<< All versions

Skill v1.0.0

currentLLM-judged scan90/100
mraakashshah/oro/context-checkpoint
──Details
PublishedOctober 21, 2026 at 06:01 PM
Content Hashsha256:b8e8f35c947e5fd8...
Git SHA79f6cad693fd
──Files
Files (1 file, 3.1 KB)
SKILL.md3.1 KBactive
SKILL.md · 90 lines · 3.1 KB

version: "1.0.0" name: context-checkpoint description: Use when task completion or context usage requires a handoff readiness check user-invocable: false


Context Checkpoint

Overview

Monitor context consumption and trigger proactive handoff before quality degrades. Called after every task completion in executing-beads.

Core principle: Better to hand off early than to lose context and produce bad work.

How It Works

Layer 1 — Native context awareness (always present). Claude Sonnet 4.5+, Sonnet 4.6, and Haiku 4.5 receive context state after every tool call:

xml
<system_warning>Token usage: 35000/200000; 165000 remaining</system_warning>

No hook needed — workers already know their context state from system_warning. Thresholds are loaded from thresholds.json at the project root:

ModelSoft thresholdHard threshold
Opus65%85%
Sonnet50%70%
Haiku40%60%

Source of truth: thresholds.json.

The Go worker enforces the hard stop as a backstop: if the agent does not respond to the soft threshold, handleContextThreshold sends a handoff and kills the subprocess at hard threshold (soft + 20%).

Quality Signals (Override Token Count)

These symptoms indicate context degradation regardless of token usage — treat as immediate handoff:

  • Repeating yourself — suggesting something already tried or discussed
  • Forgetting earlier context — asking about decisions already made
  • Hallucinating paths — referencing files or functions that don't exist
  • Growing confusion — needing to re-read files you already read

Checkpoint Protocol

Green (Continue)

system_warning shows usage well below soft threshold. Proceed to next task.

Soft Threshold Breached

Finish your current atomic step (complete the file edit, run its verification, record the result). Then:

  • Do NOT start another edit-verify cycle.
  • Follow handoff steps below.

Hard Threshold Breached

Stop after your very next tool call. Write a minimal handoff (goal + files modified + next steps). Any in-progress work will be captured by the compaction safety net.

Handoff Steps

  1. Close current task if work is complete
  2. For in-progress work: oro task update <id> --notes "Partial: <what's done, what remains>"
  3. Verify remaining work exists as tasks (create if needed)
  4. Use create-handoff skill with tasks: section
  5. git pull --rebase && git push
  • Note: The pre-commit hook automatically syncs task metadata, so manual sync is not needed

Handoff Template Addition

When handing off due to context checkpoint, add tasks: section to handoff YAML:

yaml
tasks:
completed: [oro-xxx, oro-yyy]
in_progress: [oro-zzz]
remaining: [oro-aaa, oro-bbb]
epic: oro-www

Red Flags

  • Ignoring hook messages and starting new tasks
  • Skipping checkpoint after task completion
  • Not saving in-progress state before handoff
  • Producing a handoff without the tasks: section
  • Rationalizing "just one more task" after handoff message
All versions