Repository navigation
Home
cybersenses edited this page Nov 15, 2026
·
1 revision
CodingDude is a personal, single-user coding agent. It plans, designs, implements, reviews and ships software-engineering work inside your own repository — built on pydantic-ai, driven from a loopback web UI, with tiered approvals for anything risky.
Current release: 0.4.0 · Licence: AGPL-3.0-or-later · © 2025 Cyberflex
uv sync # install
uv run codingdude serve # web UI on http://127.0.0.1:8765
uv run codingdude run "add a --verbose flag to scripts/report.py"
| Capability | In one line |
|---|---|
| Agentic engine | pydantic-ai agent loop with model failover, per-run worker processes, and context that compacts instead of dying |
| Tiered approvals | safe work runs; pushes, installs and destructive operations stop at an approval card |
| Code tools | AST-aware symbol edits, call graphs, impact analysis, dead code, mutation testing, benchmarks, coverage |
| Planning | PRDs → tech designs → roadmap items → tracer-bullet tickets, each with its own lifecycle and per-step evidence |
| Review gate | a push is refused unless a recorded two-axis review covers the exact diff being published |
| Spend control | hard max_cost_usd / max_total_tokens ceilings checked before each request, plus a session budget cap |
| Observability | JSONL session traces, and optional OTLP export to a collector you run yourself |
| Delegation | parallel sub-agents staffed by cognitive function, with their own spend slices and trace spans |
| Memory | durable facts, a solved-problem bank, project notes, and ADR-style design decisions across sessions |
| Skills | 66 seeded procedural skills, loaded on demand rather than held in context |
- Getting Started — install, configure a model, run your first task.
- Concepts — the vocabulary this project uses precisely (session, workdir, approval, delegation, budget cap).
- Web UI — every panel, and what it is for.
- Planning — the four lifecycles and where each one is stored.
- Review Gate — why publishing requires a recorded review.
- Architecture — server / worker / engine, modules, data model.
- Agents and Tools — toolsets, approvals, failover, delegation, long runs.
- Spend and Context — the three ceilings, the ledger, compaction, caching.
- Observability — traces, OTLP, cost views.
- Security — the posture, and its honest limits.
- Configuration · CLI · Deployment · Troubleshooting
- Skills and Directives — teach the agent new procedures and rules.
- Contributing — gates, review flow, living documentation.
- Release Notes — what changed, most recent first.
| Path | Contents |
|---|---|
~/.codingdude/ (CODEDUDE_HOME) |
config.toml, server-token, mcp.json, skills/, DIRECTIVES.md, SQLite DBs, traces/, logs/
|
| your project | your code, ROADMAP.md, CONTEXT.md, AGENT.md, .codegraph/, optional .codingdude/mcp.json and .codingdude.toml
|
Evidence before claims. The agent must run the gate and read its output before saying work is done — and the review gate makes publishing require a deliberate, axis-labelled review of the exact diff. Nothing here asks you to trust a summary.
This wiki is a navigable summary. The repository is canonical: README · docs/USER-MANUAL.md · docs/TECHNICAL-DESIGN-MANUAL.md.
Start
How it works
Using it
Project