Link files that agents change through shell commands

[Soph]

https://entire.io/gh/entireio/cli/trails/1534

Summary

Since #2692, a commit links to a session by the files the session recorded. Two cases recorded nothing when an agent changed a file through a shell command, so the commit carrying that file got no Entire-Checkpoint trailer.

Shell edit to a file that was already untracked. Turn end excludes untracked files that existed at prompt start from "new", and git status cannot report an untracked file as modified, so the edit was invisible. The prompt-start baseline now also records each untracked file's stat data (PrePromptState.UntrackedFileStats): size, modification time, inode and status-change time, the fields git's index compares. The change time catches a same-size rewrite that restores the modification time (cp -p, touch -r); Windows has no change time, so only size and modification time are compared there. At turn end, a baseline file whose stat data changed joins the modified list before filterToUncommittedFiles, so a file committed mid-turn is still dropped. Stat data instead of content hashes keeps UserPromptSubmit free of an extra git hash-object per prompt.

File created by a background subagent's shell command after the parent's turn ended. A background SubagentStop deliberately skips the worktree scan (it would sweep in the parent's and other agents' concurrent work), and the transcript names no file. Once the parent's next prompt starts, the file sits in that prompt's untracked baseline, so no later turn end claims it. SubagentStop now also claims untracked files created since the task launched that no session in the same worktree has recorded (unrecordedNewFilesSinceLaunch). Files the parent's own turns or another session recorded are excluded. When another background task in the worktree overlapped this one (new TaskRecord.Background flag, still live or completed after this task's launch), nothing is claimed, since the first task to stop would otherwise take both tasks' files. Accepted trade-off: a file the user creates while the parent is idle is claimed as well. Since #2692 removed line attribution, that only adds a session link to the commit carrying it, which matches what turn end already does for user edits made during an active turn.

The first commit also pins the ordering from #2653: a background task completes mid-turn and the parent commits in the same turn. The commit links, lists the file, and carries the subagent transcript. #2653's reported symptom was in line attribution, which #2692 removed.

Tests

• TestSubagentCheckpoints_BackgroundTaskCompletesMidTurn_ParentCommitLinks: the Background subagent's lines are attributed to the human when the parent commits in the same turn #2653 ordering. It fails if the subagent's completion is dropped.
• TestShellEditToPreExistingUntrackedFile_LinksCommit: the shell-edited file is recorded and linked, and an untouched pre-existing untracked file is not recorded.
• TestBackgroundSubagentShellCreatedFile_LinksCommit: the subagent's shell-created file is recorded and linked.
• TestShellRewriteKeepingSizeAndMtime_IsRecorded: a same-size rewrite with the original modification time restored is still recorded. Skipped on Windows. It fails when the change time is dropped from the comparison.
• TestOverlappingBackgroundTasks_ClaimNoNewFiles: two overlapping background tasks from different sessions each create a file, and neither session claims either file. It fails when the overlap check is removed.
• TestBackgroundSubagentStop_ClaimsOnlyUnrecordedNewFiles: of five files, the task claims only the subagent's file and the user's idle-window file. It does not claim the pre-launch file, the parent's later-turn file, or another session's file. It fails when the "already recorded" exclusion is removed.

Both gap tests failed on main before the fix. Their controls, where the transcript names the file, passed.

mise run check passes.

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

Metadata collisions can miss edits, while concurrent background tasks can claim each other’s files.

2 open findings

• Concurrent tasks can misattribute files across sessions · New
• Same-size, same-mtime rewrites evade change detection · New

What changed in this PR

Improves checkpoint linking for files changed by shell commands.

Changes:

• Detects edits to pre-existing untracked files using metadata.
• Claims unrecorded files created during background tasks.
• Adds integration coverage and architecture documentation.

File	Description
docs/​architecture/​sessions-and-checkpoints.md	Documents shell-change capture behavior.
cmd/​entire/​cli/​strategy/​manual_commit_session.go	Finds files recorded by worktree sessions.
cmd/​entire/​cli/​state.go	Stores and compares untracked-file metadata.
cmd/​entire/​cli/​lifecycle.go	Integrates shell-created files into capture.
cmd/​entire/​cli/​integration_test/​subagent_commit_in_turn_test.go	Tests mid-turn subagent commits.
cmd/​entire/​cli/​integration_test/​shell_capture_test.go	Tests shell-based file capture.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.