codedebrief 0.13.0

Visual code workflow explorer for coding agents. Builds deterministic, source-grounded Mermaid flowcharts via MCP, local-first, with no API key.

CodeDebrief

Turn source code into deterministic, source-grounded workflow flowcharts on demand.

Months into a project, asking a coding agent how a workflow works often starts a fresh reconstruction from searches, grep commands, and targeted file reads. The answer may be useful, but the process takes time, can miss less obvious paths, and is difficult to verify as a complete source-level view.

CodeDebrief builds the map before the explanation.

It statically analyzes the current source and generates persistent flowcharts of entrypoints, decisions, branches, internal calls, returns, exceptions, and outcomes. A coding agent can select, render, expand, translate, and explain a focused slice, but it does not generate the underlying map.

CodeDebrief is not a documentation generator, a bug finder, a generic graph database, or an LLM enrichment service. It complements maintained documentation with faster, on-demand workflow inspection. The analyzer, artifacts, viewer, and MCP server are local-first and do not require an LLM provider key.

Example output: a compact presentation layer generated from local CodeDebrief artifacts. Canonical workflow visuals are vertical by default; horizontal diagrams are used when the user explicitly asks for a compact overview.

Status: pre-1.0 alpha. The model is versioned, but schema and MCP payloads may evolve before 1.0.

Latest release: v0.13.0 · Website · Changelog

Quick Start

CodeDebrief requires Python 3.10 or newer. Install it from PyPI with uv:

uv tool install codedebrief
codedebrief setup-agent claude

Replace claude with codex, gemini, or cursor for another supported agent surface.

After setup, ask ordinary questions:

Show me the checkout workflow.
Which branches handle a failed payment?
What workflows are affected by this file?
Where is this status handled?
Expand this workflow one level deeper.

For manual exploration:

codedebrief view

CodeDebrief writes three deterministic artifacts:

codedebrief-out/
├── codedebrief.html   local interactive full-project viewer
├── codedebrief.md     reviewable Mermaid flowcharts
└── codedebrief.json   canonical model for MCP, CI, scripts, and the viewer

For explicit refresh during development:

codedebrief update
codedebrief validate --check-sync

To install a pinned GitHub release instead:

uv tool install "git+https://github.com/ferdinandobons/CodeDebrief.git@v0.13.0"

Or install from a source checkout:

git clone https://github.com/ferdinandobons/CodeDebrief.git codedebrief
cd codedebrief
uv tool install .

Why It Exists

A coding agent can reconstruct a workflow from raw source, but that reconstruction normally depends on the searches it performs, the files it selects, and the context it has room to read. Repeating that process is slow, and less obvious branches or cross-file paths can be left out.

CodeDebrief creates a reusable navigation layer before the agent explains anything:

• entrypoints, decisions, branches, calls, outcomes, and source ranges;
• domain concepts such as statuses, roles, permissions, enums, and feature flags;
• affected-workflow context for changed files, symbols, flows, and dependency paths;
• canonical visual slices with stable diagram hashes;
• optional language-friendly labels derived as a presentation layer from analyzer facts.

Where It Fits

CodeDebrief does not replace maintained project documentation. Documentation should capture architecture, intent, invariants, conventions, operational knowledge, and why important decisions were made.

CodeDebrief serves a faster and narrower need: inspecting the workflow encoded in the current source while brainstorming, debugging, planning a change, or reviewing impact. Use it to:

• visualize one focused part of a workflow on demand;
• follow decisions and internal calls across files;
• inspect workflows affected by a file, symbol, or planned change;
• give a coding agent a shared structural slice to reason over immediately.

Documentation explains the system over time. CodeDebrief provides an on-demand visual slice of the current source when you need to reason about it. Any agent-written explanation or language-friendly label is a presentation layer over the analyzer-generated flowchart, not the source of the flowchart itself.

The generated agent instructions treat CodeDebrief artifacts as part of done for workflow-relevant changes: after meaningful source, route, config, or agent-instruction edits, run codedebrief update and codedebrief validate --check-sync before finalizing or committing so MCP answers, Markdown summaries, and codedebrief view use current graphs.

What You Can Verify

A returned workflow_slice can expose:

• source ranges for modeled steps;
• explicit decisions, branches, calls, and outcomes;
• callers, callees, and unresolved-call context;
• omissions caused by token budget, ambiguity, stale artifacts, or unsupported capabilities;
• stable handles for expanding the slice, tracing a path, or opening the related visual.

These signals make the source-level model inspectable. They do not prove business correctness or runtime behavior.

Map First, Then Explain

Once setup-agent has configured the project MCP server and agent instructions, supported agents should start code-logic questions with MCP agent_context. The agent receives a bounded workflow_slice before deciding whether it needs more source, a deeper slice, or a path trace.

Show me a visual workflow for the invitation system.
Explain this code path with a source-grounded flowchart.
Which workflows are affected by this change?
Where is this status handled?
Expand the omitted branches in this workflow_slice.
Rewrite the diagram labels in plain English.

The agent can explain the flowchart, but the static analyzer creates the underlying map.

Setup-Agent

Supported targets:

codedebrief setup-agent codex
codedebrief setup-agent claude ../my-app
codedebrief setup-agent gemini
codedebrief setup-agent cursor --full

The selected target controls which files are written:

Target	Files
codex	AGENTS.md, .agents/skills/codedebrief/SKILL.md, project MCP config
claude	CLAUDE.md, .claude/skills/codedebrief/SKILL.md, project MCP config
gemini	GEMINI.md, .gemini/skills/codedebrief/SKILL.md, .gemini/settings.json MCP config
cursor	.cursor/rules/codedebrief.mdc, project MCP config

setup-agent <target> writes only that target's files. Run it separately for each agent surface you want to configure.

setup-agent creates codedebrief.toml when needed, installs the selected agent instruction file, installs a provider-native CodeDebrief skill where supported, registers project-scoped MCP where supported, generates initial artifacts, runs doctor, and validates the result.

The gemini target follows Gemini CLI / Antigravity conventions: GEMINI.md provides project context, .gemini/skills/codedebrief/SKILL.md provides provider-native workflow guidance, and .gemini/settings.json registers the project-scoped CodeDebrief MCP server.

Agent Workflow

For natural-language questions, agents should start with MCP agent_context.

agent_context returns a workflow_slice with:

• normalized intent and task type;
• primary and supporting flows;
• ordered source-grounded steps;
• decision nodes, branches, values, and outcomes;
• calls, callers, callees, and unresolved call context;
• domain logic for relevant state-like concepts;
• source ranges the agent can cite;
• visual handles for snapshot_slice and codedebrief view;
• omissions caused by token budget, ambiguity, stale artifacts, or unsupported capability;
• follow-up tools for expansion, path tracing, focused explanation, and snapshots.

When a user asks for a visual workflow, the agent should first render the deterministic Mermaid visual returned by CodeDebrief:

1. Render workflow_slice.presentation.canonical_visual.diagram exactly as returned only when the client renders Mermaid inline.
2. If the client cannot render Mermaid inline, use snapshot_slice with include_svg=false and provide the returned .mmd or Mermaid Markdown artifact before prose. Do not paste a long Mermaid code block as the primary visual unless the user asks for raw or copyable Mermaid. These generated files are meant to be opened with the best available local preview path, for example by opening the Markdown artifact in VS Code and using Markdown or Mermaid preview support.
3. Use SVG snapshot artifacts only when the user explicitly asks for SVG or local inspection. They are not the canonical chat visual.

The model may choose the first visible depth, but the text inside shown blocks must come from CodeDebrief payloads. After the visual, the answer should include a short high-level written flow in the user's language, derived from ordered steps, selected flows, decisions, domain logic, and source ranges. That written flow should explain the happy path first and include only the branches needed by the request. The answer should also say that the diagram is a bounded summary and can be expanded. If the user wants a more language-friendly view, the agent may rewrite both labels and the written flow in the user's language as a separate presentation layer, preserving ids or source anchors and without adding facts.

MCP Surface

Primary MCP tools:

Tool	Purpose
agent_context	Default entrypoint for natural-language questions and changed-code context.
expand_slice	Widen or deepen a workflow slice from stable flow handles.
workflow_path	Trace a deterministic path between flows, symbols, or concepts.
snapshot_slice	Render a deterministic visual snapshot for a slice.
explain_flow	Explain one flow with ordered steps, decisions, calls, and source anchors.
explain_node	Explain one flowchart node with local edge and source context.
explain_edge	Explain one modeled edge with source context.
validate_artifacts	Check generated model validity and optional JSON/Markdown sync.
update_codedebrief	Refresh JSON, Markdown, and HTML artifacts from local source.

Use codedebrief view for the manual UI. The CLI intentionally stays small: setup-agent, update, view, validate, doctor, and mcp.

Generated Artifacts

File	Commit?	Purpose
codedebrief.json	Yes	Canonical model consumed by MCP, CI, scripts, and the viewer.
codedebrief.md	Yes	Generated inspection artifact with Mermaid flowcharts. It is not a replacement for maintained project documentation.
codedebrief.html	Usually no	Local interactive viewer generated from the model.

Commit codedebrief.json and codedebrief.md when CodeDebrief is part of the project workflow. codedebrief validate --check-sync verifies that the JSON model matches current source and that codedebrief.md was rendered from that model. Regenerate HTML locally when a human needs the viewer.

codedebrief doctor also reports legacy logicchart MCP server configs left from older installs. Re-run codedebrief setup-agent <target> for the affected agent to replace them with project-scoped codedebrief MCP config.

codedebrief validate --quality reports analyzer health. Its call-resolution rate is based on project calls only: deterministic runtime, standard-library, DOM/browser, and dynamic receiver calls are counted separately as runtime_or_dynamic so they do not look like unresolved workflow edges.

Team Workflow

A simple shared workflow is:

1. Run codedebrief setup-agent <target> in the project.
2. Commit codedebrief.json and codedebrief.md.
3. Let teammates pull the same source-grounded model.
4. After workflow-relevant source, route, config, or agent-instruction changes, run codedebrief update and codedebrief validate --check-sync before finalizing or committing.

The HTML viewer is normally regenerated locally rather than committed.

Manual Viewer

codedebrief view opens the complete interactive flowchart for a human. It is the official manual experience for broad exploration:

codedebrief view
codedebrief view --render-only --no-open

Example manual view: the interactive browser canvas for exploring scopes, files, entry points, and connected workflow flows.

Use the viewer when you need to inspect the whole project graph, navigate scopes, compare neighboring flows, or visually follow callers and callees. Use MCP when an agent should answer a bounded question with a focused workflow_slice.

Domain Logic

CodeDebrief extracts and aggregates domain concepts such as:

• enum members;
• status and lifecycle states;
• roles and permissions;
• feature flags;
• handled values and the decisions that branch on them.

agent_context includes relevant domain logic inside the returned workflow_slice, with links back to flows, nodes, source ranges, snapshots, and viewer targets. Value matching uses modeled code facts, including enum-style suffix matches such as PAID for Status.PAID.

Supported Code

CodeDebrief currently extracts control flow for 11 language ids:

Language	Current coverage
Python (.py)	AST analyzer with functions, methods, decisions, loops, calls, returns, exceptions, tests, enum harvest, and import dependencies.
TypeScript / TSX (.ts, .tsx)	Tree-sitter analyzer with Next.js and React entrypoint detection, decisions, loops, calls, returns, expression-bodied arrows, tests, enum harvest, and import dependencies.
JavaScript / JSX (.js, .jsx, .mjs, .cjs)	Tree-sitter analyzer with JavaScript labeling, decisions, loops, calls, returns, expression-bodied arrows, tests, and import dependencies.
Go (.go)	Profile-driven tree-sitter analyzer with functions, methods, decisions, loops, calls, returns, tests, and import dependencies.
Java (.java)	Profile-driven tree-sitter analyzer with methods, decisions, loops, calls, returns, exceptions, tests, imports, and Spring route annotations.
C# (.cs)	Profile-driven tree-sitter analyzer with methods, decisions, loops, calls, returns, exceptions, and tests.
PHP (.php)	Profile-driven tree-sitter analyzer with functions, methods, decisions, loops, calls, returns, exceptions, and tests.
C (.c, .h)	Profile-driven tree-sitter analyzer with functions, decisions, loops, calls, returns, and tests.
C++ (.cc, .cpp, .cxx, .hh, .hpp, .hxx, .ipp, .tpp)	Profile-driven tree-sitter analyzer with functions, methods, decisions, loops, calls, returns, exceptions, and tests.
Rust (.rs)	Profile-driven tree-sitter analyzer with functions, decisions, loops, calls, match handling, returns, and tests.
Ruby (.rb)	Profile-driven tree-sitter analyzer with methods, decisions, loops, calls, returns, and tests.

Generated models include metadata.language_capabilities, with feature flags and limitation notes per language. Agents should use that contract when explaining analyzer depth.

Privacy

The analyzer, canonical artifacts, Mermaid generation, local viewer, and MCP server do not require an LLM API key. CodeDebrief itself builds the workflow model locally.

When CodeDebrief is connected to a hosted coding agent, the agent may receive the workflow slices and source ranges returned through MCP. That transfer is governed by the coding agent and provider you use.

Configuration

CodeDebrief works without config. Add codedebrief.toml only when defaults are not enough:

[codedebrief]
source_roots = ["."]
exclude = []
exclude_dirs = []
include_public_functions = true
max_call_depth = 4
output_dir = "codedebrief-out"
self_exclude = true

[codedebrief.entrypoints]
include = []
exclude = []

[codedebrief.scopes]
backend = ["backend/**", "services/**"]
frontend = ["frontend/**", "web/**"]
edge = ["edge/**", "workers/**"]

Defaults prune common VCS, dependency, cache, temporary, generated, and output directories, including .git, node_modules, virtualenv folders, .next, .turbo, .svelte-kit, dist, build, out, target, coverage, vendor, Pods, and codedebrief-out.

Limitations

CodeDebrief does not run code, observe runtime state, perform full symbolic execution, prove business correctness, or reconstruct deep framework state. It statically models source files, control flow, selected framework conventions, and resolvable internal calls.

Important practical limits:

• dynamic dispatch may remain unresolved;
• language capability varies by analyzer frontend;
• generated or unsupported files may be skipped;
• large slices are token-budgeted and report omissions;
• the displayed first slice is a bounded summary and can be expanded with MCP tools.

FAQ

Does CodeDebrief replace project documentation?

No. Maintained documentation should explain architecture, intent, design rationale, conventions, and operational knowledge. CodeDebrief is for fast, source-grounded workflow inspection during brainstorming, debugging, planning, and impact analysis.

Does an LLM generate the flowcharts?

No. The analyzer builds the canonical workflow model and Mermaid diagrams from source. A coding agent can choose a bounded slice, render it, translate labels, or explain it, but those actions remain a presentation layer over the deterministic graph.

How is this different from a generated repository summary?

A generated summary is prose produced for one request and one context window. CodeDebrief maintains a reusable structural model that can return bounded flowcharts, source anchors, omissions, and expansion handles for later questions.

Is CodeDebrief an AI code review tool?

No. CodeDebrief is for code comprehension and workflow navigation. It helps humans and coding agents understand modeled logic; it does not present possible defects as product output.

Does CodeDebrief require an LLM API key?

No. The analyzer, artifacts, Mermaid diagrams, manual viewer, and MCP server are local-first and deterministic. Coding agents can use the MCP tools without any required provider key.

How is CodeDebrief different from a call graph?

A call graph shows relationships between symbols. CodeDebrief models workflow slices with entrypoints, decisions, branches, ordered steps, source ranges, domain concepts, visual targets, and expansion tools for coding agents.

Can CodeDebrief generate Mermaid workflow diagrams from code?

Yes. agent_context returns a canonical top-to-bottom Mermaid visual for the selected workflow_slice, and snapshot_slice can persist Mermaid files for clients that cannot render Mermaid inline.

Can I use CodeDebrief only as a manual visual explorer?

Yes. Run codedebrief view to open the interactive local viewer. MCP is the primary agent surface, but the viewer remains the official manual exploration surface.

Development

For local development in this repository:

uv sync --extra dev
uv run codedebrief --help

Standard gates:

UV_CACHE_DIR=/tmp/codedebrief-uv-cache uv run ruff check .
UV_CACHE_DIR=/tmp/codedebrief-uv-cache uv run ruff format --check .
UV_CACHE_DIR=/tmp/codedebrief-uv-cache uv run mypy
UV_CACHE_DIR=/tmp/codedebrief-uv-cache uv run pytest --cov
UV_CACHE_DIR=/tmp/codedebrief-uv-cache uv run codedebrief validate . --check-sync --json

Viewer gates:

npm run viewer:typecheck
npm run viewer:test
npm run viewer:build
UV_CACHE_DIR=/tmp/codedebrief-uv-cache uv run codedebrief update
UV_CACHE_DIR=/tmp/codedebrief-uv-cache uv run codedebrief view --render-only --no-open

Schemas:

• schema/codedebrief.schema.json documents the canonical artifact model.

License

Apache License 2.0. See LICENSE.