squeez

End-to-end token optimizer for seven AI CLI hosts — Claude Code, GitHub Copilot CLI, OpenCode, Gemini CLI, OpenAI Codex CLI, Pi, and Hermes. Compresses bash output up to 95%, collapses redundant calls, preserves exact identifiers through every summary, refuses net-loss compressions, and injects a terse prompt persona — automatically, with zero new runtime dependencies.

---

📰 What's new

Latest first. Full history in CHANGELOG.md.

v1.46.0 — shipped filter pack + native git status

The filter DSL could always cover the long tail GenericHandler can't compress — it just shipped empty, so every user started from a blank filters.ini. This release fills it, adds a native git status path, and makes two guards that were only ever measured in benchmarks run for real.

• 41 built-in filter rules (assets/filters_builtin.ini) — pip/uv/poetry, bundle/composer/rubocop/phpstan, dotnet/swift/gcc, mypy/basedpyright/golangci-lint, shellcheck/hadolint/yamllint/markdownlint/pre-commit, systemctl/journalctl/rsync/ping/ansible-playbook and more. Every rule ships an inline self-test that CI gates on. Project and user rules still shadow any built-in; builtin_filters = false turns the pack off.
• git status reads as porcelain — re-run as git status --porcelain=v1 -b and rendered as a working-tree summary. Exact paths are never collapsed to counts, raw porcelain codes never reach the model, and in-progress rebase/merge/bisect (which porcelain omits) is read from .git/ markers instead of a second git status.
• The preservation guard actually runs. economy::preservation used to execute only in benchmark. It now scores anchor survival at wrap time on ≥90%-reduction calls; below the floor the verbatim original is stashed and the header carries [anchors: N%].
• The net-win gate counts its own marker. The ~40-token retrieve marker was appended after the gate decided the call was a win, so a call saving 25 tokens could ship a 40-token marker and still claim a win. It is now inside the arithmetic.
• Dispatch blind spots closed — TF_LOG=debug terraform plan and npx terraform plan reach the terraform branch; wrappers peel in a loop (sudo npx vitest); a pipeline ending in a transforming stage (cargo build | grep error) dispatches on that stage.
• Benchmarks grew to 46 scenarios, six of them covering the new rule pack, so the pack's contribution is measured rather than asserted.

---

Install

Three methods — all produce the same result (binary at ~/.claude/squeez/bin/squeez, hooks registered).

curl (recommended)

curl -fsSL https://raw.githubusercontent.com/claudioemmanuel/squeez/main/install.sh | sh

Windows: requires Git Bash. Run the command above inside Git Bash — PowerShell/CMD are not supported.

npm / npx

# Install globally
npm install -g squeez

# Or run once without installing
npx squeez

Downloads the correct pre-built binary for your platform (macOS universal, Linux x86_64/aarch64, Windows x86_64). Requires Node ≥ 16.

cargo (build from source)

cargo install squeez

Builds from crates.io. Requires Rust stable. On Windows you also need MSVC C++ Build Tools.

---

Supported hosts

squeez setup auto-detects every CLI present on disk and registers the hooks. squeez uninstall removes them. Session data and config.ini are preserved so reinstall is lossless.

Host	Memory file	Bash wrap	Session memory	Budget inject (Read/Grep)	Notes
Claude Code	~/.claude/CLAUDE.md	✅ native	✅ native	✅ native	Restart Claude Code to pick up hooks
Copilot CLI	~/.copilot/copilot-instructions.md	✅ native	✅ native	✅ native	Restart Copilot CLI after setup
OpenCode	~/.config/opencode/AGENTS.md	✅ native	✅ native	✅ native	Plugin at ~/.config/opencode/plugins/squeez.js; MCP tool calls skip hooks (upstream sst/opencode#2319)
Gemini CLI	~/.gemini/GEMINI.md	✅ native	✅ native	🟡 soft via GEMINI.md	BeforeTool rewrite schema pending upstream docs (google-gemini/gemini-cli#25629)
Codex CLI	~/.codex/AGENTS.md	✅ native	✅ native	🟡 soft via AGENTS.md	apply_patch hooks landed in 0.123.0 (#18391); updatedInput + read_file/grep hook surface still pending (openai/codex#18491)
Pi	~/.pi/agent/skills/squeez/SKILL.md	✅ native	✅ via skill	✅ native	TypeScript extension at ~/.pi/agent/extensions/squeez/index.ts; restart Pi after setup
Hermes	~/.hermes/profiles/default/SOUL.md	✅ native	✅ native	❌ no budget surface	Python plugin (__init__.py + plugin.yaml) auto-discovered by Hermes; detected via ~/.hermes/

Manage

squeez setup                  # register into every detected host
squeez setup --host=<slug>    # register into one host
squeez uninstall              # remove squeez entries from every detected host
squeez uninstall --host=<slug>

Slugs: claude-code / copilot / opencode / gemini / codex / pi / hermes.

After install, restart the CLI you use to pick up the new hooks.

Uninstall

squeez uninstall              # preserves session data + config.ini
bash ~/.claude/squeez/uninstall.sh   # (legacy) full wipe, if the script exists

Self-update

squeez update             # download latest binary + verify SHA256
squeez update --check     # check for update without installing
squeez update --insecure  # skip checksum (not recommended)

---

What it does

Feature	Description
Bash compression	Intercepts every command via PreToolUse hook, applies smart filter → dedup → log-template → relevance-truncation. Up to 95% reduction. (Per-dir grouping stays reserved for git status, where the path list is the payload.)
Reversible compression	When a large output is compressed, the verbatim original is stashed in a content-addressed blob and a [squeez: … call squeez_retrieve with key="<id>"] marker is emitted. The model recovers any dropped detail via the squeez_retrieve MCP tool — so compression can be aggressive without losing information. TTL-pruned, zero-dep.
Log-template compaction	Collapses near-identical log lines that differ only by a timestamp, id, hex hash, or 1ms-style value into one [×N] <template> line — what dedup (exact-only) leaves on the table.
Relevance-aware truncation	When the generic handler must truncate, it keeps the highest-relevance lines (error/signal words + terms drawn from the command) instead of a blind head — so a buried error survives.
Context engine	Cross-call redundancy with two paths: exact-hash match (FNV-1a, fast) and fuzzy trigram-shingle Jaccard ≥0.85 (whitespace, timestamps, single-line edits no longer defeat dedup).
Summarize fallback	Outputs exceeding 500 lines are replaced with a ≤40-line dense summary (top errors, files, test result, tail). Benign outputs get 2× the threshold so successful builds stay verbatim.
Identifier factsheet	Exact identifiers in the region a summary drops — git SHAs, UUIDs, ticket codes, versions, large ids — ride along in every dense summary (ids_preserved: / "ids":[...]). Deterministic, budget-capped (16 facts / 256 chars), so lossy summarization never silently loses a hash or ticket number.
Net-win gate	The # squeez header itself costs ~15–25 tokens. When applied compression saves less than net_win_min_tokens (default 24), the call becomes a verbatim passthrough with zero savings recorded — no net-loss invocations.
Adaptive intensity	Truly adaptive: Full (×0.6 limits) below 80% of token budget, Ultra (×0.3) above. Used to be always-Ultra; now actually responds to session pressure.
MCP server	squeez mcp runs a JSON-RPC 2.0 server over stdio exposing 17 tools (16 read-only session-memory queries + squeez_retrieve to expand a compressed output) so any MCP-compatible LLM can query session memory directly. Hand-rolled, no mcp.server dependency.
Built-in filter pack	41 shipped filter-DSL rules (assets/filters_builtin.ini) covering the long tail with no dedicated handler — pip/uv/poetry, bundle/composer/rubocop/phpstan, dotnet/swift/gcc, mypy/basedpyright/golangci-lint, shellcheck/hadolint/yamllint/markdownlint/pre-commit, systemctl/journalctl/rsync/ping/ansible-playbook, and more. Every rule carries an inline self-test that squeez filter-test and CI gate on. A project or user rule of the same name shadows a built-in; builtin_filters = false turns the pack off.
git status porcelain	git status is re-run as git status --porcelain=v1 -b and rendered as a working-tree summary (branch, ahead/behind, staged/unstaged/untracked/conflicted). Exact paths are preserved, never collapsed to counts, and raw porcelain codes never reach the model. In-progress rebase/merge/bisect/cherry-pick — which porcelain omits — is read from .git/ markers rather than a second git status run. Yields to any output flag you passed yourself; degrades to raw output on an unrecognized shape.
Preservation guard	On calls that reduce by ≥90%, the surviving fraction of navigation anchors (file paths, file:line refs, error markers, test verdicts) is scored at wrap time. Below preservation_floor (0.70) the verbatim original is stashed even outside the usual size gates and the header carries [anchors: N%] — so over-compression costs a marker, not a re-investigation.
Config CLI + /squeez	squeez config get/set/list/reset/path reads and writes config.ini safely (schema validation, comment-preserving writes). squeez setup installs a /squeez slash command that drives it in natural language from inside the session.
Post-compact re-injection	After /compact, the PostCompact hook re-injects squeez's tracked session state (recent files, error snippets, git refs, retrievable blob ids) as additionalContext — so concrete state survives compaction instead of being re-discovered.
Bash-wrap safety	Risky commands (rm -rf, git push --force, npm publish, … — configurable bash_risk_patterns) and bypassed commands run unwrapped, so the host's native permission rules evaluate the original command. wrap_bash = false disables wrapping entirely. See SECURITY.md.
Token estimate	Compression-timing decisions use a content-class calibrated estimate: output is classified Dense/Prose/Mixed and counted at chars/2.0, chars/3.7, or the code- and CJK-aware char-class estimator — dense tool output really runs ~1.9 chars/token in production, not chars/4. Flat legacy path stays available via class_density = false.
Auto-teach payload	squeez protocol (or the squeez_protocol MCP tool) prints a 2.4 KB self-describing payload — the LLM learns squeez's markers and protocol on first call.
Caveman persona	Injects an ultra-terse prompt at session start so the model responds with fewer tokens.
ADHD focus mode	focus = adhd shapes output structure rather than length: next action first, numbered steps, one-line state restatement, advisories capped at 5, squeez doctor sorted failures-first with a single closing command. Stacks with any persona.
Memory-file compression	squeez compress-md compresses CLAUDE.md / AGENTS.md / copilot-instructions.md in-place — pure Rust, zero LLM. i18n-aware: set lang = pt (or --lang pt) for pt-BR article/filler/phrase dropping and Unicode-correct matching.
Session memory	On SessionStart, injects a structured summary of the previous session: files investigated, learned facts (errors + git events), completed work (builds, test passes), and next steps (unresolved errors, failing tests). Summaries carry temporal validity (valid_from/valid_to).
Token tracking	Every PostToolUse result (Bash, Read, Grep, Glob, Monitor, SubagentStop) feeds a SessionContext so squeez knows what the agent has already seen. Read/Grep/Glob/Monitor outputs are also rewritten via updatedToolOutput (Claude Code v2.1.119+) when content is redundant or oversized.
Token economy	Sub-agent cost tracking (~200K tokens/spawn), burn rate prediction ([budget: ~N calls left]), session efficiency scoring, tool result size budgets.
Auto-calibration	squeez calibrate runs benchmarks on install and generates an optimized config.ini (aggressive / balanced / conservative profiles).

---

How squeez compares

There are now several token-reduction tools targeting AI coding CLIs. They make different bets — the right one depends on what you care about: zero deps, lossless filtering, structural reformatting, or task-conditioned ML.

Tool	Approach	Hosts	Deps	Key wins	Trade-off
squeez (this project)	Hook + filter pipeline + context engine (MinHash dedup, log-template, relevance truncation, summarize + identifier factsheet, adaptive intensity, net-win gate) + reversible compression (retrieve) + MCP server	Claude Code, Copilot CLI, OpenCode, Gemini CLI, Codex CLI, Pi, Hermes	Zero runtime deps (libc only on Unix)	Up to 95% on bash; cross-call dedup; reversible squeez_retrieve; exact ids (SHAs/UUIDs/tickets) survive every summary; signature-mode for source files; TOON re-encoder (incl. nested JSON); 17 MCP tools; post-compact state re-injection; cache-aware savings proof; enterprise (Bedrock/Vertex) USD-saved estimate	Heuristic, not ML — no per-task understanding
chopratejas/headroom	Library + HTTP proxy + MCP; compresses tool output, logs, RAG chunks and conversation history at the API layer with real tokenizers and ML (Kompress/Magika)	Any (OpenAI/Anthropic/Bedrock/Vertex via proxy)	Python + Rust; PyTorch/HF models	60-95%; reaches conversation history (the biggest sink) via the proxy; reversible CCR; image compression	Heavier (proxy + ML deps); not a zero-dep drop-in hook. squeez adopts its reversible-retrieve and post-compact ideas within the zero-dep hook model.
teamchong/pxpipe	Local API proxy that renders bulky request context (system prompt, tool docs, old history, big tool_results) as dense PNG images — image tokens are priced by pixel area, so dense text packs ~3.1 chars/image-token vs ~1 as text	Claude Code (/v1/messages), Codex (/v1/responses)	Node/pnpm; canvas at build time	~59–70% end-to-end bill cut on Fable 5 traffic; reaches surfaces hooks can't (system prompt, tool docs, history); rigorous per-request count_tokens counterfactual measurement	Lossy: exact-string recall from images fails on most models (0/15 verbatim on Opus 4.8; silent confabulation, not errors) — default scope is Fable 5 only. Complementary to squeez, not competing: squeez compresses losslessly at the hook layer before content enters context; pxpipe cheapens what remains on the wire. They stack — squeez v1.35.0's factsheet, class-density and cache-aware accounting are ported from its measurement discipline.
rtk-ai/rtk	Hook proxy that rewrites bash commands (git status → rtk git status), then compresses 100+ command outputs	Claude Code, Cursor	Zero deps (Rust)	60-90% on 100+ commands; rtk read -l aggressive for signature mode	Rewriting a command changes what runs, not just how its output is printed. Compression also has a floor: on low-entropy content there is little to remove, and a transformed format can cost more than it saves.
KRLabsOrg/squeez	Task-conditioned ML (Qwen 2B / ModernBERT 150M) — pipe tool output + task description, get back only relevant lines	Any (CLI tool)	Python, PyTorch / vLLM server	92% compression, F1 0.80; task-aware (same log slices kept differently per query)	Requires running an LLM locally; not zero-dep. Same project name, different design.
ojuschugh1/sqz	CLI context compressor	Any	Python	Single-command compression	Lower coverage than the others.
LLMLingua-2 (Microsoft)	Neural prompt compressor that removes 50-80% of a prompt while preserving meaning	API / library	Python, transformers	Strong on long static prompts	Latency + model dep; not a CLI hook.
TOON	Schema-aware JSON replacement (users[100]{id,name,role}:) — ~40% fewer tokens on arrays of uniform objects	Library, not a CLI	TypeScript SDK	Lossless on the right shape; squeez embeds a TOON encoder for gh/kubectl/aws/gcloud/az JSON outputs	Only helps on uniform JSON shapes.

If you want a CLI hook that just works, never needs a Python runtime, and never silently inflates your output tokens, squeez is the safe default. If you can run an LLM next to your shell and want task-aware filtering, KRLabsOrg/squeez is worth a look as a complement. The two squeez projects share a name but are independent.

---

Scope & Limits

squeez optimizes what it can reach — the surfaces exposed by each host's hook API. It cannot fix token leaks outside those surfaces.

Coverage table

Surface	How	When	Supported hosts
Bash stdout/stderr	PreToolUse wraps command w/ the filter pipeline (smart-filter → dedup → log-template → relevance-truncation; original stashed for squeez_retrieve). Risky/bypassed commands run unwrapped under native permission rules.	Every Bash invocation	all 5
Read / Grep / Glob limits	PreToolUse injects limit / head_limit per read_max_lines / grep_max_results	Every Read/Grep/Glob call	Claude Code, Copilot, OpenCode (hard); Gemini + Codex soft via GEMINI.md / AGENTS.md
Read / Grep / Glob / Monitor output rewrite	PostToolUse runs squeez compress-output and returns updatedToolOutput when content is redundant or oversized	Claude Code v2.1.119+	Claude Code
Agent / Task prompt	PreToolUse compresses tool_input.prompt (markdown-aware, via compress-prompt)	When prompt > agent_prompt_max_tokens	Claude Code (post–v1.8.0)
Sub-agent output	SubagentStop hook feeds last_assistant_message into SessionContext for cross-call dedup	On every sub-agent completion	Claude Code
Compaction lifecycle	PreCompact logs the event; PostCompact re-injects tracked session state (files, errors, git refs, retrievable blob ids) as additionalContext so it survives compaction	On context compaction	Claude Code
Session memory	SessionStart injects prior session summary + file-access cache	Once per session start	all 5
Markdown viewing	Bash handler routes .md reads through compress-md when auto_compress_md=true	Viewer commands on .md paths	all 5

What squeez CANNOT compress

Agent/Task returned output. No hook API surface exists to rewrite an Agent's return value. PostToolUse updatedToolOutput (Claude Code v2.1.119+) covers built-in tools (Read, Grep, Glob, Monitor) but not the Agent/Task result. Workaround: keep agent prompts compact (squeez compresses at dispatch time via PreToolUse), and use squeez_agent_costs MCP tool to monitor spawn overhead.

Skills & slash-command files. Claude Code loads these into the system prompt before any hook fires. squeez has no visibility into session-start system prompt construction.

User's top-level prompt. squeez runs per tool call, not on user turns.

Tools whose host doesn't expose PreToolUse / BeforeTool. E.g. Codex apply_patch hooks landed in 0.123.0, but updatedInput is explicitly unsupported and read_file/grep still have no hook surface (openai/codex#18491) — so Read/Grep caps for Codex are soft hints in AGENTS.md, not hard injections.

Secondary wins (not compression, but token-saving)

• Cross-call redundancy dedup — exact-hash and fuzzy-trigram collapsing across 16 recent calls (see Context engine)
• Skill re-injection dedup — when the same skill body is injected by the Skill tool more than once in a session, the repeat collapses to [squeez: identical to Skill #N]. Keyed by body hash in a session-long store (not the 16-call window), so it fires even when injections recur far apart
• File-access cache — subsequent Bash commands trimmed when re-reading a file squeez has already fingerprinted
• Burn-rate warnings — [budget: ~N calls left] nudges so the user changes behavior before context pressure spikes

Reducing overall session cost

squeez cannot automate these, but you can:

• Fewer Agent/Task dispatches per session → use squeez_agent_costs to track, then refactor tasks to batch work
• Smaller prompts injected into agents → squeez compresses them at dispatch, but smaller is better
• Shorter CLAUDE.md / AGENTS.md files → run squeez compress-md --ultra to drop abbreviations and filler

---

Benchmarks

Measured on macOS (Apple Silicon). Token count = chars / 4 (matches Claude's ~4 chars/token). Run squeez benchmark to reproduce.

Per-scenario results — 46 scenarios × 5 iterations

Scenario	Before	After	Reduction	Latency
curl_json	18,904 tk	36 tk	-100%	1.1 ms
az_json	23,479 tk	74 tk	-100%	549 µs
summarize_huge	82,257 tk	467 tk	-99%	83.9 ms
xcode_build	1,881 tk	17 tk	-99%	253 µs
go_test_ndjson_failures	7,176 tk	106 tk	-99%	1.0 ms
read_reread_distant	717 tk	17 tk	-98%	7.2 ms
rsync_transfer	912 tk	26 tk	-97%	188 µs
repetitive_output	4,692 tk	134 tk	-97%	354 µs
pytest_failures	3,402 tk	108 tk	-97%	259 µs
jest_json_failures	5,643 tk	218 tk	-96%	663 µs
systemctl_status	732 tk	41 tk	-94%	171 µs
ps_aux	40,373 tk	2,338 tk	-94%	1.0 ms
cargo_test_failures	1,934 tk	157 tk	-92%	202 µs
git_log_200	2,692 tk	275 tk	-90%	358 µs
tsc_errors	731 tk	101 tk	-86%	197 µs
eslint_json_failures	1,553 tk	233 tk	-85%	343 µs
pip_install	407 tk	62 tk	-85%	164 µs
high_context_adaptive	4,418 tk	729 tk	-84%	1.5 ms
cargo_build_noisy	2,106 tk	439 tk	-79%	404 µs
bundle_install	121 tk	28 tk	-77%	144 µs
docker_logs	665 tk	181 tk	-73%	223 µs
curl_html_response	2,181 tk	626 tk	-71%	229 µs
git_status	50 tk	16 tk	-68%	175 µs
ruff_json_failures	1,261 tk	494 tk	-61%	326 µs
verbose_app_log	4,957 tk	1,978 tk	-60%	767 µs
npm_install	524 tk	218 tk	-58%	215 µs
crosscall_redundancy_3x	486 tk	222 tk	-54%	52.5 ms
ls_la	1,782 tk	872 tk	-51%	226 µs
mypy_errors	650 tk	349 tk	-46%	165 µs
shellcheck_run	335 tk	187 tk	-44%	149 µs
agent_directory_output	3,348 tk	1,937 tk	-42%	927 µs
env_dump	441 tk	287 tk	-35%	157 µs
agent_heavy	2,306 tk	1,514 tk	-34%	688 µs
git_copilot	640 tk	421 tk	-34%	241 µs
find_deep	424 tk	279 tk	-34%	179 µs
adversarial_tiny_output	4 tk	3 tk	-25%	131 µs
md_prose	187 tk	142 tk	-24%	119 µs
md_claude_md	316 tk	270 tk	-15%	224 µs
claude_md_overhead	717 tk	635 tk	-11%	160 µs
next_build_output	902 tk	884 tk	-2%	208 µs
git_diff	502 tk	497 tk	-1%	193 µs
jest_failures	451 tk	448 tk	-1%	183 µs
state_first_simulation	182 tk	181 tk	-1%	138 µs
kubectl_pods	1,513 tk	1,513 tk	-0%	172 µs
adversarial_dense_json	485 tk	485 tk	-0%	143 µs
adversarial_reread_tiny	4 tk	4 tk	-0%	5.6 ms

Aggregate

Metric	Value
Total token reduction	91.2% — 229,443 tk → 20,249 tk
Bash output	-88.7%
Markdown / context files	-18.1%
Wrap / cross-call engine	-99.2%
Quality (signal terms preserved)	46 / 46 pass
Latency p50 (filter mode)	3.6 ms
Latency p95 (incl. wrap/summarize)	7 ms

Estimated cost savings — Claude Sonnet 4.6 · $3.00 / MTok input

Usage	Baseline / month	Saved / month
100 calls / day	$18.00	$16.41 (91%)
1,000 calls / day	$180.00	$164.11 (91%)
10,000 calls / day	$1800.00	$1641.06 (91%)

Independently verified with a real tokenizer

The table above uses chars / 4 for reproducibility (no vocab, no deps). A fair objection: "that's a made-up unit — show me the reduction under a real byte-pair tokenizer." So bench/verify_tokens.py re-tokenizes every fixture — the raw output and the squeez-compressed output — with cl100k_base (the real GPT-4 family BPE tokenizer, via tiktoken) and compares against chars/4:

Model	Aggregate reduction (22 filter/markdown fixtures)
Real BPE (cl100k_base)	83.5% — 112,557 tk → 18,590 tk
chars / 4 (benchmark unit)	83.0%
Divergence	0.5 pts → the reported reduction is not a token-model artifact

Reduction ratios are near model-invariant, so a real tokenizer confirms the claim rather than inflating it. Reproduce end-to-end:

cargo build --release
pip install tiktoken
python3 bench/verify_tokens.py          # human-readable table
python3 bench/verify_tokens.py --json   # machine-readable (see bench/verify_tokens.json)

Nothing is cherry-picked: fixtures that squeez can't help (e.g. kubectl_pods, a 61-line output below the truncation threshold) stay in the aggregate at 0%.

---

Commands

squeez wrap <cmd>                        # compress a command's output end-to-end
squeez filter <hint>                     # compress stdin (piped usage)
squeez config <get|set|list|reset|path>  # inspect/change config.ini (also via the /squeez skill)
squeez compress-md [--ultra] [--dry-run] [--all] <file>...   # compress markdown files
squeez benchmark [--json] [--output <file>] [--scenario <name>] [--iterations <n>]
squeez mcp                               # JSON-RPC 2.0 MCP server over stdin/stdout
squeez protocol                          # print the auto-teach payload (markers + protocol)
squeez update [--check] [--insecure]     # self-update
squeez init [--copilot]                  # session-start hook (called by hook, not manually)
squeez calibrate                         # auto-tune config from benchmarks
squeez budget-params <tool>              # output JSON budget patch for tool
squeez compact-summary                   # PostCompact hook: re-inject session state (called by hook)
squeez --version

Escape hatch — bypass compression for one command

--no-squeez git log --all --graph

Prefix any command with --no-squeez to run it raw without squeez touching it.

squeez wrap

Runs a command, compresses its output, and prints a savings header:

# squeez [git log] 2692→289 tokens (-89%) 0.2ms [adaptive: Ultra]

wrap re-executes the command through a shell: sh -c on Unix, and on Windows bash -c, then sh -c, when one of them resolves on PATH — every agent host writes its terminal commands for bash there, so re-running them under cmd.exe corrupts quoting, $(…), backticks and ;. cmd /C remains the fallback when no POSIX shell is present, and is also retried automatically if the preferred shell cannot be spawned. %SystemRoot%\System32\bash.exe is skipped: that is the WSL launcher, which would run the command inside a Linux distro rather than against the host filesystem. CLAUDE_CODE_GIT_BASH_PATH is honoured when set. Set SQUEEZ_SHELL to force a specific shell:

SQUEEZ_SHELL="C:/Program Files/Git/bin/bash.exe" squeez wrap 'ls -la'

squeez filter

Reads from stdin. Use for manual pipelines:

git log --oneline | squeez filter git
docker logs mycontainer 2>&1 | squeez filter docker

squeez compress-md

Pure-Rust, zero-LLM compressor for markdown files. Preserves code blocks, inline code, URLs, headings, file paths, and tables. Compresses prose only. Always writes a backup at <stem>.original.md.

squeez compress-md CLAUDE.md             # Full mode (English default)
squeez compress-md --ultra CLAUDE.md    # + abbreviations (with→w/, fn, cfg, etc.)
squeez compress-md --lang pt CLAUDE.md  # pt-BR locale (articles, fillers, phrases)
squeez compress-md --dry-run CLAUDE.md  # preview, no write
squeez compress-md --all                # compress all known locations automatically

When auto_compress_md = true (default), squeez init runs --all silently on every session start.

squeez benchmark

Reproducible measurement of token reduction, cost, latency, and quality across 19 scenarios:

squeez benchmark                          # human-readable report
squeez benchmark --json                   # JSON to stdout
squeez benchmark --output report.json     # save JSON report
squeez benchmark --scenario git           # run only git scenarios
squeez benchmark --iterations 5           # more iterations per scenario
squeez benchmark --list                   # list all scenarios

Quality is scored by checking that signal terms (words from error/warning/failed lines in the baseline) survive compression. 19/19 pass at ≥ 50% threshold.

squeez benchmark --efficiency-proof additionally reports cache-aware effective costs: list-price ratios (input ×1.0, cache write ×1.25, cache read ×0.1, output ×5) applied identically to both sides of every row under the same cache state — so the provider's caching discount is never miscounted as compression savings. Negative savings are representable, never floored; JSON output is schema_version: 2. Savings are reported per compressed slice; end-to-end depends on workload.

squeez mcp

Runs a Model Context Protocol JSON-RPC 2.0 server over stdin/stdout. Hand-rolled, no mcp.server / fastmcp dependency — keeps the libc-only constraint intact. Wire it into Claude Code:

claude mcp add squeez -- /path/to/squeez mcp

Thirteen read-only tools become available to the LLM:

Tool	Returns
squeez_recent_calls	Last N bash invocations with hash + length + cmd snippet — check before re-running
squeez_seen_files	Files this session has touched, with access type (Read/Write/Created/Deleted), sorted by recency
squeez_seen_errors	Distinct error fingerprints observed this session (FNV-1a hashes of normalized errors)
squeez_seen_error_details	Error fingerprints with the first 128 chars of message text — find what the error was
squeez_session_summary	Token accounting + call counts (tokens_bash / tokens_read / tokens_other / seen_files / seen_errors / seen_git_refs)
squeez_session_stats	Dedup hit counts (exact + fuzzy), summarize triggers, Ultra-mode calls, tokens saved per category
squeez_agent_costs	Sub-agent usage: spawn count, cumulative estimated tokens, per-call breakdown
squeez_session_efficiency	Session efficiency scores: compression ratio, tool choice, context reuse, budget conservation (basis points)
squeez_prior_summaries	Last N finalized prior-session summaries with structured fields: investigated / learned / completed / next_steps
squeez_search_history	Full-text search across all session summaries — find when you last saw an error or touched a file
squeez_file_history	Sessions where a given file path was touched, with token-savings and commit status
squeez_session_detail	Full structured view of a past session by date: calls, files, errors, git events, test summary
squeez_protocol	Auto-teach payload — read once per session to learn squeez's markers + memory protocol

All read-only. Backed by SessionContext::load(), memory::read_last_n(), and memory::search_history(). No side effects.

squeez protocol

Prints the auto-teach payload — a 2.4 KB self-describing block covering:

• The 5-rule memory protocol (what to do with [squeez: ...] markers, when to call the MCP tools)
• The output marker spec (# squeez [...], [squeez: identical to ...], [squeez: ~95% similar to ...], squeez:summary, # squeez hint:)

Same content the MCP squeez_protocol tool returns. Pipe it into a system prompt or paste it into a one-shot session that doesn't have the MCP server connected.

---

Configuration

Optional config file — all fields have defaults, none are required.

Platform	Config path
Claude Code / default	~/.claude/squeez/config.ini
Copilot CLI	~/.copilot/squeez/config.ini

# ── Compression ────────────────────────────────────────────────
max_lines              = 200     # generic truncation limit
dedup_min              = 3       # collapse lines appearing ≥N times
git_log_max_commits    = 20
git_diff_max_lines     = 150
docker_logs_max_lines  = 100
find_max_results       = 50
bypass                 = docker exec, psql, mysql, ssh   # never compress these

# ── Context engine ─────────────────────────────────────────────
adaptive_intensity         = true    # truly adaptive: Full <80% budget, Ultra ≥80%
context_cache_enabled      = true    # track seen files/errors across calls
redundancy_cache_enabled   = true    # collapse identical OR fuzzy-similar recent outputs
read_dedup_session_long    = true    # also collapse a re-read of an unchanged file past the 16-call window
summarize_threshold_lines  = 500     # outputs above this trigger summarize fallback (×2 if benign)
compact_threshold_tokens   = 120000  # session token budget — drives adaptive intensity

# ── Session memory ─────────────────────────────────────────────
memory_retention_days = 30

# ── Output / persona ───────────────────────────────────────────
persona          = ultra    # off | lite | full | ultra
focus            = off      # off | adhd — output *structure*, orthogonal to persona
auto_compress_md = true     # run compress-md on every session start
lang             = en       # compress-md locale: en | pt (pt-BR) — more languages extensible

# ── Advanced tuning (rarely needed) ───────────────────────────
max_call_log              = 32    # rolling call log depth (also caps redundancy window)
recent_window             = 16    # how many recent calls are eligible for redundancy lookup
similarity_threshold      = 0.85  # Jaccard threshold for fuzzy dedup (0.0–1.0)
ultra_trigger_pct         = 0.80  # fraction of context budget at which Full → Ultra
class_density             = true  # content-class token estimate: dense chars/2.0, prose chars/3.7 (false = legacy flat path)
net_win_min_tokens        = 24    # skip compression + header when it saves less than this (0 = gate off)
mcp_prior_summaries_default = 5   # default n for squeez_prior_summaries
mcp_recent_calls_default    = 10  # default n for squeez_recent_calls

# ── Token economy ─────────────────────────────────────────────
agent_warn_threshold_pct  = 0.50  # warn when agent cost > 50% of budget
burn_rate_warn_calls      = 20    # warn when < 20 calls remaining
agent_spawn_cost          = 200000 # estimated tokens per Agent/Task spawn
read_max_lines            = 0     # max lines injected into Read tool_input (0 = off)
grep_max_results          = 0     # max results injected into Grep tool_input (0 = off)

# ── Auto-curation nudges ──────────────────────────────────────
nudge_enabled              = true   # emit [squeez: hint ...] markers on recurring patterns
nudge_error_threshold      = 3      # fingerprint repeats before a nudge fires
nudge_file_mod_threshold   = 5      # writes/creates to same path before nudge fires
nudge_cmd_repeat_threshold = 4      # expensive-command repeats before nudge fires

# ── Continuous handler calibration ────────────────────────────
handler_stats_enabled      = true   # accumulate per-handler savings across sessions

Adaptive intensity — Full / Ultra split

When adaptive_intensity = true (default), squeez actually adapts to session pressure rather than always running Ultra:

Used / budget	Tier	Scaling
< 80%	Full	×0.6 limits, dedup_min ×0.66 (floor 2)
≥ 80%	Ultra	×0.3 limits, dedup_min ×0.5 (floor 2)
adaptive_intensity = false	Lite	passthrough — no scaling

Floors are enforced so we never reduce to zero: max_lines ≥ 20, git_diff_max_lines ≥ 20, dedup_min ≥ 2, summarize_threshold_lines ≥ 50.

The active level is shown in every bash header: [adaptive: Full] or [adaptive: Ultra].

Pre-0.3 squeez was effectively always-Ultra. The new behavior preserves more verbatim text in the common case (empty / mid-session) and only graduates to aggressive compression when the context budget is genuinely under pressure.

Caveman persona

Three intensity levels (lite, full, ultra) and off. Default is ultra. The persona prompt is injected into:

• The Claude Code session banner (printed at SessionStart)
• The <!-- squeez:start -->…<!-- squeez:end --> block in ~/.copilot/copilot-instructions.md for Copilot CLI

ADHD focus mode

focus = adhd is a second, independent axis. Persona decides how terse the prose is; focus decides how it is ordered. They stack — persona = ultra plus focus = adhd keeps maximum compression while forcing action-first structure.

Turning it on changes three surfaces:

1. The model's prose — a 10-rule block (EN or pt-BR, per lang) rides alongside the persona block: lead with the next action, number multi-step work, restate state in one line, concrete time estimates, cap lists at 5, no preamble or closing pleasantries.
2. The session banner — the pending next step becomes line 1, remaining steps are numbered, one prior session is shown instead of three, and the stats line sinks to the bottom (state, not an action).
3. Advisories and doctor — [squeez: …] bursts are capped at 5 per command with an honest +N more advisories suppressed tail, and squeez doctor sorts failures first and ends with one command to run.

squeez config set focus adhd    # then `squeez init` to rewrite the memory block
squeez config set focus off     # back to the default ordering

The ruleset is adapted from the MIT-licensed i-have-adhd skill by ayghri.

---

How it works

Compression pipeline

Each bash command passes through four strategies in order:

1. smart_filter — strips ANSI codes, progress bars, spinner chars, timestamps, and tool-specific noise (npm download lines, stack frame noise, etc.)
2. dedup — lines appearing ≥ dedup_min times are collapsed to one entry annotated [×N]
3. grouping — files in the same directory (≥5 siblings) are collapsed to dir/ N modified [squeez grouped]
4. truncation — Head (keep first N) or Tail (keep last N) depending on handler; truncated portion noted

Supported handlers

Category	Commands
Git	git
Docker / containers	docker, docker-compose, podman
Package managers	npm, pnpm, bun, yarn
Build systems	make, cmake, gradle, mvn, xcodebuild, cargo (build), next build/dev/start
Test runners	cargo test, jest, vitest, pytest, nextest, playwright, bun test
TypeScript / linters	tsc, eslint, biome
Cloud CLIs	kubectl, gh, aws, gcloud, az, wrangler
Databases	psql, prisma, mysql, drizzle-kit
Filesystem	find, ls, du, ps, env, lsof, netstat
JSON / YAML / IaC	jq, yq, terraform, tofu, helm, pulumi
Text processing	grep, rg, awk, sed
Network	curl, wget
Runtimes	node, python, ruby
Generic fallback	everything else

Hooks (Claude Code & Copilot CLI)

Six hooks work together automatically after install on Claude Code (three on Copilot CLI):

• PreToolUse — rewrites safe Bash calls: git status → squeez wrap git status (risky/bypassed commands pass through to native permission rules); injects Read/Grep/Glob limits; compresses Agent/Task prompts
• SessionStart — runs squeez init: finalizes previous session into a memory summary, injects the persona prompt
• PostToolUse — tracks every tool result; rewrites Read/Grep/Glob/Monitor output via updatedToolOutput when content is redundant or oversized (Claude Code v2.1.119+)
• SubagentStop (Claude Code only) — feeds last_assistant_message into SessionContext so the parent agent can dedup against what the sub-agent saw
• PreCompact (Claude Code only) — logs compaction events for session efficiency metrics; allows compaction to proceed
• PostCompact (Claude Code only) — re-injects tracked session state (files, errors, git refs, retrievable blob ids) as additionalContext so it survives compaction (squeez compact-summary)

Cross-call redundancy

Two-path dedup across the last 16 calls:

Exact match — FNV-1a hash of the compressed output. When a subsequent call produces the same bytes, it collapses to:

[squeez: identical to 515ba5b2 at bash#35 — re-run with --no-squeez]

Fuzzy match — bottom-k MinHash over whitespace-token trigrams (k=96, Jaccard ≥ 0.85, length-ratio guard ≥ 0.80). Survives timestamp changes, added/removed blank lines, and single-line edits. Collapses to:

[squeez: ~92% similar to 515ba5b2 at bash#35 — re-run with --no-squeez]

Minimum 6 lines to attempt fuzzy match (below that, exact-only).

Summarize fallback

When raw output exceeds summarize_threshold_lines (default 500), the full pipeline is bypassed and replaced with a ≤40-line dense summary:

squeez:summary cmd=docker logs app
total_lines=5003
top_errors:
  - error: connection refused on tcp://10.0.0.1:5432
top_files:
  - /var/log/app/error.log
test_summary=FAILED: 3 of 248
ids_preserved:
  - 3f2ec81ea3e4ce24
  - 550e8400-e29b-41d4-a716-446655440000
  - PROJ-1482
tail_preserved=20
[last 20 lines verbatim...]

Identifier factsheet: the ids_preserved: block carries exact identifiers found in the region the summary drops — git SHAs / hex ids (≥7 chars with a digit), UUIDs, ticket codes, versions, and large integers — so the most dangerous failure mode of lossy summarization (a silently lost hash) can't happen. Extraction is deterministic and budget-capped at 16 facts / 256 chars; bulk generated sequences (hundreds of version strings in a build log) are recognized as noise and dropped wholesale while opaque ids survive. The Structured summary shape carries the same list as "ids":[...].

Benign-aware threshold: before summarizing, squeez scans for error markers (error:, panic, traceback, FAILED, EXCEPTION, Fatal). If none are found, the threshold is doubled (1,000 lines default) so successful builds, clean test runs, and uneventful logs stay verbatim unless they are genuinely huge.

---

Platform notes

OpenCode

Plugin installed at ~/.config/opencode/plugins/squeez.js. OpenCode auto-loads plugins on startup. All Bash commands are automatically compressed via squeez wrap.

Hermes

Directory plugin installed at $HERMES_HOME/plugins/squeez-fallback/ — ~/.hermes/plugins/ when HERMES_HOME is unset (a leading ~ is expanded). Set HERMES_HOME if Hermes lives elsewhere (e.g. %LOCALAPPDATA%\hermes on Windows), otherwise squeez setup --host=hermes will not detect it. Directory plugins are opt-in; enable it once with hermes plugins enable squeez-fallback.

GitHub Copilot CLI

Hooks registered in ~/.copilot/settings.json. Session memory written to ~/.copilot/copilot-instructions.md (Copilot CLI reads this automatically). State stored separately at ~/.copilot/squeez/.

Refresh memory manually:

SQUEEZ_DIR=~/.copilot/squeez ~/.claude/squeez/bin/squeez init --copilot

Pi

TypeScript extension installed at ~/.pi/agent/extensions/squeez/index.ts. Pi auto-discovers extensions from that directory — no settings patching needed. Session memory is injected via a skill at ~/.pi/agent/skills/squeez/SKILL.md; Pi includes the skill description in every system prompt and loads full instructions on demand. Unlike other hosts, Pi achieves BUDGET_HARD output compression via the tool_result event (return-patch API), not just soft hints.

---

Local development

Requires Rust stable. Windows requires Git Bash.

git clone https://github.com/claudioemmanuel/squeez.git
cd squeez

cargo test                  # run all tests (356 tests, 37 suites)
cargo build --release       # build release binary

bash bench/run.sh           # filter-mode benchmark (14 fixtures)
bash bench/run_context.sh   # context-engine benchmark (3 wrap scenarios)
./target/release/squeez benchmark   # full 19-scenario benchmark suite

bash build.sh               # build + install to ~/.claude/squeez/bin/

---

Contributing

git checkout -b feature/your-change
cargo test
cargo build --release
bash bench/run.sh
git push -u origin feature/your-change
gh pr create --base main --title "Short title" --body "Description"

CI runs cargo test, bench/run.sh, bench/run_context.sh, and squeez benchmark on every push and pull request.

See CONTRIBUTING.md for coding standards.

---

Similar projects

There is an unrelated project with the same name at KRLabsOrg/squeez. This project (claudioemmanuel/squeez) is a hook-based token compressor for AI coding CLIs.

---

License

Licensed under the Apache License 2.0 — see LICENSE + NOTICE.

Contributions require a DCO sign-off (git commit -s …) rather than a CLA. You keep copyright on what you contribute; sign-off is a lightweight affirmation that you have the right to submit it under Apache 2.0. See CONTRIBUTING.md for details.