CavsSatyamKhatri

Claude Session Recall — Development skill for Claude Code

Development community

A Claude Code skill that reads your session's own transcript, so a fact that arrived through a compaction summary can be checked instead of repeated.

How to install Claude Session Recall

This entry records only its repository, not the path inside it, so there is no exact command to give. Open CavsSatyamKhatri/claude-session-recall and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Claude Session Recall does

A Claude Code skill that reads your session's own transcript, so a fact that arrived through a compaction summary can be checked instead of repeated. trace tells you whether a claim had any evidence behind it before a summary asserted it.

Alternatives in Development

Claude Instructions For Local Skill Loading — This file tells Claude how to load skills locally 845 ★
Public APIs Live — Every public API in one place — aggregated, deduped, and auto-checked for reachability 137 ★
Agentjail — Policy guardrails for coding agents (Claude Code, Codex, Cursor) — every tool call is checked locally, before 79 ★

README

session-recall

**Your AI coding session has been compacted. It did not tell you how many times, and it cannot tell which of the things it "knows" it actually measured.**

Run this on your own session, before reading any further:

git clone https://github.com/CavsSatyamKhatri/claude-session-recall.git
node claude-session-recall/recall.mjs compactions

  Compacted 28 time(s). Everything you "remember" from before the last one
  reached you through a summary somebody wrote, not through the record.

    # 1  line    2816  2026-08-04 06:47  summary of 16,517 chars
    # 2  line    5885  2026-08-05 08:14  summary of 21,831 chars
    ...

Your number will be different. That is the point — it is *your* number, and nothing else was going to tell you.

A Claude Code skill that reads the session's own transcript, so a fact that arrived through a compaction summary can be checked instead of repeated.

The problem

When a conversation is compacted, what survives into the next context window is a **summary**: chosen, compressed and written by the assistant. The full record is not deleted — it stays in the transcript on disk — but nothing points at it.

So a fact that arrived through a summary is **indistinguishable** from a fact that was measured. It reads the same, it is stated with the same confidence, and it gets acted on the same way. Repeat that through several compactions and a claim nobody ever checked becomes something everybody believes.

Find your own inherited claims

You do not have to guess what to look for. Ask what your last summary asserted:

node recall.mjs claims

  Summary #28 of 28, line 103216, 17,109 chars.
  16 sentence(s) in it assert something checkable.

  This is a starting list, not a verdict: a sentence here is not wrong, and one that is
  missing is not cleared. Trace the ones your next decision depends on.

   1. [version] ... the release chore is still pending ...
      recall trace "the release cho