v0.1.2 — All 3 hooks now fire correctly

Hook fix — Stop and PreToolUse

v0.1.1 fixed the packaging so hooks registered with Claude Code's hook loader. But stop.sh and pre-tool-use.sh still read environment variables (CLAUDE_TRANSCRIPT_PATH, CLAUDE_TOOL_NAME) — Claude Code actually passes hook input as JSON on stdin with fields transcript_path, tool_name, etc.

v0.1.2 fixes both scripts to parse stdin JSON via jq (with python3 fallback).

Hook status — all green

Hook	Status
UserPromptSubmit (signal-detector)	✅ fires (since v0.1.1)
Stop (synthesis)	✅ now fires (fixed in v0.1.2)
PreToolUse (think-first)	✅ now fires (fixed in v0.1.2)

Benchmark (v0.1.1 numbers, n=3 per cell)

Task	Baseline	Plugin	Δ
01-fizzbuzz	27.8±1.4s	16.6±2.1s	-40%
02-bug-fix-surgical	17.8±2.2s	14.7±0.7s	-18%
03-refactor-restraint	25.6±3.1s	26.9±8.8s	+5% (noise)

100% pass rate. Diff sizes identical. Full methodology in BENCHMARK.md.

Upgrade

claude plugins update claude-mind

v0.1.1 — critical hook packaging fix

Critical bug fix

v0.1.0 shipped with non-functional hooks. The plugin manifest used the wrong format and location, so Claude Code's hook loader never registered any of the three hooks (UserPromptSubmit, PreToolUse, Stop). They never fired in any user's environment.

v0.1.1 fixes the packaging. Hooks now register correctly via hooks/hooks.json in the standard Claude Code format used by other working plugins (Warp, superpowers, etc.).

What's now verified

Hook	Fires?
UserPromptSubmit (signal-detector)	✅ Yes — verified via SDK probe (24 entries written to _signals.md) and claude -p smoke test (8 hits)
Stop (synthesis)	❌ Still broken — separate stdin parsing bug, fix in v0.1.2
PreToolUse (think-first)	❌ Still broken — separate stdin parsing bug, fix in v0.1.2

Benchmark (n=3 per task × mode, 18 runs total)

Task	Baseline	Plugin	Δ
01-fizzbuzz	27.8±1.4s	16.6±2.1s	-40%
02-bug-fix-surgical	17.8±2.2s	14.7±0.7s	-18%
03-refactor-restraint	25.6±3.1s	26.9±8.8s	+5% (noise)

100% pass rate in both modes. Diff sizes byte-identical across modes (code-quality claims still unsupported by data).

See BENCHMARK.md for full caveats and methodology.

Upgrade

If you installed v0.1.0:

claude plugins update claude-mind

This is mandatory — v0.1.0's hooks are non-functional.

Roadmap

• v0.1.2: fix Stop and PreToolUse hooks (stdin JSON parsing)
• v0.2 benchmark: n≥10 per cell; bigger tasks; component-level evals

v0.1.0 — initial release

Claude Mind v0.1.0

The opinionated senior-engineer brain for Claude Code.

Synthesizes Karpathy + Chang + Cherny + IndyDevDan + Tan into one tightly-curated plugin.

What's inside

• 10 skills: signal-detector, brain-first, think-first, minimalism, surgical-editing, memory-protocol, uv-python, bun-javascript, taste, using-claude-mind
• 3 hooks: UserPromptSubmit (signal capture), PreToolUse (think-first reminder), Stop (session synthesis)
• 4 commands: /sc-init, /remember, /recall, /ship
• Brain format: compiled-truth + timeline (see docs/brain-format.md)
• Optional bridge to garrytan/gbrain for hybrid retrieval

Install

git clone https://github.com/iamgagan/claude-mind.git ~/.claude/plugins/claude-mind

Compatibility

• macOS, Linux, Windows-with-WSL: full support
• Native Windows (no WSL): hooks skip silently; skills still work

Philosophy

See PHILOSOPHY.md for the five-thinker synthesis.