AnaxiGraph

Understand the system. Guide the agent. Keep the architecture coherent.
Shared architecture intelligence for people and AI coding agents.

Get started · Agent plugin · Docker · Advanced · Contribute

AI makes it easy to add code faster than a team can understand the architecture absorbing it. Hidden coupling, duplicated responsibilities, inconsistent abstractions, and one-off agent changes quietly become spaghetti code.

AnaxiGraph is the shared architecture intelligence layer for humans and AI agents. It turns a repository and its Git history into a living model of what the software does, how its parts work together, why they exist, and how the design is changing. A person can explore that model in the dashboard; a coding agent can use the same evidence through AnaxiMCP before and after it edits code.

AnaxiGraph does not hand out a magic architecture score and does not edit the analyzed repository. It keeps facts read from code, AI-created interpretations, and recommendations distinct so that beginners can read a plain-language conclusion and experts can inspect the evidence behind it.

🚀 Start in four steps

You need Git, Python 3.11+, and uv.

1. Run one command in the repository

cd /path/to/your/repository
uvx anaxigraph up . --open --semantic agent --connect codex

Use --connect claude for Claude Code. Omit --semantic agent --connect codex when you only want the code-only map.

This command creates or loads repository policy, stores AnaxiIndex outside the target, completes the current scan, starts the loopback dashboard and AnaxiMCP, and builds representative Git history in the background. Stop it with Ctrl-C; restart with the same command.

2. Open the dashboard

Visit http://127.0.0.1:8765. Current architecture is ready before background history finishes.

3. Restart Codex in the repository

The explicit --connect codex option configures http://127.0.0.1:8765/mcp on the machine where Codex runs. Restart it after first-time setup:

cd /path/to/your/repository
codex

4. Ask it to build the AI-created code map

Use AnaxiGraph to build or resume the AI-created code map for this repository. Choose an explicit economical worker model, start or reuse the background run, and return its run ID. Let the worker manage the queue. Check compact progress when I ask; do not supervise it in an LLM loop or edit source while mapping.

The durable command survives the invoking Codex session:

anaxigraph understand . --executor codex --model '<worker-model>' --parallel-jobs 32 --background --json
anaxigraph semantic-status . --compact --json

Replace <worker-model> with your chosen Codex model. For Claude Haiku workers, use --executor claude --model haiku. The repository's max_parallel_jobs setting limits total concurrency across executors; raise it to at least 32 for 32 simultaneous jobs.

With semantic.provider: agent, understand auto-detects an invoking Codex or Claude session and uses that authenticated local CLI as a read-only semantic executor. Use --executor codex or --executor claude to select one explicitly. Use --executor mcp when the already-connected agent should perform the MCP work loop itself; that mode returns status: agent_action_required—meaning the agent still has work to do—until it has submitted every task. --background includes the complete saved task list, records a durable run handoff in user state, and keeps the host worker alive if the coding-agent session exits. semantic-status reports whether that worker is really running, where its log is, whether it finished, which saved index it is using, and its model and reasoning effort. Choose --model explicitly to avoid inheriting an expensive CLI default; optional --reasoning-effort is also a runtime choice. Changing either never makes an existing AI description stale. already_running returns the existing run. Return that handle and check compact status on request or no more often than every five minutes; ordinary software handles waiting, scheduling, and retries. Direct MCP looping handles one limited task at a time when no authenticated host worker is available; it is not the default way to build the complete map.

When the loopback dashboard is already running, understand matches the checkout to its service by Git remote identity and executes against that sidecar's AnaxiIndex—even when the container sees the checkout at /repo. It never creates a second default database in that case. Without a matching service it uses the same stable per-checkout user-state path as anaxigraph up; a timeout or invalid inventory fails closed instead of silently choosing another index. --db explicitly selects a standalone index, and --service-url explicitly selects a service. Every command result reports the chosen index.authority and physical/service identity for unambiguous handoff.

That is the key cost model: the chosen workers spend model tokens on code understanding. The caller can launch cheaper workers without passing its conversation to them. AnaxiGraph needs no model key in provider: agent mode. It gives each worker a limited page of evidence for one file or code area at a time, checks the returned structured description, records which worker and model created it, and resumes unfinished work in a later session. Once every file has a current description, the same workflow automatically proposes a responsibility-based area/subsystem map, runs independent AI critic/revision passes, and applies exact membership and size checks in code. It then produces a versioned Living Architecture Charter: purpose, actors, observable capabilities, responsibilities, important flows, public contracts, invariants, extension points, patterns, coherence concerns, conflicts, unknowns, and a behavior-only Capability Brief for fresh-context review. There is no human approval gate: the result is versioned map metadata and never edits or controls the analyzed code. Unchanged fingerprints avoid rereading unchanged files or rebuilding unchanged higher-level understanding. Read the same Charter in Overview, anaxigraph charter ., or ANAXIGRAPH_OVERVIEW; use the Map selector or the responsibility map embedded in ANAXIGRAPH_OVERVIEW for its area/subsystem structure.

A deterministic scan exposes an honest provisional Charter immediately. AI synthesis replaces it only when current evidence is complete; a prior Charter is labeled stale after relevant evidence changes. Optional declared context can clarify an inferred claim without overwriting it:

anaxigraph charter . \
  --correct-section purpose \
  --statement "Help teams keep AI-assisted code architecturally coherent." \
  --author "repository owner" \
  --rationale "This intended outcome is not fully visible in source structure."

The overlay retains the inferred statement, author, time, and rationale in AnaxiIndex. Repeat with --withdraw to remove it from the current presentation. Human input is never required for Charter generation, refresh, or agent use.

The complete onboarding guide explains the normal coding loop and setup diagnostics.

The installed AnaxiGraph skill carries the normal coding loop. The dashboard's /api/glossary response exposes the same finding-state and measurement meanings for clients that render the human interface; they are not duplicated as another agent tool.

🔁 Use one coding loop

Keep one persistent AnaxiGraph service running throughout the coding session. The service supervises its structural watcher internally, so ordinary saves update cheap structural facts without a second container or another model-backed repository pass. Build the complete AI map once when needed.

Give the connected agent one concrete goal. ANAXIGRAPH_GUIDE is the front door: use intent="understand" for the system map, build for placement, improve for a bounded structural change, redesign for the capability-first fresh-eyes review, and reassess after editing.

Use AnaxiGraph to guide “add saved prompt exports.” Find the smallest relevant file set, tell me where the code belongs, inspect what depends on shared files, identify the focused checks, and refresh the shared map after implementation.

The agent follows one sequence:

1. Guide — ANAXIGRAPH_GUIDE(intent="build"|"improve") returns one evidence-backed recommendation, likely files, placement, counter-reasons, bounded impact, tests, and risks.
2. Impact — ANAXIGRAPH_IMPACT shows direct dependants of the shared files before they change.
3. Change — edit source and run focused tests through the normal coding workflow. AnaxiGraph observes the repository; it does not edit it.
4. Refresh and reassess — request ANAXIGRAPH_SCAN(refresh_semantics=true), finish any prepared AI work, then call ANAXIGRAPH_GUIDE(intent="reassess"). The scan compares code fingerprints, reuses unchanged file meanings, and prepares only structurally changed files plus context whose interfaces or relationships changed. The shared before/after response explains observed changes, architectural consequences, possible improvements or regressions, reasons to leave the design alone, and the smallest safe verification step. It creates no approval or change-management state. The same result is visible under Changes or from anaxigraph reassess .. Use History when you need the wider introduction, recurrence, churn, or co-change context.

With the default semantic.refresh: on_scan, an explicit scan prepares those changed and affected scopes automatically. anaxigraph update . --prepare-semantics is the explicit equivalent for a repository that keeps manual refresh. Run anaxigraph understand . --executor codex --background to execute the prepared work. Static reassessment is available immediately; ask again after semantically_ready when the decision needs refreshed responsibility, duplication, pattern, or possible-unused-code evidence.

The reviewed responsibility map has its own incremental validation boundary. It is carried forward when the included file inventory and map policy still match and the fraction of unchanged intrinsic module roles meets semantic.taxonomy.stability_bias (default 0.8). New or removed files, changed map constraints, or broader responsibility drift trigger a fresh proposal and its configured agent reviews. A carried taxonomy does not suppress affected context, group, repository, or pattern work. Planning responses expose this decision and the currently queued cascade under work_plan.

Each guidance and impact reply includes server time, payload size, and model-token use. Semantic status groups AI jobs by action with current-snapshot and lifetime time/token totals, while scan results and detached execution records show wall-clock duration. Successful and failed attempts contribute token totals when the executor reports them. A process killed before it emits usage is still labeled unreported; zero is never presented as proof that the model call was free.

A changed metric or finding is not automatically an improvement. Expected behavior, focused tests, and architecture evidence must agree. The onboarding guide explains the same loop; lower-level and operator workflows stay in the advanced guide.

👀 Challenge the design with fresh eyes

For a major refactor, AnaxiGraph can deliberately step outside the current package layout instead of asking an agent already immersed in the repository to redesign what it just read. Start the fixed review explicitly:

anaxigraph fresh-eyes . --start --proposals 2
anaxigraph understand . --executor codex --background
anaxigraph fresh-eyes .

The clean-sheet agents receive only the behavior-only Capability Brief and external constraints—no current paths, frameworks, findings, history, or architecture map. A blind adjudicator preserves meaningful disagreement, then a repository-aware pass compares that reference design with what is actually built. A final mission filter keeps only small, justified recommendations and records reasons not to proceed. One proposal is the lower-cost mode, two is the recommended default, and three is optional.

Two providers can answer one review. Pin an executor family to each proposal slot, then start one host worker per executor; each executor owns its own background run slot for the repository:

anaxigraph fresh-eyes . --start --proposals 2 --proposal-executors codex,claude
anaxigraph understand . --executor codex --background
anaxigraph understand . --executor claude --background

A pinned slot is only handed to the executor family it names, so a worker that finds only the other executor's slot reports waiting_for_executor with the exact command that starts it and keeps polling instead of calling the queue complete. Both workers claim at the same time only when semantic.max_parallel_jobs is 2 or more. The review reports cross_provider: true only when two different executor families actually produced the proposals; two sessions of one provider stay false with the recorded caveat. A pin applies to a new review or a --restart generation, and anaxigraph fresh-eyes . --unpin releases every assignment when the second executor is never started.

The connected Codex or Claude executor supplies the model context and tokens; AnaxiGraph supplies bounded evidence, validates each result, and resumes the saved stages after interruption. The Codex adapter delivers oversized requests through temporary evidence pages in the same isolated, read-only workspace, preserving the complete payload without exceeding the CLI's inline-input limit. The pages add no review stages and are removed on completion or a handled failure. The dashboard exposes the same review under Improve → Fresh eyes. A connected agent can read or start it through ANAXIGRAPH_GUIDE(intent="redesign", start=true, proposal_count=2). The older fresh_eyes=true form remains accepted. Starting a review never edits source or automatically accepts a recommendation.

Changing executor, model, or reasoning effort does not silently invalidate a completed review. When you deliberately want a new architectural judgment, run anaxigraph fresh-eyes . --restart (a connected agent requests the same rerun with ANAXIGRAPH_GUIDE(intent="redesign", start=true, restart=true)) and then run anaxigraph understand with your explicit runtime model and reasoning-effort choices. The new generation reruns every Fresh-Eyes stage while preserving the earlier documents for audit; it does not force unchanged module dossiers to be regenerated.

Every review lists the generations it has recorded, with the executor models, stage timings, output volume, token counts, and attempts of each. Read an earlier one in full with anaxigraph fresh-eyes . --generation 2, GET /api/fresh-eyes?generation=2, ANAXIGRAPH_GUIDE(intent="redesign", generation=2), or the generation selector in Improve → Fresh eyes. A recorded generation is history: it reports superseded and cannot be started, retried, or restarted.

🐳 Durable Docker sidecar

If you prefer an isolated, persistent container beside the repository:

cd /path/to/your/repository
uvx anaxigraph init . --start --semantic agent --connect codex

The generated Compose service mounts source read-only, drops Linux capabilities, enables no-new-privileges, persists AnaxiIndex in a named volume, and publishes only to loopback by default. Use --connect claude for Claude Code. Preview the full repository and client change with --dry-run --json.

See Docker operation for manual Compose review, updates, watchers, and the experimental multi-repository registry.

🔌 Install the guided agent workflow

The shared plugin teaches Codex and Claude Code how to select the right indexed repository, build or resume the AI-created code map, find a small set of likely files and affected callers, hand off a planned finding, and verify a completed change.

Codex:

codex plugin marketplace add hcekne/anaxigraph && \
  codex plugin add anaxigraph@anaxigraph

Invoke $anaxigraph. Claude Code:

claude plugin marketplace add hcekne/anaxigraph && \
  claude plugin install anaxigraph@anaxigraph --scope user

Invoke /anaxigraph:anaxigraph. The plugin includes the default loopback MCP connection, so plugin users may omit --connect from the start command. See the agent plugin guide for the safety contract and custom endpoint behavior.

How it works

source + Git ── facts, hashes, relationships, history ──→ versioned AnaxiIndex
                                                               │
                                  ┌────────────────────────────┴──────────────────────┐
                                  ▼                                                   ▼
                         human dashboard                                      AnaxiMCP for agents
                    understand and investigate                         plan, inspect impact, verify
                                  │                                                   │
                                  └────────────────────┬──────────────────────────────┘
                                                       ▼
                                             better shared decisions

changed or stale modules ── bounded evidence ──→ connected agent using its own tokens
                                                       │
                                                       └──→ checked descriptions in AnaxiIndex

Structural refresh and semantic execution are separate operations. A dashboard Refresh scan runs asynchronously with observable progress and safe cancellation; semantic prepare/resume uses the already-current snapshot and never hides a structural rescan inside the command.

Three named surfaces share one index:

• AnaxiGraph is the scanner, dashboard, and overall project.
• AnaxiIndex is the SQLite record of repositories, files, named code parts, direct code links, findings, history, and AI-created descriptions.
• AnaxiMCP gives coding agents size-limited repository evidence and controlled ways to update the external index.

AnaxiGraph does not execute target code and does not edit repository source. A generated sidecar mounts the target read-only. The target needs only optional .anaxigraph.yml policy; analysis state stays external.

Facts are not opinions

AnaxiGraph deliberately separates:

1. facts read directly from repository data—hashes, code structure, named code parts, direct links, Git changes, branch counts, imported test coverage, and which analyzer produced them;
2. AI explanations—purpose, responsibilities, role in the repository, related behavior, and pattern opportunities, each with the model, instructions, evidence, and evidence-strength rating that produced it; and
3. recommendations—reviewable proposals with evidence, counter-evidence, cost, safety, and lifecycle state.

Relationship edges say whether they are resolved, ambiguous, unresolved, or external. Dynamic runtime wiring can still be invisible, so a missing edge is never presented as proof of dead code.

One index, several views

The map selector distinguishes four sources instead of blending them: Current view uses optional declared intent first, then the AI-reviewed Responsibility map, then the deterministic Path map fallback; Declared map shows repository policy alone. Stable group identities are kept separate from their display labels, and history uses today's current-view frame by default so the same regions visibly fill and connect over time.

• Understand combines the Living Architecture Charter, overview, file inventory, and graph.
• Guide answers where to build and how to refactor with the same recommendation for a person or coding agent.
• Improve combines the ranked finding record and reviewed pattern intelligence.
• Changes replays representative first-parent commits and supports reassessment after edits.
• Settings owns repositories, readiness, refresh, and progressively disclosed operations.
• Pattern intelligence also exposes finalized evaluations through anaxigraph patterns and the paged /api/patterns endpoint. Guidance includes relevant recommendations directly for coding agents. Each result leads with a conclusion, evidence, action, cautions, and verification; its nine exact ratings are grouped and explained instead of shown as a number wall. Candidate results likewise explain why a pair was selected or skipped, what evidence is missing, and why the internal selection order is not itself a pattern recommendation.

anaxigraph search "goal or code name" . uses the same bounded SQLite FTS ranking as dashboard search, ANAXIGRAPH_SEARCH, and the first step of architecture guidance. It searches paths, filenames, symbols, summaries, responsibilities, contracts, and normalized aliases, then reports the semantic provenance that contributed to each result.

🎯 Findings are a workflow, not a wall

The default attention list shows at most 20 useful findings and excludes routine long-function notes. The complete record remains filterable and paginated; no evidence is deleted merely to quiet the UI.

Every finding says what AnaxiGraph saw, why it may matter, what to do, when the code may be fine as it is, and how to check the result. The dashboard, REST API, MCP tools, guidance results, and copied agent prompt use the same wording. Exact rule IDs, evidence values, and ordering scores remain structured fields for automation, but each field has an adjacent ordinary-language meaning; they are never dumped into a jargon-filled “technical details” section. Plan agent work selects a finding for implementation; its handoff also says which retained code map first shows the problem, where it disappears, and whether it later returns. Resolution and regression normally come from a later scan. Retained maps are samples, so the named frame bounds the change rather than claiming that every Git commit was analyzed.

Current support boundary

Python uses its built-in AST. JavaScript, JSX, TypeScript, and TSX use pinned Tree-sitter grammars for structural symbols, imports and re-exports, CommonJS, modern dynamic imports, source spans, TypeScript declarations, and syntax-level type evidence. Workspace targets resolve from indexed package.json, tsconfig.json, and jsconfig.json evidence with explicit provenance and honest ambiguity; AnaxiGraph does not run the target build or pretend it has compiler/type-checker or runtime certainty. Other recognized source and text formats have heuristic or inventory support. The roadmap deliberately does not call extension recognition “full language support.” See the capability matrix.

Linux x86-64 is release-gated. Linux ARM64, macOS, and WSL2 are best effort; Docker Desktop is the recommended macOS path. Native Windows is not supported—use WSL2. See the platform matrix.

The REST and MCP service is a local sidecar. Keep it bound to loopback or access it through a trusted SSH tunnel; do not expose the port to an untrusted network.

Advanced operation

The advanced guide covers local Codex/Claude/custom executors, semantic cost and privacy, SSH forwarding, custom ports/state, optional coverage imports, durable history controls, watchers, integrity diagnostics, upgrades, resets, lower-level CLI commands, and several repositories.

🛠️ Development

uv sync --extra dev
uv run pre-commit install --install-hooks
uv run python scripts/run_quality_gate.py --base origin/main

The quality gate includes a fresh, repeatable AnaxiGraph scan of this repository. Its full report is retained in CI and compared with quality/self-analysis-baseline.json; new or worsened warning/error findings fail while unchanged information-level diagnostics remain visible and non-blocking. Run it directly with uv run python scripts/check_self_analysis.py --output /tmp/anaxigraph-self-analysis.json.

The product brief is repo_instructions.md, the consecutive roadmap is docs/feature-development-plan.md, and the release contract is docs/releasing.md. Contributions are welcome; see CONTRIBUTING.md.