The context intelligence layer for AI coding agents.

The layer between your tools and your AI. PandaFilter understands what's noise and what matters — compressing, routing, and preserving the right context so your agent thinks faster, costs less, and never loses its place.

brew tap AssafWoo/pandafilter
brew install pandafilter

Linux / any platform:

curl -fsSL https://raw.githubusercontent.com/AssafWoo/homebrew-pandafilter/main/install.sh | bash

First run: PandaFilter downloads the BERT model (~90 MB, all-MiniLM-L6-v2) from HuggingFace and caches it at ~/.cache/huggingface/. Subsequent runs are instant.

Then wire it in — one command installs for every AI agent you have:

panda init --agent all

Auto-detects Claude Code, Cursor, Gemini CLI, Codex, Windsurf, Cline, OpenClaw, and VS Code Copilot. Skips anything that isn't installed. Or target one specifically:

panda init                          # Claude Code (default)
panda init --agent cursor           # Cursor
panda init --agent gemini           # Gemini CLI
panda init --agent codex            # Codex (CLI + VS Code extension)
panda init --agent windsurf         # Windsurf
panda init --agent cline            # Cline
panda init --agent openclaw         # OpenClaw
panda init --agent copilot          # VS Code Copilot

Raw command output is filtered, deduplicated, and semantically compressed by a BERT-powered pipeline that understands what matters for your task — not just what matches a regex. When your AI agent runs pip install, cargo build, or npm install, PandaFilter intercepts the output and strips download progress, module graphs, and passing test lines. The agent sees a clean summary with errors, warnings, and results. Nothing useful is dropped.

A content-aware router analyzes each output and activates only the strategies relevant to it: error-focus for test failures, dedup for log streams, structural digest for unchanged file re-reads, semantic summarization for prose. No more one-size-fits-all.

Enable with use_router = true in panda.toml.

PandaFilter learns your codebase's noise patterns across sessions. It tracks what you've read, what you've changed, and where the context pressure is building — adapting its compression strategy in real time.

New in v1.3.0: File re-reads now send diffs instead of full file content. Unchanged re-reads return structural digests (function/class signatures). Both happen automatically — no config needed.

When your agent's context fills up and auto-compacts, PandaFilter preserves what matters: edited files, error signatures, key decisions. The next session starts oriented, not blank.

Requires Claude Code and is installed automatically with panda init.

When your AI agent runs a command — pip install, cargo build, npm install — PandaFilter intercepts the output and removes everything the model doesn't need: download progress, module graphs, passing test lines, spinners. The agent sees a clean summary with errors, warnings, and results. Nothing useful is dropped.

No config changes. No workflow changes. Runs 100% locally.

Run panda gain after a session to see your cumulative savings:

Numbers from ccr/tests/handler_benchmarks.rs. Run panda gain to see your own live data.

panda gain — see your token savings:

panda gain                    # overall summary
panda gain --breakdown        # per-command table
panda gain --history          # last 14 days
panda gain --insight          # categorized savings + top saves

panda doctor — diagnose the full installation in one command.

panda init --uninstall — remove hooks:

panda init --uninstall                  # Claude Code
panda init --agent cursor --uninstall   # Cursor

panda focus — opt-in: tells the agent which files matter for the current prompt, preventing unnecessary reads:

panda focus --enable             # enable for this repo
panda focus --disable            # disable (keeps index data)
panda focus --status             # show status + index age
panda focus --dry-run            # preview guidance without enabling

panda index — manually rebuild the file-relationship index:

panda index                      # full/incremental build for current repo

Other commands:

panda verify                            # check hook integrity
panda discover                          # scan history for unfiltered commands
panda run git status                    # run a command through PandaFilter manually
panda proxy git status                  # run raw (no filtering), record baseline
panda read-file src/main.rs --level auto  # preview read filtering
panda expand ZI_3                       # restore a collapsed block
panda noise                             # show learned noise patterns; --reset to clear
panda compress --scan-session           # compress current conversation context

0. Hard input ceiling (200k chars — truncates before any stage)
1. Strip ANSI codes
2. Normalize whitespace
3. Global regex pre-filter (progress bars, spinners, download lines, decorators)
4. NDJSON streaming compaction (go test -json, jest JSON reporter)
5. Command-specific pattern filter
6. If over summarize_threshold_lines:
   6a. BERT noise pre-filter
   6b. Entropy-adaptive BERT summarization (up to 7 passes)
7. Hard output cap (50k chars)

[global]
summarize_threshold_lines = 50
head_lines = 30
tail_lines = 30
strip_ansi = true
normalize_whitespace = true
deduplicate_lines = true
input_char_ceiling = 200000
output_char_cap = 50000
# cost_per_million_tokens = 15.0

# v1.3.0: Adaptive MoE router (opt-in)
# use_router = true                  # enable adaptive expert routing
# router_exploration_noise = true    # prevent expert collapse in long sessions

[tee]
enabled = true
mode = "aggressive"   # "aggressive" | "always" | "never"

[read]
mode = "auto"   # "passthrough" | "auto" | "strip" | "aggressive" | "structural"
# v1.3.0: delta mode (re-reads send diffs) and structural mode (signature-only) are
# now active automatically — no config change needed.

[focus]
enabled = false       # disabled by default — enable with `panda focus --enable` after testing
min_files = 25        # skip for repos smaller than this
min_lines = 2000      # skip for repos with fewer source lines

[commands.git]
patterns = [
  { regex = "^(Counting|Compressing|Receiving|Resolving) objects:.*", action = "Remove" },
]

[commands.cargo]
patterns = [
  { regex = "^\\s+Compiling \\S+ v[\\d.]+", action = "Collapse" },
  { regex = "^\\s+Downloaded \\S+ v[\\d.]+", action = "Remove"   },
]

[commands.myapp]
patterns = [
  { regex = "^DEBUG:",            action = "Remove" },
  { regex = "^\\S+\\.ts\\(",     action = "TruncateLinesAt", max_chars = 120 },
]
on_empty = "(no relevant output)"

[commands.myapp.match_output]
pattern        = "Server started"
message        = "ok — server ready"
unless_pattern = "error"

ccr/        CLI binary (panda) — handlers, hooks, session state, commands
ccr-core/   Core library (no I/O) — pipeline, BERT summarizer, config, analytics
ccr-sdk/    Conversation compression — tiered compressor, deduplicator, Ollama
ccr-eval/   Evaluation suite — fixtures against Claude API
config/     Embedded default filter patterns

panda init --uninstall                            # Claude Code
panda init --agent cursor   --uninstall           # Cursor
panda init --agent gemini   --uninstall           # Gemini CLI
panda init --agent codex    --uninstall           # Codex
panda init --agent windsurf --uninstall           # Windsurf
panda init --agent cline    --uninstall           # Cline
panda init --agent copilot  --uninstall           # VS Code Copilot

brew uninstall pandafilter && brew untap AssafWoo/pandafilter   # Homebrew
# or: cargo uninstall panda

rm -rf ~/.local/share/panda                   # analytics + sessions
rm -rf ~/.cache/huggingface/hub/models--sentence-transformers--all-MiniLM-L6-v2

Does PandaFilter change what the agent can see? It removes noise — build progress, passing test lines, module download logs. Errors, file paths, and results are always kept.

What if I don't want a specific command filtered? Add a rule to .panda/filters.toml to customize or override any handler. See the User-defined filters section. You can also use panda proxy <cmd> to run a command raw with no filtering.

What about commands PandaFilter doesn't know? Output passes through unchanged. PandaFilter never silently drops output from unknown commands.

How do I verify it's working? Run panda gain after a session. To see exactly what the agent received from a specific command: panda run git log --oneline -20.

Does PandaFilter send any data outside my machine? No. All processing is fully local. BERT runs on-device.

What is Context Focusing? An opt-in feature that tells the agent which files are relevant for the current prompt, preventing it from reading unrelated files. Enable with panda focus --enable after running panda doctor to confirm the index is ready.

AI coding sessions are expensive — not because of what you ask, but because of what the agent reads back. Every cargo build or npm install dumps thousands of tokens of noise into the context window. I built PandaFilter to strip that out automatically.

The name comes from how a panda eats: it consumes enormous amounts of raw material and extracts only what it needs.

— Assaf Petronio · github.com/AssafWoo

Open an issue or PR on GitHub. To add a handler: implement the Handler trait and register it in ccr/src/handlers/mod.rs — see git.rs as a template.

MIT — see LICENSE.