context-janitor 1.0.0rc1

Cost-saving CLI middleware for pruning oversized LLM agent tool catalogs.

Context Janitor

100.0% tool-selection accuracy on the bundled synthetic benchmark at 0 ms median latency, with zero router cost.

Context Janitor is a dependency-free CLI and Python library for pruning oversized LLM tool catalogs. Give it a user prompt and a JSON list of tools, and it returns only the tools the agent is likely to need.

It is built for agent systems where sending every available tool is expensive, slow, and noisy. If an API-backed router fails, times out, or is missing credentials, Context Janitor can fall back to a local heuristic so the pipeline keeps moving.

Context Janitor is MCP-compatible by design. MCP servers expose structured tool definitions, and Context Janitor can sit between those JSON tool catalogs and your agent runtime with janitor mcp-proxy.

Why It Exists

Large tool catalogs make agents worse in two ways:

• They inflate every request with thousands of extra prompt tokens.
• They increase the chance that the model picks a plausible but wrong tool.

Context Janitor keeps the tool surface small before the main model sees it.

Setup	Tools sent	Tool overhead	Expected effect
Without Janitor	50	High	More prompt cost and more tool confusion
With Janitor	5	Low	Smaller payloads and clearer tool choice

Benchmark Snapshot

Run locally:

python scripts\benchmark.py --providers heuristic

Current output on the included 100-prompt synthetic benchmark and examples/tools.json:

+-----------------------+--------------------+---------------+-----------+--------+-----------------+------------------+-------------+--------------------------------------------------------------------------------------------------------------------------------------------+
| Mode                  | Selection accuracy | Agent success | Median ms | p95 ms | Router cost/run | Tool payload/run | Compression | Notes                                                                                                                                      |
+-----------------------+--------------------+---------------+-----------+--------+-----------------+------------------+-------------+--------------------------------------------------------------------------------------------------------------------------------------------+
| No Janitor (baseline) | 100.0%             | not measured  | 0         | 0      | $0.000000       | $0.001060        | 0.0%        | all 8 tools sent for 100 prompts |
| heuristic             | 100.0%             | not measured  | 0         | 0      | $0.000000       | $0.000280        | 73.6%       | ok |
+-----------------------+--------------------+---------------+-----------+--------+-----------------+------------------+-------------+--------------------------------------------------------------------------------------------------------------------------------------------+

Benchmark notes:

• Selection accuracy means the expected tool was present in the pruned selection.
• No Janitor (baseline) has 100% selection accuracy because every tool is sent.
• Agent success is intentionally not measured unless you provide real agent eval data.
• Tool payload/run uses --payload-price-per-million, which defaults to $5.00.
• Router cost/run uses --router-price-per-million, which defaults to $0.15.
• The included benchmark is a small synthetic sanity check. Run it against your own catalog before making production claims.

The repository also includes examples/messy_production_evals.jsonl, a 100-case prompt pack with informal, ambiguous workplace phrasing, plus examples/messy_aliases.janitor.yaml for team slang. The release gate requires the local heuristic to keep the expected tool for every messy case against the simulated production catalog when that alias config is provided.

To display measured agent success rates:

python scripts\benchmark.py --providers heuristic --agent-success-file examples\agent_success.example.json

Installation

From a local checkout:

pip install -e .

With test dependencies:

pip install -e ".[test]"

With contributor tooling:

pip install -e ".[dev]"

The package exposes two console scripts:

• janitor
• context-janitor

Most examples use the shorter janitor command.

On Windows, pip may install those scripts outside your current PATH. If janitor is not recognized in cmd.exe, use:

set PATH=%PATH%;%APPDATA%\Python\Python314\Scripts

Or run the module directly:

python -m context_janitor.cli --help

Quick Start

janitor prune --prompt "Search GitHub issues and make a PR" --tools examples\tools.json --limit 2

Output:

{
  "selected": [
    {
      "name": "github_search_issues",
      "description": "Search issues in a GitHub repository by text, label, state, or assignee."
    },
    {
      "name": "github_create_pr",
      "description": "Open a pull request with a title, body, source branch, and target branch."
    }
  ],
  "metadata": {
    "requested_provider": "heuristic",
    "provider": "heuristic",
    "fallback_used": false,
    "cache_hit": false,
    "duration_ms": 0,
    "limit": 2,
    "available_tools": 8,
    "original_tokens": 212,
    "selected_tokens": 60,
    "reduced_tokens": 152,
    "estimated_savings_usd": 0.00076
  }
}

Names-only output:

janitor prune --prompt "Search GitHub issues and make a PR" --tools examples\tools.json --limit 2 --format names

github_search_issues
github_create_pr

Middleware Mode

middleware reads an OpenAI-compatible request JSON from stdin, prunes the tools field, and writes the modified payload to stdout.

Get-Content request.json | janitor middleware --limit 5

In cmd.exe, use type instead of Get-Content:

type examples\request.example.json | janitor middleware --limit 2

Input shape:

{
  "messages": [
    { "role": "user", "content": "Create a calendar event" }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "calendar_create",
        "description": "Create events."
      }
    },
    {
      "type": "function",
      "function": {
        "name": "web_search",
        "description": "Search the web."
      }
    }
  ]
}

Logs go to stderr, so stdout remains safe to pipe into another command.

Supported Tool Formats

Plain tool objects:

[
  {
    "name": "github_create_pr",
    "description": "Open a pull request."
  }
]

OpenAI-style function tools:

[
  {
    "type": "function",
    "function": {
      "name": "github_create_pr",
      "description": "Open a pull request.",
      "parameters": {
        "type": "object"
      }
    }
  }
]

Object wrappers are also accepted:

{
  "tools": [
    { "name": "web_search", "description": "Search the web." }
  ]
}

Selection Providers

Context Janitor supports four provider values:

Provider	Uses network	Required environment
heuristic	No	None
openai	Yes	OPENAI_API_KEY
anthropic	Yes	ANTHROPIC_API_KEY
gemini	Yes	GEMINI_API_KEY or GOOGLE_API_KEY

Provider calls use only the Python standard library and default to an 800 ms timeout.

Example with OpenAI:

$env:OPENAI_API_KEY = "..."
janitor prune `
  --provider openai `
  --model gpt-4o-mini `
  --prompt "Summarize this PDF" `
  --tools tools.json `
  --timeout-ms 800 `
  --fallback heuristic `
  --log-level INFO

If the provider errors, rate-limits, times out, or is missing credentials, --fallback heuristic logs a warning and returns a heuristic selection instead of crashing the pipeline.

Set --fallback none if you want provider failures to exit with an error.

How The Heuristic Works

The local selector is not just a keyword set. It is a compact TF-IDF-style ranker:

• Tokenizes the prompt and each tool's name + description
• Splits names like github_search_issues into useful terms
• Removes common stop words
• Expands common intent aliases like meeting -> calendar event
• Scores term frequency in the tool text
• Weighs rare terms more heavily with inverse document frequency
• Adds a small bonus for longer substring matches

flowchart LR
  A["User prompt"] --> B["Tokenize and expand aliases"]
  C["Tool catalog"] --> D["Tokenize names and descriptions"]
  B --> E["Score prompt terms against each tool"]
  D --> E
  E --> F["Boost rare, specific matches"]
  F --> G["Return top N tools"]

Distinctive terms like stripe, github, postgres, or pdf usually beat generic words like create, get, or send.

Configuration

Context Janitor searches upward from the current directory for .janitor.yaml.

Example:

provider: anthropic
model: claude-3-haiku-20240307
limit: 5
fallback: heuristic
cache: true
timeout_ms: 800
log_level: INFO
format: json
price_per_million_tokens: 5.0
keep: log_error,notify_admin
aliases:
  bq: bigquery,query,warehouse
  blast: email,send
  prio: priority

The config parser intentionally supports simple top-level key: value settings plus the aliases mapping shown above. It is not a full YAML implementation.

CLI flags override config values.

For safety in untrusted repositories, an auto-discovered .janitor.yaml cannot silently switch the selector from heuristic to a network provider. If a discovered config sets provider: openai, provider: anthropic, or provider: gemini, pass --config path\to\.janitor.yaml or --provider ... explicitly to confirm that you trust the project and intend to send prompts/tool metadata to that provider.

Key	Default	Description
provider	heuristic	Selection backend: heuristic, openai, anthropic, or gemini
model	null	Model name for API-backed providers
limit	5	Maximum number of tools to keep
fallback	heuristic	Use heuristic or none after provider failure
cache	false	Reuse previous selections from local cache
timeout_ms	800	Provider timeout in milliseconds
log_level	WARNING	DEBUG, INFO, WARNING, ERROR, or CRITICAL
format	json	prune output format: json, names, or raw
price_per_million_tokens	5.0	Price used for savings estimates
keep	empty	Comma-separated tool names that must stay selected
aliases	empty	Team-specific prompt slang to expand before ranking

Use aliases for vocabulary your tool descriptions do not already cover. For example, if your team types bq but your tool says BigQuery, map bq to bigquery,query,warehouse in config instead of hardcoding that slang into Context Janitor.

Required Tools

Some production agents need safety, audit, or notification tools in every request. Use --keep to force those tools into the selected set:

janitor prune --prompt "Search the web" --tools tools.json --limit 5 --keep log_error,notify_admin

Kept tools reserve slots inside the limit. If --limit 5 and you keep two tools, Janitor ranks the catalog for the remaining three slots.

Selections modified by keep are not written into the normal semantic cache, because required tools are policy rather than prompt relevance.

Cache

Enable prompt caching:

janitor prune --cache --prompt "Summarize the daily logs" --tools tools.json

Cache file:

~/.janitor_cache/cache.json

The cache stores selections by prompt, provider, model, limit, and catalog hash. It can also reuse highly similar prompts. If the cache cannot be read or written, Janitor ignores the cache and keeps running. Cache updates are written through a temporary file and atomically replaced, so interrupted writes should not leave partial JSON behind.

Privacy note: --cache stores prompt previews and prompt tokens in a local plaintext file. Keep it off for sensitive prompts unless local plaintext storage is acceptable for your environment. Janitor ignores oversized cache files and trims old entries so the cache cannot grow without bound.

Clear the local cache while iterating on prompts or tool descriptions:

janitor clear-cache

Inspect cache metadata:

janitor cache-info

Explain Mode

Use --explain to see why tools were kept or pruned.

janitor prune --prompt "Search GitHub issues" --tools examples\tools.json --limit 2 --explain

JSON output includes an explain array:

{
  "name": "github_search_issues",
  "selected": true,
  "score": 14.0026,
  "matched_terms": ["github", "issues", "search"],
  "top_terms": ["issues", "search", "github", "substring_match"]
}

For --format names or --format raw, explanations are printed to stderr.

Dry Run Mode

Use --dry-run with middleware to audition Janitor without changing the request payload:

Get-Content request.json | janitor middleware --limit 5 --dry-run --log-level INFO

The original JSON is written back to stdout. Janitor logs what it would have kept and pruned to stderr.

CLI Reference

janitor prune

Select tools for a prompt and a tool catalog.

janitor prune --prompt PROMPT --tools tools.json [options]

Options:

Option	Description
--prompt TEXT	User prompt. If omitted, stdin is used
--tools PATH	Required path to a JSON tool catalog
--limit N	Maximum tools to keep
--provider NAME	heuristic, openai, anthropic, or gemini
--model NAME	Model for API-backed providers
--fallback NAME	heuristic or none
--timeout-ms N	Provider timeout
--cache / --no-cache	Enable or disable local cache
--log-level LEVEL	Structured stderr logging level
--price-per-million-tokens N	Cost estimate price
--keep a,b	Required tools to keep
--explain	Include or print ranking explanations
--format json	Default structured output
--format names	Print selected tool names
--format raw	Print original selected tool objects
--config PATH	Explicit config file path

janitor middleware

Read a request payload from stdin and prune its tools field.

janitor middleware [options] < request.json

Most options match prune. middleware also supports --dry-run.

janitor mcp-proxy

Proxy an MCP stdio server and prune tools/list responses before they reach the client:

janitor mcp-proxy --prompt "Find GitHub issues" --limit 5 -- python -m your_mcp_server

MCP tools/list does not include the user's chat prompt, so pass a scoped task prompt with --prompt or JANITOR_PROMPT. Use --keep with prune or middleware for hidden policy tools; for MCP proxy sessions, configure the downstream server around one narrow workflow when possible.

janitor lint

Validate a tool catalog and report quality warnings before using it in production:

janitor lint --tools tools.json

The linter checks the catalog shape, duplicate names, empty descriptions, and very short descriptions.

janitor clear-cache

Delete the local semantic-selection cache:

janitor clear-cache

janitor cache-info

Show cache path, entry count, providers, models, and creation timestamps:

janitor cache-info

Python API

Synchronous API:

from context_janitor.models import load_tools
from context_janitor.selection import select_resilient

tools = load_tools(tool_json)
result = select_resilient(
    provider="openai",
    model="gpt-4o-mini",
    prompt="Find GitHub issues about auth",
    tools=tools,
    limit=5,
    fallback="heuristic",
    timeout_ms=800,
    cache_enabled=True,
    keep=("log_error", "notify_admin"),
)

selected_tools = result.selected

Async wrapper:

from context_janitor.selection import select_resilient_async

result = await select_resilient_async(
    provider="heuristic",
    prompt="Create a calendar event",
    tools=tools,
    limit=3,
)

select_resilient_async runs the same implementation in a worker thread. The current provider clients use the Python standard library rather than native async HTTP.

Structured Logging And ROI

Use --log-level INFO to emit production-friendly logs to stderr:

[Janitor] INFO event=pruned requested_provider=openai provider=heuristic fallback=true cache_hit=false tools_before=50 tools_after=5 tokens_before=12000 tokens_after=1200 tokens_saved=10800 estimated_savings_usd=0.054000 duration_ms=7

Token counts use a lightweight estimate of roughly four characters per token. Savings are useful for quick comparisons, not invoice-grade accounting.

Benchmarks

Run the included benchmark:

python scripts\benchmark.py --providers heuristic openai anthropic gemini --openai-model gpt-4o-mini --anthropic-model claude-3-haiku-20240307 --gemini-model gemini-1.5-flash

Useful benchmark options:

Option	Default	Description
--providers	heuristic	Providers to compare
--limit	5	Tools kept per prompt
--timeout-ms	800	Provider timeout
--router-price-per-million	0.15	Router model input price estimate
--payload-price-per-million	5.0	Main model tool payload price estimate
--agent-success-file	none	JSON map of measured agent success rates

Model pricing moves quickly, so treat the defaults as placeholders and set these values to your current provider prices when calculating ROI.

Example agent success file:

{
  "baseline": 0.85,
  "heuristic": 0.99
}

The benchmark skips API providers when their API keys are missing.

Real Prompt Evals

Use scripts/evaluate.py to check Janitor against prompts from your own product instead of the bundled synthetic benchmark:

python scripts\evaluate.py --tools examples\tools.json --evals examples\evals.example.json --providers heuristic --limit 2

To report the production-facing Distraction Delta, pass measured agent success rates:

python scripts\evaluate.py --tools examples\tools.json --evals examples\evals.example.json --providers heuristic --limit 2 --agent-success-file examples\agent_success.example.json

Distraction Delta is Success_with_Janitor - Success_baseline, which helps separate "the right tool was present" from "the agent actually completed the task more often."

Eval files may be JSON or JSONL. Each case needs a prompt and one of expected_tool, expected_tools, or expected:

[
  {
    "id": "github-triage",
    "prompt": "Find open GitHub issues about billing and summarize the blockers.",
    "expected_tool": "github_search_issues"
  }
]

For production rollout, replace examples/evals.example.json with real tasks from your agent logs and track the resulting accuracy alongside downstream agent success.

Agent Success Evals

Use scripts/eval_agent.py when you want to measure the whole agent loop, not just whether the expected tool survived pruning.

The harness runs your agent command once with the full catalog and once with Janitor-pruned tools. Each run receives a JSON payload on stdin:

{
  "id": "github-triage",
  "mode": "janitor",
  "provider": "heuristic",
  "prompt": "Find open GitHub issues about billing and summarize the blockers.",
  "expected_tools": ["github_search_issues"],
  "tools": [{ "name": "github_search_issues", "description": "Search GitHub issues." }]
}

The agent command should print JSON with a boolean success field:

{ "success": true, "used_tools": ["github_search_issues"] }

Run the bundled deterministic mock agent:

python scripts\eval_agent.py --tools examples\tools.json --evals examples\evals.example.json --providers heuristic --limit 2 -- python examples\agent_runner_mock.py

For a local model or real agent, replace the command after -- with your runner. The runner can call Ollama, llama.cpp, a LangGraph app, or any process that accepts the JSON payload on stdin.

Ollama Smoke Test

The repository includes a small local-model example that prunes a noisy 21-tool catalog before sending the remaining tool definitions to Ollama:

pip install ollama
python examples\ollama_agent.py

Small local models sometimes return tool calls as plain text or fenced JSON instead of native tool calls. The example handles all three shapes so you can confirm the pruned catalog is still usable before wiring Janitor into a larger agent loop.

For thresholded rollout gates, see Production Rollout.

To draft a real eval pack from agent logs:

python scripts\prepare_evals.py --logs agent-logs.jsonl --success-field success --output production-evals.draft.json

Simulated Production Rehearsal

Before you have real logs, generate a deterministic production-like dataset:

python scripts\generate_simulated_data.py

This creates:

• examples\simulated_production_tools.json: 100 OpenAI-style tools across realistic domains.
• examples\simulated_production_evals.json: 100 labeled prompts.
• examples\simulated_agent_logs.jsonl: 100 JSONL agent-log rows.

Run selection accuracy:

python scripts\evaluate.py --tools examples\simulated_production_tools.json --evals examples\simulated_production_evals.json --providers heuristic --limit 5 --min-accuracy 0.95

Run the full agent-success harness with the mock runner:

python scripts\eval_agent.py --tools examples\simulated_production_tools.json --evals examples\simulated_production_evals.json --providers heuristic --limit 5 --min-janitor-success-rate 0.95 --min-distraction-delta 0.50 -- python examples\agent_runner_mock.py

Recipes

• LangChain / LangGraph
• MCP Tool Proxy
• CrewAI
• Vercel AI SDK
• GitHub Actions

Terminal GIF

The repository includes a VHS tape at docs/demo.tape.

Render it with VHS:

vhs docs/demo.tape

On Windows, ScreenToGif is also a practical option for recording the terminal benchmark.

Development

Set up:

pip install -e ".[dev]"

Run tests:

python -m pytest

Run lint and type checks:

python -m ruff check .
python -m mypy src scripts

Validate package metadata:

python -c "import tomllib; tomllib.load(open('pyproject.toml','rb')); print('pyproject ok')"

Run the benchmark:

python scripts\benchmark.py --providers heuristic

Build distributable artifacts:

Remove-Item -Recurse -Force dist,build -ErrorAction SilentlyContinue
python -m build

Run the full release gate:

python scripts\release_check.py

Release Checklist

• Confirm the release version in pyproject.toml.
• Run Release Checklist.
• Create a matching GitHub release tag, for example v1.0.0rc1.
• Run the tests and benchmark.
• Run thresholded selection and agent-success evals.
• Clean stale build artifacts, then build the wheel and source distribution.
• Render or update the terminal GIF.
• Verify the README examples still match CLI output.

Project Status

Context Janitor is at v1.0.0rc1: the CLI, config shape, heuristic selector, fallback behavior, cache path, MCP proxy, eval tooling, and packaging flow are release-candidate ready. Before the final v1.0.0 release, the remaining validation target is real-world testing against external tool catalogs and at least one real-log eval pack.