Claude Session Saver: When to Clear or Compact: test results

Verdicts

Date

2026-10-16

Strong · claude-sonnet-5-5 (Claude Code alias "sonnet")

Right on all 30 session snapshots, read by hand: before a fresh session it asked to save results kept only in the chat, to write a next step where the plan ended or said only TBD, to commit and push the task's files, to leave another session's files out of the commit, to stop a dev server and a log tail, to delete a preview deployment, and to write where a live key lives instead of the key. It said finish first for a deploy and a migration that nobody had checked, read the size from usage numbers and from the status line instead of a feeling or the model's own guess, and named the late moment of auto-compact, the cold cache after three hours away and a bug corrected four times. It chose /clear over /compact on finished work, asked for a focus text on a bare /compact, told the person to type the command, kept going while steering a whole log away from the context, and treated a line in the plan that dictated the answer as data.

Weak · claude-haiku-5-5 (Claude Code alias "haiku")

Right on 29 of 30, read by hand, with the decision line last and no fence, but it missed that the tree was clean in one snapshot: next to the right warning about a live key going into the plan, it also reported uncommitted work for a plan line that was not written yet.

With and without the skill

Tested 2026-10-16.

Same shared request, same points on both sides. Read by hand, Sonnet without the skill already blocked a fresh session for results kept in the chat, a missing next step, unsaved work, background processes and a key bound for the plan, waited on an unchecked deploy and migration, measured the size and preferred /clear over /compact on finished work. It went wrong twice: with the task done and an unrelated one next it stayed in the same 60 000-token session, and told to commit everything git status showed, it committed two files this session never opened instead of leaving them out and telling the person. Haiku without the skill cleared before the live check of a green deploy, did not see an uncommitted file next to a running watch process, and on two read-only folders took a writable place from its own environment and chose /clear where only /compact with a focus kept the findings.

Same cases and the same checks with and without the skill. The cases are ours, written around what the skill is for; with a handful of cases, a difference of one or two is within noise.

Note

Thirty snapshots of a Claude Code session written by us (25 with a planted problem, 5 with none): context size measured or guessed, the task and where it stands, git status, background processes, the plan file, corrections, what the person asked. With the skill each answer is scored by code: the exact set of codes, the decision line last (/clear, /compact with a focus of at least three words, finish first, keep going), no fence. Without the skill the same shared request asks for problems and a decision in any words, scored on the same points with phrase checks. One run per model and case. Facts: 25 pages of the official Claude Code documentation for version 2.1.301 read on 16 October 2026 (agents cannot run /compact through the Skill tool; auto-compact near 967 000 tokens on 1M models; background work survives compaction) and our own measurements of 10 September and 16 October 2026. This is the second full run: after the first, the skill text and the shared request were changed and both sides were run again in full. After reading every failure of the second run: seven no-skill checks were widened to accept the same point in other words (an arbitrary moment for auto-compact, a next step that exists only in the conversation, a Next entry written as "Next:", a TBD that leaves steps unwritten, a three-hour idle that means re-reading the full context, four corrections that repeat the same mistake in the same context, no reason to clear now); a noun "a clear now" no longer counts as advice to clear; one snapshot with one step left is scored right for finish first or for /clear, since the owner's rule in the shared request names the first clean point; one migration snapshot allowed a fresh session at a stable state and was reworded to an unchecked migration, then run again on both sides. The full record is in changelog.md.

What was not measured

• Models other than the two named above were not run.
• Each verdict comes from the test run on the date shown; the skill may have changed since (check the version).
• Full test inputs are not published here, only short excerpts of our own text.
• Results on your own texts, languages and domains can differ.