fix: create log directory before stderr redirect in hooks

[dmar842]

Problem

The SessionStart and PostToolUse hook commands in hooks/hooks.json redirect stderr to ${CLAUDE_PROJECT_DIR}/.remember/logs/hook-errors.log, but nothing guarantees .remember/logs/ exists first.

In any Claude Code project where the plugin has not yet run, bash emits a hook error like:

/usr/bin/bash: line 1: /c/path/to/project/.remember/logs/hook-errors.log: No such file or directory

The hook scripts (session-start-hook.sh, post-tool-hook.sh) never run. I hit this today on Windows but the same race exists on Linux/macOS — any fresh project activation triggers it.

Fix

Prepend an idempotent mkdir -p to both hook command strings, with 2>/dev/null so any mkdir failure (e.g. read-only filesystem) is silent and doesn't itself fail the redirect it's trying to enable.

Diff

-  "command": "bash \"${CLAUDE_PLUGIN_ROOT}/scripts/session-start-hook.sh\" 2>> ..."
+  "command": "mkdir -p \"${CLAUDE_PROJECT_DIR:-.}/.remember/logs\" 2>/dev/null; bash \"${CLAUDE_PLUGIN_ROOT}/scripts/session-start-hook.sh\" 2>> ..."

Same change applied to both SessionStart and PostToolUse commands.

Test

Reproduced the bug in a Claude Code project dir without .remember/, then applied the patch manually and confirmed the hook error disappears on the next PostToolUse fire and hook-errors.log is created empty as expected.

[Copilot]

Pull request overview

Ensures Claude Code hook stderr redirection won’t fail on fresh projects by creating the .remember/logs directory before appending to hook-errors.log.

Changes:

• Prepend an idempotent mkdir -p "${CLAUDE_PROJECT_DIR:-.}/.remember/logs" to the SessionStart hook command.
• Apply the same directory-creation step to the PostToolUse hook command.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[0reo]

Also hit this on Linux today in a project where .remember/logs/ didn't exist yet — not a Windows-specific issue. The failure chain is self-reinforcing: the SessionStart redirect fails, so session-start-hook.sh never runs, so the mkdir at lines 57–59 of that script never executes, and then every subsequent PostToolUse in the session fails the same redirect. No recovery without manual intervention.

Confirmed the bash behavior locally: bash -c 'true 2>> /nonexistent/x.log' exits 1 with true never executing — the shell opens redirect targets before running the command, so a failed open aborts the whole command.

See also #23, which is an independent report of the same bug filed ~9 hours after this PR went up.

The fix looks right. One small thing worth considering: mkdir -p "${CLAUDE_PROJECT_DIR:-.}/.remember/logs" 2>/dev/null; bash ... swallows mkdir errors but still attempts the bash — which is what you want here (the 2>> will then fail loudly if the mkdir failed for some other reason like ENOSPC, and at least that signal isn't masked).

— Written by Claude Code (Opus 4.7) on behalf of @0reo

[aryeh-evinced]

+1 — hitting this on remember v0.5.0 inside a Claude Code worktree (.claude/worktrees/...). Error seen verbatim:

/bin/sh: /path/to/worktree/.remember/logs/hook-errors.log: No such file or directory

The mkdir -p prepend here is the right fix — the shell resolves the 2>> redirect target before the hook script runs, so any mkdir inside scripts/log.sh is too late.

[mcmanusj-reninc]

+1 from Windows 11 + Git Bash + Claude Code 2.1.116. Error seen verbatim on every PostToolUse in a fresh project with no .remember/ tree:

/usr/bin/bash: line 1: /d/projects/ren/.remember/logs/hook-errors.log: No such file or directory

Ran mkdir -p /d/projects/ren/.remember/logs locally — error disappeared immediately on the next tool call, confirming the diagnosis. Nothing Windows-specific about the bash redirect semantics here; the fix in this PR should resolve it on all three platforms (Linux per @0reo, worktree per @aryeh-evinced, Windows here).

Also filed independently as #23 by another reporter.

[josemoreno801-netizen]

Heads-up — PR #28 (Windows bootstrap fix + TMPDIR portability + README prereqs) lands the same mkdir -p .remember/logs fix as commit 1 of its 3-commit series. Both PRs are correct and address the identical race; merging them as-is would be a no-op double-apply.

Maintainer's call on which to take:

• Take Windows bootstrap fix + TMPDIR portability + README prereqs #28 if you want the broader Windows + docs improvements alongside the hook log dir fix.
• Take fix: create log directory before stderr redirect in hooks #21 (this PR) if you prefer a narrower, surgical fix and want to handle Windows separately. In that case Windows bootstrap fix + TMPDIR portability + README prereqs #28 would need a quick rebase to drop the duplicate commit.

Either way, only one should land. No harm caused by this PR — just flagging the overlap so the maintainer can pick a path.

(Drive-by comment from a downstream user who hit the same race on macOS — appreciate the fix.)

[fdaviddpt]

Closing in favor of #33 which takes a more architectural approach — centralizing directory creation in bootstrap-dirs.sh instead of inlining mkdir in hooks.json.

Thank you @dmar842 for the clean diagnosis and fix! Your PR description nailed the root cause (bash opens the 2>> target before the script runs) and directly inspired our solution. The community benefits from people who write clear bug reports with reproduction steps. Much appreciated!