broadcast, never forked privately.
- Verified entries carry their verifier and coverage. A claim without verifier + coverage + gap does not belong here.
- Next is replaced, never accumulated. If it can't be executed cold, make it more specific first.
- On resume (pause/handoff/restart): read the ledger before anything else, confirm Goal still matches the original request, then execute Next.
- Checkpoints at seams: after each tool chain boundary, file switch, or subagent return, refresh against the ledger before acting.
Completion
The task is done when combined Verified coverage matches Goal scope (gaps explicitly accepted), Open is empty or deferred with stated reason, and the final report cites verifier + coverage per claim.
Source: adapted from J-Space Cognition Suite V3.6 mechanisms (capability-realization report, Tiger3807861189, 2026) — runtime-agnostic subset: 5-state ledger, verifier binding, failure inheritance, broadcast-once anchors. First-person grammar and first-turn persona anchoring intentionally excluded as model-specific overfits.
1---2name: loop-ledger3description: Scaffold a J-Space-style 5-state task ledger (Goal/Core/Verified/Open/Next) for long multi-step tasks and goal runs. Use when starting loop-tier work that spans multiple tools, files, sessions, or subagents, or when resuming such work after a pause or handoff. Pairs with the loop-ledger-verification-task guideline.4---56# Loop Ledger78A five-section persistent artifact that lets a task survive tool-call seams, context switches, compression, and restarts without re-deriving state.910## When to use1112- Loop-tier work: multi-file, multi-tool, multi-session, goal-run, or subagent-coordinated tasks.13- Not for fast single-step work — overhead exceeds value there.1415## Scaffold1617Create `threads/<thread_id>/artifacts/specs/ledger.md`:1819```markdown20# Ledger: <one-line task name>2122## Goal23<One sentence: the deliverable, not the activity.>2425## Core26<Anchors established exactly once, broadcast to every branch/child:>27- Paths: <repo dirs, key files>28- Names/IDs: <goal_run_id, thread_id, session IDs, task IDs>29- Constraints: <hard requirements, acceptance criteria>30- Style/conventions: <language, format, coding conventions in force>3132## Verified33<Append-only. One line per claim:>34- <claim> | verifier: <cmd/test/review + exit code> | covers: <scope> | not: <gap>3536## Open37<Unresolved questions/blockers with owner or next probe.>3839## Next40<Exactly ONE next action, executable cold without further context.>41```4243## Maintenance rules44451. **Core is broadcast-once.** Subagent descriptions, child task briefs, and resume turns reference Core entries verbatim instead of paraphrasing. New anchors get added to Core and re-broadcast, never forked privately.462. **Verified entries carry their verifier and coverage.** A claim without verifier + coverage + gap does not belong here.473. **Next is replaced, never accumulated.** If it can't be executed cold, make it more specific first.484. **On resume** (pause/handoff/restart): read the ledger before anything else, confirm Goal still matches the original request, then execute Next.495. **Checkpoints at seams:** after each tool chain boundary, file switch, or subagent return, refresh against the ledger before acting.5051## Completion5253The task is done when combined `Verified` coverage matches `Goal` scope (gaps explicitly accepted), `Open` is empty or deferred with stated reason, and the final report cites verifier + coverage per claim.5455Source: adapted from J-Space Cognition Suite V3.6 mechanisms (capability-realization report, Tiger3807861189, 2026) — runtime-agnostic subset: 5-state ledger, verifier binding, failure inheritance, broadcast-once anchors. First-person grammar and first-turn persona anchoring intentionally excluded as model-specific overfits.
mkurman