feat: async captured checkpoints for bash tool hooks

[jwiegley]

Summary

• Adds daemon snapshot.watermarks API to query per-file mtime watermarks
• Pre-hook captures content of stale files (mtime > watermark + grace) as Human checkpoint
• Post-hook captures bash command's changed files as AiAgent checkpoint
• Both submitted as fire-and-forget captured checkpoints via existing pipeline
• Graceful degradation: non-daemon mode, query failures, large trees all fall back safely

Design

See docs/superpowers/specs/2026-03-27-async-bash-snapshots-design.md

Changes

• 17 files changed, 853 insertions across daemon infrastructure, bash tool, presets, and handler
• New SnapshotWatermarks control API with 500ms timeout
• FamilyState.file_snapshot_watermarks with #[serde(default)] for backward compat
• BashToolResult wrapper type replacing raw BashCheckpointAction returns
• captured_checkpoint_id field on AgentRunResult for early-return handler path
• 8 new unit tests + 2 family actor watermark tests

Test plan

• Unit tests for system_time_to_nanos, find_stale_files, capture_file_contents
• Family actor watermark round-trip tests (get/update/monotonic advance)
• All 54 bash_tool_provenance integration tests pass
• Full cargo test suite passes (1312 lib tests)
• cargo clippy --all-targets clean

🤖 Generated with Claude Code

---

[devin-ai-integration]

Devin Review found 1 new potential issue.

View 8 additional findings in Devin Review.

[devin-ai-integration]

🟡 Live checkpoint watermark computation ignores will_edit_filepaths for Human checkpoints

In src/daemon.rs:5257-5261, when computing checkpoint_file_paths for watermark updates on the CheckpointRunRequest::Live path, only edited_filepaths is read. Human checkpoints store their target paths in will_edit_filepaths (not edited_filepaths), as seen in src/commands/checkpoint.rs:183-187 (explicit_capture_target_paths selects will_edit_filepaths for Human kind). This means any Human checkpoint processed via the Live daemon path will produce an empty checkpoint_file_paths, causing compute_watermarks_from_stat at src/daemon.rs:5282 to return no watermarks. Consequently, the watermark map is never updated for these files, and find_stale_files (src/commands/checkpoint_agent/bash_tool.rs:849-876) will perpetually treat them as stale (no watermark → always stale), causing unnecessary pre-hook captures on every subsequent bash tool invocation.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

Devin Review found 1 new potential issue.

View 15 additional findings in Devin Review.

[devin-ai-integration]

🔴 Bash tool captured checkpoints lose real agent identity, transcript, and metadata

When the bash tool's attempt_post_hook_capture (bash_tool.rs:1012-1027) prepares a captured checkpoint, it builds a synthetic AgentRunResult with agent_id: { tool: "bash-tool", id: "post-hook", model: "" } and transcript: None. This captured checkpoint's ID is then propagated via captured_checkpoint_id in the real preset's AgentRunResult (e.g., Claude's result with the correct agent identity and transcript). However, when run_checkpoint_via_daemon_or_local (git_ai_handlers.rs:1092-1112) detects a captured_checkpoint_id, it submits only the capture_id and repo_working_dir as a CapturedCheckpointRunRequest, completely discarding the real AgentRunResult. When the daemon executes this captured checkpoint via execute_captured_checkpoint (checkpoint.rs:1025-1035), it uses manifest.agent_run_result — the synthetic bash-tool identity — for authorship attribution. The real AI agent's identity, transcript, and metadata are permanently lost, undermining the core purpose of AI code provenance tracking.

Prompt for agents

The fix requires passing the real agent context into attempt_post_hook_capture so the captured checkpoint manifest stores the correct agent identity and transcript. Specifically:

1. In src/commands/checkpoint_agent/bash_tool.rs, change attempt_post_hook_capture to accept additional parameters for the real agent_id, transcript, and author (or accept the full AgentRunResult). Use these when building the synthetic AgentRunResult at lines 1013-1027, replacing the hardcoded bash-tool/post-hook/empty values.

2. Alternatively, in src/commands/git_ai_handlers.rs at lines 1092-1112, when the early captured checkpoint path is taken, re-write the captured checkpoint manifest on disk to replace the synthetic AgentRunResult with the real one from the preset before submitting to the daemon. This approach avoids changing the bash_tool API but adds disk I/O.

3. A third approach: instead of using CapturedCheckpointRunRequest for bash-tool captures, submit a Live request that includes the real agent_run_result but references the pre-captured file contents. This would require extending the checkpoint execution to support a hybrid mode.

---

Was this helpful? React with 👍 or 👎 to provide feedback.