v0.7.0 install/update wires the legacy blocking PreToolUse gate, not hook_augment (regresses #214)

[cryptomaltese]

Summary

On a clean v0.7.0 install -y --standard (and update), the Claude Code PreToolUse hook written to ~/.claude/hooks/cbm-code-discovery-gate is the old blocking shell gate (exit 2, gate file keyed on bare $PPID) — not the non-blocking augmenter shipped in the binary (src/cli/hook_augment.c, commit c29e6d5). So the exact symptom of #214 returns on every install/update, even though #214 is marked fixed in v0.7.0.

Why it misbehaves (same root as #214)

$PPID is the shell spawned for each hook invocation, so it rotates per tool call. The gate file /tmp/cbm-code-discovery-gate-$PPID therefore never matches a prior one → the "first call per session" check is always true → every Grep/Glob/Read is blocked with exit 2, not just the first.

The installed hook (v0.7.0):

#!/bin/bash
# Gate hook: nudges Claude toward codebase-memory-mcp for code discovery.
# First Grep/Glob/Read/Search per session -> block. Subsequent -> allow.
# PPID = Claude Code process PID, unique per session.
GATE=/tmp/cbm-code-discovery-gate-$PPID
...
touch "$GATE"
echo 'BLOCKED: ...' >&2
exit 2

Repro

1. codebase-memory-mcp install -y --standard (or update) on v0.7.0, Linux, Claude Code.
2. In a session, issue several Read/Grep calls.
3. Each is blocked by the gate (not nudged once) — the claude process reads the hook fresh each call and $PPID differs every time.

Expected

install/update should wire the non-blocking augmenter (the hook_augment path that's already in the binary), so Read is no longer gated and the nudge is emitted at most once per session — the behaviour #214 was closed as delivering.

Note

The binary genuinely has the fix; this is a packaging/install gap — the install template still emits the legacy blocking script. A session_id-keyed (from the PreToolUse stdin JSON) non-blocking augmenter that exit 0s avoids both the $PPID rotation and any ps/process-name assumptions.

Environment

• v0.7.0, Linux (cgroup v2), Claude Code, standard stdio variant.

[DeusData]

Thanks for the detailed write-up — the $PPID analysis is spot-on for the old gate.

I'm having trouble reproducing this on a clean v0.7.0. On main (and in v0.7.0, which already contains the augmenter fix c29e6d5), both install and update write ~/.claude/hooks/cbm-code-discovery-gate as the non-blocking augmenter shim, not the old gate:

#!/bin/bash
# codebase-memory-mcp search augmenter (Claude Code PreToolUse).
BIN="<path>"
[ -x "$BIN" ] || exit 0
"$BIN" hook-augment 2>/dev/null
exit 0

No $PPID, no exit 2; settings.json's PreToolUse command points at that script; the file is written with fopen(..., "w") (truncate, so it overwrites any older gate); and a regression test guards it (cli_hook_gate_script_no_predictable_tmp_issue384 — asserts no PPID, delegates to hook-augment).

To pin down what's different in your environment, could you share:

1. The verbatim contents of your installed ~/.claude/hooks/cbm-code-discovery-gate.
2. codebase-memory-mcp --version.
3. The PreToolUse block (matcher + command) from ~/.claude/settings.json.
4. Was this a fresh install or an upgrade from a pre-v0.7.0 version (roughly which)?

Leading suspicion: a stale PreToolUse entry in settings.json from an older install still pointing at an inline/$PPID gate rather than the rewritten shim — (3) + (4) would confirm it.