agentsclimarketplace

Context checkpoint

Skill prasadmogulothu/agent-skills/context-checkpoint

Reusable, token-optimized agent skills for autonomous development with Hermes, Claude Code & OpenCode. Multi-model orchestration patterns — context-checkpointing, human review gates, RLS security, GDPR consent, AI cost control, and test-fix loops. Framework-agnostic, agentskills.io compatible.

Install
npx -y skills add prasadmogulothu/agent-skills --skill context-checkpoint

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use whenever a session's context approaches its working limit OR the current atomic task completes. Cleanly checkpoint progress to a task file, commit, and end the session so a fresh one can resume — preventing quality loss from a bloated context window.

SKILL.md

2.5 KB, as published. Nobody here has run it

context-checkpoint

Treat the context window like a budget. One task ≈ one session. Checkpoint and resume cleanly instead of letting a session drift toward its limit mid-task. Progress lives in a shared tasks.md, so any session can stop and a fresh one can pick up with no lost work.

Thresholds

  • Set a soft ceiling (e.g. ~70-80% of your model's context) and a hard ceiling you never cross. When you approach the soft ceiling OR finish the current atomic unit, checkpoint and end.
  • If you can't read token usage directly, estimate from files read + edits + tool output, and checkpoint early. Early is always safe — resuming is cheap.

Checkpoint sequence (run at end of every session)

  1. Reach a safe stopping point — a complete function / file / passing test, not a half edit.
  2. Run your project's done-checks for what's complete (build, typecheck, lint, relevant tests).
  3. git commit with a message referencing the task id (wip: if intentionally partial). Uncommitted work is lost work.
  4. Update tasks.md for the current task: set Status (DONE / BLOCKED / IN_PROGRESS); if not done, record a Checkpoint (what's complete) and Resume notes (exact next step, files in play, decisions made). Append a line to a Session Log.
  5. Append durable learnings to a MEMORY.md (short bullets) so future sessions don't re-derive them.
  6. End / clear the session. Do not keep working in the same context.

Resume sequence (new session)

  1. Read tasks.md + MEMORY.md and only the files the next task needs — not the whole repo.
  2. Pick the first task whose dependencies are met. Set IN_PROGRESS, record the session id, work.
  3. Checkpoint again when near the soft ceiling or the task completes.

Splitting a too-large task

If a task won't fit one session, split it into independently completable sub-tasks in tasks.md, carry over dependencies, note why, and do the first sub-task. Never carry a half-done task across a session boundary without a written checkpoint.

Token tactics

Near-zero model cost — this is bookkeeping the agent runs itself. Keeping sessions small is the single biggest quality + cost win: a fresh lean context beats a bloated one every time.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.