Add Hidden Checkpoint Create for Snapshots

Pi·GPT-6-astra·Soph·1w ago·118hr 53min·1 Checkpoint·35 file changes·+2525/-55·42.7K tokens

can you review the changes in this branch

1w ago·4.4s

The branch has uncommitted changes adding a hidden checkpoint create command. I’ll review those changes and their interactions with session and checkpoint handling.

Found one issue:

  • [P2] Snapshot checkpoints record misleading code attribution — cmd/entire/cli/strategy/manual_commit_snapshot.go:56
    CondenseSession calculates attribution assuming HEAD contains the newly committed work. This command creates no commit, so HEAD still contains the old code. If a shadow checkpoint contains uncommitted agent additions, the shadow→HEAD comparison treats those additions as human removals. Omit commit attribution for snapshots, or calculate it against an appropriate snapshot tree instead.

The new command and strategy tests pass. They don’t currently verify attribution.

can you check again

1w ago·4.6s

Rechecked the updated changes. The attribution issue is fixed: snapshots omit commit attribution, and the new regression test verifies that.

No additional findings. Targeted command, snapshot, condensation, and attribution tests pass. I haven’t run the full CI suite.

bef1c24Add hidden `entire checkpoint create` for snapshot checkpoints Creates a checkpoint from a session's current transcript and prints its ID. With no argument it checkpoints the calling agent's session, and only when the caller was identified (caller-env or ancestry); most-recent fallbacks are refused so a guess cannot checkpoint someone else's work. The write goes through CondenseSession, so extraction, redaction, and the store write (push-queue enqueue / unpushed v1 commit) are unchanged. It is a snapshot: state is loaded, condensed in memory, and never saved back, so a mid-turn call cannot race the session's hooks and the next commit still condenses the same range. Skill telemetry is not emitted, since the next commit would count those events again. Commit attribution is omitted (new condenseOpts.noCommitAttribution): it diffs the shadow tree against HEAD assuming HEAD holds the committed work, which without a commit counts uncommitted agent edits as human edits. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Entire-Checkpoint: 01M3D8JE8P1DAB87QY1YT4BDDS+505/-11
dfb704bAdd hidden `entire checkpoint create` for snapshot checkpoints Creates a checkpoint from a session's current transcript and prints its ID. With no argument it checkpoints the calling agent's session, and only when the caller was identified (caller-env or ancestry); most-recent fallbacks are refused so a guess cannot checkpoint someone else's work. The write goes through CondenseSession, so extraction, redaction, and the store write (push-queue enqueue / unpushed v1 commit) are unchanged. It is a snapshot: state is loaded, condensed in memory, and never saved back, so a mid-turn call cannot race the session's hooks and the next commit still condenses the same range. Skill telemetry is not emitted, since the next commit would count those events again. Commit attribution is omitted (new condenseOpts.noCommitAttribution): it diffs the shadow tree against HEAD assuming HEAD holds the committed work, which without a commit counts uncommitted agent edits as human edits. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Entire-Checkpoint: 01M3D8JE8P1DAB87QY1YT4BDDS+505/-11
69d3875Add hidden `entire checkpoint create` for snapshot checkpoints Creates a checkpoint from a session's current transcript and prints its ID. With no argument it checkpoints the calling agent's session, and only when the caller was identified (caller-env or ancestry); most-recent fallbacks are refused so a guess cannot checkpoint someone else's work. The write goes through CondenseSession, so extraction, redaction, and the store write (push-queue enqueue / unpushed v1 commit) are unchanged. It is a snapshot: state is loaded, condensed in memory, and never saved back, so a mid-turn call cannot race the session's hooks and the next commit still condenses the same range. Skill telemetry is not emitted, since the next commit would count those events again. Commit attribution is omitted (new condenseOpts.noCommitAttribution): it diffs the shadow tree against HEAD assuming HEAD holds the committed work, which without a commit counts uncommitted agent edits as human edits. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Entire-Checkpoint: 01M3D8JE8P1DAB87QY1YT4BDDS+505/-11
b7a1493Add hidden `entire checkpoint create` for snapshot checkpoints Creates a checkpoint from a session's current transcript and prints its ID. With no argument it checkpoints the calling agent's session, and only when the caller was identified (caller-env or ancestry); most-recent fallbacks are refused so a guess cannot checkpoint someone else's work. The write goes through CondenseSession, so extraction, redaction, and the store write (push-queue enqueue / unpushed v1 commit) are unchanged. It is a snapshot: state is loaded, condensed in memory, and never saved back, so a mid-turn call cannot race the session's hooks and the next commit still condenses the same range. Skill telemetry is not emitted, since the next commit would count those events again. Commit attribution is omitted (new condenseOpts.noCommitAttribution): it diffs the shadow tree against HEAD assuming HEAD holds the committed work, which without a commit counts uncommitted agent edits as human edits. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Entire-Checkpoint: 01M3D8JE8P1DAB87QY1YT4BDDS+505/-11
30c1012Add hidden `entire checkpoint create` for snapshot checkpoints Creates a checkpoint from a session's current transcript and prints its ID. With no argument it checkpoints the calling agent's session, and only when the caller was identified (caller-env or ancestry); most-recent fallbacks are refused so a guess cannot checkpoint someone else's work. The write goes through CondenseSession, so extraction, redaction, and the store write (push-queue enqueue / unpushed v1 commit) are unchanged. It is a snapshot: state is loaded, condensed in memory, and never saved back, so a mid-turn call cannot race the session's hooks and the next commit still condenses the same range. Skill telemetry is not emitted, since the next commit would count those events again. Commit attribution is omitted (new condenseOpts.noCommitAttribution): it diffs the shadow tree against HEAD assuming HEAD holds the committed work, which without a commit counts uncommitted agent edits as human edits. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Entire-Checkpoint: 01M3D8JE8P1DAB87QY1YT4BDDS+505/-11