Redcon

Redcon selects, compresses, and budgets repository context for coding-agent workflows. It is deterministic, local-first, and built to produce machine-readable artifacts that can be reused in CI, local tooling, and agent middleware.

What It Does

• ranks repository files against a natural-language task
• plans step-by-step context usage across multi-step agent workflows
• packs relevant context under an explicit token budget
• records stable run.json and run.md artifacts
• aggregates historical run.json artifacts into file and directory heatmaps
• reuses cached summaries and an incremental scan index
• supports local multi-repo and monorepo-package workspaces
• exposes an adapter-ready middleware layer for external agent tools

Quickstart

# Install
python3 -m pip install -e .[dev]

# Optional exact local tokenizer backend
python3 -m pip install -e .[tokenizers]

# Rank likely-relevant files
redcon plan "add caching to search API" --repo .

# Plan context across a multi-step agent workflow
redcon plan-agent "refactor auth middleware" --repo .

# Pack context for one repository
redcon pack "refactor auth middleware" --repo . --max-tokens 30000

# Pack context across multiple local repositories or packages
redcon pack "update auth flow across services" --workspace workspace.toml

# Summarize an existing run artifact
redcon report run.json

# Compare two runs
redcon diff old-run.json new-run.json

# Audit a pull request for context growth
redcon pr-audit --repo . --base origin/main --head HEAD

# Compare packing strategies
redcon benchmark "add rate limiting to auth API" --repo .

# Aggregate historical token hotspots
redcon heatmap .

# Refresh scan state once without entering watch mode
redcon watch --repo . --once

Workspaces

Workspace files let one task span multiple local repositories or monorepo packages while keeping the same scan, score, and pack pipeline.

name = "backend-services"

[scan]
include_globs = ["**/*.py", "**/*.ts"]

[budget]
max_tokens = 28000
top_files = 24

[[repos]]
label = "auth-service"
path = "../auth-service"

[[repos]]
label = "billing-service"
path = "../billing-service"
ignore_globs = ["tests/fixtures/**"]

Workspace artifacts add provenance fields without changing single-repo flows:

• workspace
• scanned_repos
• selected_repos
• repo-qualified file paths such as auth-service:src/auth.py

See docs/workspace.md and the examples in examples/workspaces/.

Agent Middleware

The middleware layer sits on top of RedconEngine; it does not duplicate packing logic. It prepares context, optionally enforces policy, and records additive metadata for agent loops.

from redcon import RedconEngine, enforce_budget, prepare_context, record_run

result = prepare_context(
    "update auth flow across services",
    workspace="workspace.toml",
    max_tokens=28000,
    metadata={"agent": "local-demo"},
)

policy = RedconEngine.make_policy(
    max_estimated_input_tokens=28000,
    max_quality_risk_level="medium",
)

checked = enforce_budget(result, policy=policy)
record_run(checked, "agent-run.json")

LocalDemoAgentAdapter is included as a local simulation of how an external tool can call the middleware without introducing any vendor API dependency.

Extension Points

Redcon stays deterministic by default but exposes explicit hooks for local extensions:

• scorer plugins
• compressor plugins
• token-estimator plugins
• summarizer adapters
• telemetry sinks
• agent adapters

Artifacts record active implementations under implementations, along with additive cache, summarizer, token-estimator, workspace, and middleware metadata when those features are active.

Migration Notes

Recent additions are additive rather than disruptive:

• existing single-repo CLI flows stay unchanged
• multi-repo analysis is opt-in through --workspace <workspace.toml> or workspace=...
• workspace TOML files can carry shared config plus [[repos]] entries
• the public Python API now exports RedconMiddleware, AgentTaskRequest, prepare_context(...), enforce_budget(...), record_run(...), and LocalDemoAgentAdapter
• machine-readable artifacts can now include workspace, scanned_repos, selected_repos, implementations, token_estimator, summarizer, and agent_middleware

Detailed upgrade notes: docs/migration.md.

Documentation

• Getting Started
• CLI Reference
• Configuration
• Workspace
• Python API
• Agent Integration
• Plugins
• Architecture
• Migration Notes

Examples and sample outputs: examples/README.md.

License

Redcon uses a dual-license model:

Component	License
Core engine (redcon/core/, redcon/compressors/, redcon/scanners/, redcon/scorers/, redcon/stages/, redcon/schemas/, redcon/plugins/, redcon/cache/, redcon/sdk/, redcon/cli.py, redcon/engine.py)	MIT
CLI and benchmark framework	MIT
Gateway server (redcon/gateway/)	Proprietary
Control plane (redcon/control_plane/)	Proprietary
Agent middleware (redcon/agents/)	Proprietary
LLM integrations (redcon/integrations/)	Proprietary
Runtime and telemetry (redcon/runtime/, redcon/telemetry/)	Proprietary

The open-source core builds community adoption and developer trust. The commercial layer powers Redcon Cloud and enterprise deployments.

For commercial licensing: natjiks@gmail.com