Compact guard

Skill dhruba-datta/claude-papercuts/skills/compact-guard

Ten Claude Code skills that fix bugs Anthropic closed as 'not planned'.

Install

npx -y skills add dhruba-datta/claude-papercuts --skill compact-guard

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

Survive /compact and auto-compact without losing the plan. Use this skill when the user runs /claude-papercuts:compact-guard, asks why Claude lost their plan after a compact, mentions plan-mode state vanishing mid-session, or wants to see the latest pre-compact snapshot. A PreCompact hook silently captures the current task, active todos, files edited, and the last assistant plan to .papercuts/compact-snapshots/ right before compaction runs; a SessionStart hook (only when source=compact) injects that snapshot back as context so the post-compact session knows what it was doing.

SKILL.md

5.0 KB, ~1.1k tokens by cl100k_base, as published. Nobody here has run it

compact-guard — survive /compact

Fixes: #24686, #26061

What this prevents

"After context compaction occurs during plan mode execution, Claude Code fails to re-read or reference the existing plan." — issue #24686

"Plan mode state lost after context compression." — issue #26061

You're 30 turns into a careful refactor. /compact runs (manually or because you hit the auto-compact threshold). The summarizer condenses everything, including the plan you were halfway through. Claude wakes up on the other side without the plan, and you have to re-prompt from scratch.

compact-guard snapshots the plan before compaction touches it, then re-injects the snapshot as context immediately after, so the post-compact turn starts knowing the current task, active todos, files touched, and the most recent plan.

How it works (zero manual work for the user)

/compact (manual)  or  auto-compact at threshold
        │
        ▼
   PreCompact hook fires
        │
        ▼
Read transcript → extract:
  • Most recent user message     (the current task)
  • Latest TodoWrite todos       (pending + in_progress only)
  • Files edited this session    (Edit / Write / MultiEdit)
  • Last assistant text block    (likely the working plan)
        │
        ▼
Write .papercuts/compact-snapshots/<timestamp>.md
(keeps 5 most recent — oldest pruned)


Compaction runs. The summarizer eats most of the context.


Next session start with source="compact"
        │
        ▼
   SessionStart hook fires
        │
        ▼
Read the latest snapshot → print to stdout
Claude Code auto-injects stdout as additional context

When you (the model) should invoke this skill manually

User runs /claude-papercuts:compact-guard
User asks "what was I doing before the compact?"
User asks to see the latest pre-compact snapshot
User says they lost their plan / their todos / their work after a compact ran

Manual invocation procedure

When asked to show compact-guard state:

Check if .papercuts/compact-snapshots/ exists in the project root.
If not, tell the user no compaction has happened yet in this project — the directory will populate on the first PreCompact. Then stop.
Otherwise, list the snapshots (sorted newest first). Tell the user:
- How many snapshots are stored
- The timestamp of each
If asked for a specific snapshot, cat it to the user verbatim. The format is plain markdown — don't summarize.
Never modify or delete snapshots automatically. If the user asks you to clean them up, ask for explicit confirmation first.

Snapshot format

Each snapshot is a markdown file in .papercuts/compact-snapshots/<UTC-timestamp>.md:

# compact-guard snapshot

- **When:** 2026-05-16 09:14 UTC
- **Trigger:** auto-compact
- **Session:** abc123def456

## Current task (last user message)
> <the most recent user prompt>

## Active todos
- [in_progress] <todo content>
- [pending] <todo content>

## Files edited this session
- `src/middleware/auth.ts`
- `src/__tests__/auth.test.ts`

## Last assistant message (plan / summary)

<the last assistant text block — usually the working plan, capped at 1500 chars>

Sections that have nothing useful are omitted. Total snapshot file is uncapped because it's read on demand, not auto-injected into the system prompt.

What this skill does NOT do

It does not block compaction. PreCompact runs read-only and exits 0.
It does not modify the transcript. The snapshot is a separate file.
It does not deduplicate snapshots. Five most-recent are kept, in case you want to inspect a series of compactions.
It does not survive across projects. Snapshots live in .papercuts/compact-snapshots/ relative to the project's cwd.
It does not inject context on /clear, /resume, or startup — only post-compact. Use amnesia-fix for cross-session continuity.

Privacy

Snapshots are written to a local file in your project's .papercuts/ directory. No network calls. Add .papercuts/ to your .gitignore if not already.

Deprecation plan

If Anthropic ships a compact-resilient plan-mode (per issues #24686 and #26061), this skill becomes a no-op and gets deprecated in the next monthly release with the date.