Skip to content

Hooks adopt <cwd>/data as a legacy data dir, so the capture spool is written inside the user's project #1810

Open

Description

@masa6161

Summary

resolveDataDir() adopts <cwd>/data as a legacy store when it contains state_store.db or iii-config.yaml. The check came in with #314 (the fix for #303), the only commit that has touched src/cli-data-dir.ts, and its test treats it as data-dir migration. The capture spool (#1460) resolves its folder through the same function, with the agent session's working directory as cwd. So in a project that still has such a folder, the hooks write their capture spool into <project>/data/capture-spool, even when the daemon uses a different data dir and never reads that folder.

In my case the folder had the layout described in #303 (state_store.db and stream_store, both last modified on 2026-06-04, before #303 was closed on 2026-07-25). The daemon uses %USERPROFILE%\.agentmemory\data.

Environment: agentmemory 0.9.33 (npm), iii 0.25.1, Windows 11, Claude Code hooks, daemon started with --data-dir %USERPROFILE%\.agentmemory\data. The code references are from main (54bbb7cd351e).

Why it happens

  1. src/cli-data-dir.ts:114-120: the legacy check runs before the platform default and uses cwd. For hooks that is process.cwd(), the agent's working directory.
  2. spoolDir() (src/capture/spool.ts) resolves the data dir this way when AGENTMEMORY_CAPTURE_SPOOL_DIR and AGENTMEMORY_DATA_DIR are unset, which is the normal setup for hooks.
  3. test/cli-data-dir.test.ts ("adopts legacy data without sharing it with another instance", marker comment "Legacy marker recognized during data-dir migration") expects <repo>/data to be adopted. I found nothing that limits the adoption to the process that owns the store.

What I observed

  • From 2026-11-09 to 2026-11-17, hook runs whose working directory was the project root spooled 828 records into <project>/data/capture-spool (the cumulative counter in its stats file), including the records from an outage on 2026-11-17. Hook runs from other working directories used %APPDATA%\agentmemory\capture-spool, so the spool was split by working directory.
  • The spooled records contain tool input and output (after the hooks' redaction). They ended up inside the project tree, which in my case is a cloud-synced folder. data/ is also a common name for a project's own files; in my project, the build packages that folder.
  • A hook-side drain only drains the folder its own working directory resolves to, so the two folders were drained independently. The daemon's boot drain read neither of them (Windows: with a non-default data dir, the boot-time spool drain misses the hooks' default spool because APPDATA is not in the worker environment #1809).
  • 43 files named local-3111.stats.json.<pid>.tmp were left in <project>/data/capture-spool. writeStats() (src/capture/spool.ts:228-235) swallows a failed renameSync and leaves its temp file behind. I assume the renames failed because the sync client held local-3111.stats.json, but I did not capture the error.

Suggested change

  • Do not adopt <cwd>/data in the hooks' spool resolution. Spool to the platform default, AGENTMEMORY_DATA_DIR or AGENTMEMORY_CAPTURE_SPOOL_DIR. Hooks run in arbitrary project directories, and the daemon reads <cwd>/data only if it was itself started there and adopted it. Keep the adoption for the CLI, where fix(cli): add --data-dir flag + AGENTMEMORY_DATA_DIR so engine state lives outside repos #314 put it.
  • Smaller, related: in writeStats(), remove the temp file when the rename fails.

I have not built or run a patched version.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions