qdf-hook

Stop paying for the same tokens twice. qdf-hook is a set of Claude Code hooks that intercept tool output and collapse the redundant, repeated, and re-read parts before they ever reach the model's context — cutting 50–99 % of the tokens your dev tools would otherwise burn.

Re-read an unchanged 500-line file   →  not read at all  (−100 %)
1000-row JSON from a Bash command     →  ~300 chars        (−99.5 %)
Repeated command output (any tool)    →  §ref token        (−98.2 %)
go test -v, 50 tests pass             →  3 lines           (−99 %)

Measured across 18,000+ real tool calls in day-to-day Claude Code usage: 88 % of tool-output tokens eliminated — 34 MB of raw output collapsed to 4 MB, about 7.9M tokens saved. Every hot path runs in tens of microseconds — a sub-1 % sliver of the process-spawn cost, so the savings are effectively free. An optional resident daemon (qdf-hookd) cuts the per-hook cost further still: a warm daemon roundtrip is ~59.7 µs versus ~6.97 ms for a fresh CLI spawn — about 117×.

qdf-hook stats — real numbers from one machine's usage. Run it anytime to see yours.

• ⚡ Fast — re-read of an unchanged file: 49 µs; a §ref cache hit: 3.6 µs; JSON analysis of 1000 rows: 263 µs; warm daemon roundtrip: ~59.7 µs (vs ~6.97 ms for a fresh CLI spawn).
• 🧠 Context-aware — tracks what the model already saw this session and across sessions and serves a reference instead of the bytes.
• 🔒 Safe — never changes what a tool does; only compresses what the model sees. A cache miss is always a fresh, correct read.
• 🚀 Two-command setup. brew install, then qdf-hook init to wire the hooks — nothing to hand-edit. After that the resident daemon starts (and self-upgrades) on its own; there's nothing to run by hand.

---

Quick start

Three steps — and steps 2 and 3 are required. brew install only drops the binaries; nothing compresses until you run qdf-hook init and restart.

# 1. install — Homebrew places BOTH the qdf-hook daemon/CLI and the qdf-hookc client
brew install alex60217101990/tap/qdf-hook

# 2. REQUIRED — wire the hooks into Claude Code (idempotent; preserves your other settings)
qdf-hook init

# 3. REQUIRED — restart Claude Code so it loads the hooks

After the restart you're done. The resident daemon starts (and version- upgrades) itself on every session — nothing else to run. Confirm it's working: qdf-hook stats.

⚠️ Skipping qdf-hook init (or the restart) = zero compression. That's the one step people miss — brew install alone does not activate anything.

macOS: “cannot verify developer” on first run

The binaries are ad-hoc–signed (not Developer ID). Homebrew usually clears the download quarantine, but if Gatekeeper still blocks one, allow it once:

xattr -d com.apple.quarantine "$(command -v qdf-hook)" "$(command -v qdf-hookc)" 2>/dev/null || true

or approve it under System Settings → Privacy & Security.

Upgrade later with brew upgrade qdf-hook (re-run qdf-hook init once after a major upgrade to refresh the hook wiring). Prefer Go? go install github.com/alex60217101990/terse/cmd/qdf-hook@latest installs the daemon/CLI; the native qdf-hookc client then comes from a release asset or make install.

qdf-hook init merges every hook into your global ~/.claude/settings.json (or qdf-hook init --project for the current repo's .claude/settings.json), using the binary's absolute path so it resolves without PATH juggling. Re-running is a no-op; your other hooks and settings are left untouched (and backed up).

Check it's paying off after a few tool calls:

qdf-hook stats            # aggregate token savings this session

Manual configuration (if you'd rather edit settings.json yourself)

A single catch-all PostToolUse hook now routes every tool through post, which dispatches internally — new tools need no config change. PostToolUse and PreToolUse talk to the resident daemon first through the tiny native client qdf-hookc (installed next to qdf-hook), falling back to the plain CLI subcommand when the daemon isn't up or the client is absent:

{
  "hooks": {
    "PreToolUse": [
      {"matcher": "Read", "hooks": [{"type": "command", "command": "qdf-hookc ~/.qdf-hook/d.sock 2>/dev/null || qdf-hook pretooluse"}]}
    ],
    "PostToolUse": [
      {"matcher": ".*", "hooks": [{"type": "command", "command": "qdf-hookc ~/.qdf-hook/d.sock 2>/dev/null || qdf-hook post"}]}
    ],
    "PreCompact":   [{"matcher": ".*", "hooks": [{"type": "command", "command": "qdf-hook precompact"}]}],
    "PostCompact":  [{"matcher": ".*", "hooks": [{"type": "command", "command": "qdf-hook postcompact"}]}],
    "SessionStart": [{"matcher": "",   "hooks": [{"type": "command", "command": "qdf-hook daemon --ensure"}]}]
  }
}

---

Why it saves tokens

Every byte a tool returns to Claude Code enters the model's context and stays there for the rest of the session — you pay for it again on every turn. The three biggest offenders in a normal dev session:

1. Re-reading files. Read main.go, edit elsewhere, read it again — the full file lands in context twice.
2. Repeated command output. go test, git status, a deploy log — run again, and the identical output is re-ingested in full.
3. Structured & noisy dumps. A 1000-row JSON array, a verbose test log, or a progress bar redrawing 500 times carries far more bytes than information.

qdf-hook attacks all three (see docs/ARCHITECTURE.md for the deep dive):

• A PreToolUse interceptor denies re-reading a file whose mtime and size are unchanged — the file is never read, so the tokens are never spent.
• A content-addressed §ref store replaces any byte-identical repeated output (any command, any tool, across sessions) with a 13-token reference.
• Structural summarizers turn JSON arrays, go test -v, git log, and benchmark output into compact schemas and counts.
• A Grep grouper collapses matches per file; a squeezer strips ANSI and run-length-collapses repeated lines; a delta engine serves a unified diff when a re-read file changed only slightly.

The only lever local tooling has is how many text tokens enter the context window — that is exactly what qdf-hook minimizes.

---

Benchmarks

Full methodology and raw numbers: docs/BENCHMARKS.md.

Token savings (measured through the real binary)

Scenario	Before	After	Reduction
Re-read unchanged 500-line file (PreToolUse deny)	500 lines	0 (not read)	−100 %
Re-read unchanged file (PostToolUse)	~12 KB	~150 B	~−99 %
1000-row JSON array from Bash	~60 KB	~300 B	−99.5 %
Repeated unstructured output (7 KB log)	7 179 B	131 B	−98.2 %
go test -v, 50 tests pass	300 lines	3 lines	−99 %
Write a 500-line file	500 lines	1 line	−99.8 %
Realistic mixed session (7 ops)	216.9 KB	60.2 KB	−72.3 %

Latency (per hook op, Apple Silicon; benchstat, n ≥ 12)

Operation	Time	Allocs
PreToolUse, unchanged file (deny)	49 µs	54
§ref cache hit	3.6 µs	12
Session state Save + Load	69 µs	29
JSON analysis (1000 rows)	269 µs	2 068
Daemon roundtrip (warm qdf-hookd)	~59.7 µs	—
CLI roundtrip (fresh process spawn)	~6.97 ms	—

Launching the hook process itself costs Claude Code ~1–6 ms of spawn + runtime init; qdf-hook's own work is a sub-1 % sliver of that. The resident daemon (qdf-hookd) removes the spawn cost entirely for warmed-up sessions: a request answered from its shared in-RAM store is ~117× faster than a fresh CLI invocation of the same pipeline.

---

How it works

System topology — every Claude Code tool event flows through a hook that prefers the resident daemon and falls back to the one-shot CLI:

flowchart LR
    CC[Claude Code]
    subgraph hook["hook command per tool event"]
      direction LR
      QC["qdf-hookc sock"] -.->|fallback if down| CLI["qdf-hook CLI"]
    end
    CC -->|tool JSON on stdin| hook
    QC ==>|warm ~60µs| D["qdf-hookd<br/>resident daemon<br/>in-RAM store"]
    CLI -->|cold ~7ms| P["compression pipeline<br/>noise-strip · summarize<br/>§ref · delta · squeeze"]
    D --> P
    P -->|fewer tokens back| CC
    D <--> C[("~/.qdf-hook<br/>refs · sessions · analytics")]
    CLI <--> C

Loading

Per-tool pipeline — what each tool's output becomes:

flowchart LR
    R[Read] --> PRE{PreToolUse}
    PRE -->|mtime+size match| DENY["deny — §unchanged§ (file not read)"]
    PRE -->|changed / new| RD[read handler]
    RD -->|first read| CACHE[cache + full]
    RD -->|changed| DELTA[unified diff]
    RD -->|unchanged| U["§unchanged§"]
    B[Bash] --> DET{detect}
    DET -->|structured| SUM[JSON/test/log summary]
    DET -->|repeat| REF["§ref:HASH§"]
    DET -->|noisy| SQ[ANSI strip + RLE]
    G[Grep] --> GRP[group by file + cap]
    GL[Glob] --> TREE[directory tree]
    WR[Write/Edit] --> SUP[suppress echo, cache file]

Loading

• Read — PreToolUse denies unchanged re-reads (mtime+size); PostToolUse serves §unchanged§, a unified diff, or full content on first read. Windowed reads (an offset/limit or a partial startLine/numLines window) always pass through uncached — caching a slice would poison delta tracking for the whole file.
• Bash — structural summarizers → §ref dedup of repeats → ANSI/RLE squeeze → pass through.
• Grep — group content matches per file (capped) or delegate a file list to the tree compressor.
• Glob — flat list → directory tree. Write/Edit — suppress echo, cache the real file for the next delta (Write/Edit/MultiEdit tool responses carry no plain-text field — the hook resolves originalFile/structuredPatch instead). Compaction — force full re-read + file manifest.

Resident daemon (qdf-hookd)

Every hook invocation above still pays Claude Code's exec cost — spawning a fresh process and re-decoding the on-disk cache — even though qdf-hook's own logic runs in tens of microseconds. qdf-hookd removes that spawn: it's a long-lived process, started (and version-checked) at SessionStart via qdf-hook daemon --ensure, that answers hook requests over a unix socket (~/.qdf-hook/d.sock) against one shared in-RAM state store instead of a per-invocation disk round trip.

The PostToolUse/PreToolUse hooks are a hybrid client: qdf-hookc <sock> 2>/dev/null || <exe> <sub> talks to the daemon when it's up and falls back to the plain CLI when it isn't. qdf-hookc is a tiny native AF_UNIX socket client (a single C file cross-compiled with zig cc, shipped as a release asset and in the qdf-hookc container image) installed next to qdf-hook; it does one thing — stream stdin to the socket and the reply back — with no nc portability quirks. If it's missing (e.g. Windows, which has no native client), the shell || runs <exe> <sub> directly. A PING handshake carrying the binary's version lets daemon --ensure detect and replace a stale daemon after an upgrade; QUIT requests a clean shutdown; an idle daemon exits on its own after 30 minutes. The daemon restores GOMAXPROCS(NumCPU) and the garbage collector on start — the one-shot CLI disables both for startup speed, but a long-lived process needs them.

Every persisted store is a rebuildable cache written with a single plain write and serialized with qdf (~38× faster decode than JSON); decode and hashing are zero-copy. A §ref is emitted only when the current output's hash already has a stored blob, so it can never point at stale or wrong content. The refs/ and rerun-delta last/ stores are bounded: gc (run automatically, throttled to at most once per 24h at SessionStart, and on a 10-minute sweep inside the daemon) evicts the lowest-utility blobs — hits × exponential time-decay — once total size exceeds --cache-max-size (default 128 MiB) or an entry is older than --cache-ttl (default 720h / 30 days).

---

Subcommands

Subcommand	Purpose
init	Install all hooks into Claude Code settings.json (idempotent)
post	Universal PostToolUse hook — routes any tool through the pipeline
pretooluse	Deny re-read of an unchanged file (mtime+size)
read	§unchanged§ / unified-diff compression (standalone; superseded by post)
bash	Structured summaries + §ref dedup + ANSI/RLE squeeze (standalone; superseded by post)
write	Suppress content echo, prime delta cache (standalone; superseded by post)
glob / grep	File tree / grouped matches (standalone; superseded by post)
precompact / postcompact	Full re-read + file manifest around compaction
sessionstart	SessionStart hook logic (invoked via daemon --ensure)
daemon --serve	Run the qdf-hookd serve loop in the foreground
daemon --ensure	Start/refresh qdf-hookd, replacing a stale version; throttled disk-cache gc
stats	Token-savings + latency analytics (--json, --style line|braille)
gc	Evict low-utility sessions + bounded refs//last/ blobs (--dry-run)
expand <hash>	Print the full content behind a §ref:HASH§ token
version	Print version

Global flags: --cpuprofile FILE, --memprofile FILE.

qdf-hook init installs a single catch-all PostToolUse hook (post) plus the PreToolUse/SessionStart hooks; the per-tool subcommands (read, bash, write, glob, grep) still work standalone but are superseded by post and pruned from settings.json on re-init.

Configuration

Variable	Default	Effect
QDF_CACHE_MAX_SIZE	128 MiB	Combined size cap for refs/ + last/ before gc evicts
QDF_CACHE_TTL	720h (30 d)	Max age of a cache entry before gc prunes it regardless of size
QDF_DECAY_LAMBDA	0.1	Decay rate for utility-score eviction (sessions and blobs)
QDF_SKIP_TOOLS	(none)	Comma-separated tool names to always pass through verbatim

--cache-max-size and --cache-ttl flags on daemon and gc override the env vars for a single invocation.

State lives under ~/.qdf-hook/ (sessions/, refs/, last/, d.sock, analytics.jsonl) and is a rebuildable cache — safe to delete anytime (the daemon will restart itself and the caches will refill).

FAQ

Does it change what my commands do? No. Hooks run after a tool executes, or only deny a redundant re-read. Never mutates files or command behavior.

A §ref blob I need is gone? qdf-hook expand <hash>. In practice the model rarely needs it — a §ref means it already saw that exact output.

Safe across crashes / concurrent hooks? Yes — state is content-addressed and rebuildable; a torn write is treated as a cache miss. A daemon connection that panics is recovered per-connection and just closes without a reply, so the hybrid client's CLI fallback covers it.

Do I need to manage the daemon myself? No. SessionStart runs daemon --ensure, which starts it if it's not running, replaces it if it's a stale version, and is a no-op otherwise. It exits on its own after 30 minutes idle. To stop it early, send it QUIT (e.g. printf 'QUIT\n' | qdf-hookc ~/.qdf-hook/d.sock) — it flushes and shuts down cleanly.

What if qdf-hookc isn't installed? The hybrid command falls back to the plain CLI subcommand whenever the native client is missing or fails to connect — you lose the daemon speedup, not correctness. qdf-hookc ships as a release asset (and container image) per OS/arch; qdf-hook init wires it in only when it's found next to the qdf-hook binary.

Plays nice with other hooks (sqz, atuin, …)? Yes; qdf-hook init preserves them. They compose.

Contributing

See CONTRIBUTING.md. The one hard rule: measure before you optimize — every perf change is gated by an interleaved benchstat run.

go test -race ./...
go test -bench=. -benchmem -count=6 ./...

License

Apache 2.0 © 2026 the qdf-hook authors.