Codexa

Codexa is an edit-lifecycle governance layer for AI coding agents — plan conformance, drift review, and verification crediting — built on a local, deterministic codebase map.

If Codex or Claude Code is the editor, Codexa is the proof layer: it shows what the agent read, what plan it saved, what changed, which checks would earn verification credit, and which gaps still need an honest handoff.

In plain English: it reads a repository, builds a compact index of the files, symbols, imports, tests, risks, workflows, process traces, and graph clusters it can prove, then gives Codex, Claude Code, or another MCP client small evidence-backed packets before and after edits. It is meant to help an agent answer questions like:

• What should I read first?
• What could this change break?
• Which tests are relevant?
• Did my final dirty tree match the plan I saved before editing?
• Did the verification commands the agent reported actually prove anything?

It is not an autonomous coding agent. It does not edit your source files through MCP. It is a context compiler, query server, and verification guide.

Why Codexa

Four capabilities are deliberately hard to find elsewhere:

• A drift loop. change_plan snapshots per-file hashes plus symbol and risk baselines before editing; post_edit_review diffs the real dirty tree against that plan afterwards, rename-aware. When no plan was saved, the pre-edit hook saves an implicit baseline automatically, so the review always has a pre-edit reference; an explicit change_plan upgrades it with planned scope and tests. Blocking is opt-in: only reviews against an explicit plan can surface a blocking verdict to the host — implicit baselines keep the loop informational.
• A verification ledger. Commands the agent reports are parsed against a faithful POSIX-shell subset before earning coverage credit: npm test || true earns nothing, tsc --help is vetoed as non-compiling, sh -c wrappers are unwrapped with ambiguity failing closed. Scope stated plainly: this detects structural exit-masking in reported commands — it cannot detect a wholesale fabricated report. The opt-in AutoVerify lane exists for execution-backed evidence.
• Graph-aware relational packets. v0.7.0 precomputes bounded process packets, functional module clusters, graph-view exports, and opt-in summary prompts. search now reports raw exact-hit counts beside Codexa-ranked anchors, so a "zero grep hits" task can still surface one ranked symbol or file target plus related process and cluster context.
• A fail-closed eval. The eval harness runs real rg/git baselines and fails a scenario outright if the raw baseline does the job better. The archived v0.2.0 release run passed 20/20 scenarios with packets averaging 0.66x the raw baseline output size — and the harness ships in this repo, so you can re-run it yourself. See Public Proof.

Limits, stated up front: TypeScript/JavaScript and Python are the deep lanes (Rust/Go/Java are shallow; other languages get light file facts). Impact expansion caps at graph depth 3. The tested envelope is repos around the ~50K-LOC scale of Codexa itself — expect slower cold indexing and shallower ranking on large monorepos. Everything runs locally: zero API keys and zero network calls in the core paths.

Maintainer Expectations

Codexa is maintained by one person, in spare time, with a deliberately narrow scope. That shapes how this repo works:

• Response times are days to weeks, not hours.
• Scope is narrow on purpose. Deep native language indexers, new LLM analysis layers, broad IDE products, and general-purpose search modes are usually out of scope.
• Not every working PR will be merged. Open an issue first for anything beyond a typo or small docs fix.
• Security issues go through private advisories, not public issues. See SECURITY.md.
• Questions and "is this the right tool?" discussions belong in Discussions, not the issue tracker.

Quick Start

Codexa requires Node.js 22 or newer.

Install from npm:

npm install -g @mirnoorata/codexa

Or work from a checkout:

git clone https://github.com/mirnoorata/codexa.git
cd codexa
npm install
npm run build
npm link

Wire Codexa into another repository:

codexa init /path/to/project --policy-pack            # Codex CLI: .codex/config.toml + hooks + local proof policies
codexa init /path/to/project --claude --policy-pack   # also writes a repo-root .mcp.json for Claude Code
codexa session-start /path/to/project
codexa prove /path/to/project --task "make this change safely"

For a start-to-finish first setup, see the new user tutorial. For a host-focused install choice, see No-Brainer Install Guide.

After codexa init, the target repository gets a repo-local .codex/config.toml entry that lets Codex discover the Codexa MCP server automatically, and with --claude a repo-root .mcp.json so Claude Code discovers the same server (only the codexa entry is managed; other servers in an existing .mcp.json are preserved, and malformed JSON aborts the write). When init runs from an evictable npx cache, generated configs pin npx -y @mirnoorata/codexa@<version> instead of the cache path so they keep working after a cache prune.

Useful flags: the default tool profile for fresh installs is core — only the primary-loop tools (plus impact/freshness) are exposed, which cuts per-turn schema token cost; --tools full exposes all 20 tools, and re-running plain codexa init preserves whichever profile the repo already uses. On the Codex side the core profile relies on Codex CLI honoring enabled_tools (older versions ignore the key and simply expose every tool); the Claude Code .mcp.json path filters server-side via serve --tools core and needs no client support. --agents-md (opt-in) writes a managed Codexa workflow block into the repo's AGENTS.md for Codex, and --claude-md (opt-in) writes the same managed block into CLAUDE.md for Claude Code. The region between the <!-- >>> codexa managed --> / <!-- <<< codexa managed --> markers is reserved: Codexa replaces it in place on every re-run (so the block stays current) and never edits anything outside it. Unbalanced or malformed markers abort the write instead of silently truncating the file.

The installed command is codexa, and the server can also run ad hoc:

npx -y @mirnoorata/codexa serve /path/to/project --auto-refresh

Codexa is also listed in the official MCP registry as io.github.mirnoorata/codexa for MCP clients that discover servers there.

For shared workspace launches such as codexa serve /srv, Codexa can route to the active project recorded in .codex/WORKING.md. Selected session rows win; conflicting active focus, workspace default, or active-session evidence fails closed instead of silently choosing the wrong repo. Use CODEXA_WORKSPACE_SESSION=<session> or --workspace-session <session> when serving a shared workspace root with multiple live workers.

Proof cards and policy packs

codexa prove is the compact "should I trust this agent handoff?" view:

codexa prove /path/to/project --task "change auth timeout behavior" --diff

It reports:

• index freshness and current dirty-tree state;
• read-first files selected from the task and graph context;
• saved change-plan snapshot status, including planned edit targets and planned tests when a snapshot exists;
• verification commands, ledger preview, and reported commands/tests/reports classified with the same command-credit rules as post-edit-review;
• local policy-pack status and remaining proof gaps.

codexa policy-init /path/to/project writes a small local policy pack under .codex/policies/ (verification.json, complexity.json, security.json). codexa init /path/to/project --policy-pack creates the same pack during initial setup. The files are plain JSON, are not executable, and are consumed by codexa prove as bounded local evidence. Neither init nor policy-init overwrite existing policy files unless policy-init --force is passed.

Works with any MCP host

Codexa is deterministic and model-agnostic — its core indexing, ranking, and query paths call no model and need no API keys, so it serves the same evidence-backed context to any agent host that speaks MCP: the OpenAI Codex CLI (repo-local .codex/config.toml), Claude Code (codexa init --claude writes a repo-root .mcp.json; the bundled plugin under integrations/claude-code/ ships its own MCP server entry, hooks that auto-save the pre-edit baseline and surface blocking drift verdicts to the model, and slash commands; --claude-md adds workflow steering — pick the plugin or init --claude for MCP wiring, not both, or Claude Code will register the codexa server twice), and any client that discovers it through the MCP registry. There is no per-model integration to do — the model lives in the host, and Codexa is the host's context server. (The one exception is the opt-in, off-by-default semantic lane, which can call a configured embedding provider such as OpenAI — see Optional Lanes.)

Token discipline is built in: every tool description states its typical output cost, structured results are budget-compacted with truncation records naming dropped fields, hosts with small MCP result limits can set CODEXA_MCP_STRUCTURED_BUDGET_BYTES, and the big retrieval tools accept responseFormat: "concise" for a summary-tier packet that compacts both the structured payload and the text block. The tools/list surface is budgeted too: the per-tool output schema defaults to a compact top-level contract (measured on this repo: 123KB -> 54KB for the full 20-tool surface, 21KB with the core profile; CODEXA_MCP_OUTPUT_SCHEMA=full restores the deep schema), and codexa serve --tools core registers only the primary-loop tools for hosts without a client-side allowlist. Because the budget caps tokens rather than dollars, the savings scale with the host model's price — they matter most on frontier-tier models.

Managed cloud agents

Codexa's stdio transport is for a host running on the same machine as the repository (Codex CLI, Claude Code). Its HTTP transport is loopback-only by design — non-loopback bind addresses and non-loopback Origin headers are rejected — so a hosted agent whose container runs in someone else's cloud (for example a Claude Managed Agents session) cannot reach a local Codexa server over the public network.

The supported way to give a managed cloud agent Codexa context is a self-hosted sandbox: run the agent's tool-execution container in your own infrastructure, alongside a Codexa server, and point the agent's MCP config at Codexa on 127.0.0.1. The agent loop stays on the provider's orchestration layer; tool execution — and the Codexa connection — stay inside your trust boundary, where loopback HTTP is safe. An authenticated remote HTTP mode that would let a provider-hosted container dial into Codexa directly is intentionally not shipped: exposing a codebase context server to the network needs an auth/origin policy Codexa does not yet have, so it is deferred rather than shipped insecure.

The Everyday Workflow

Use Codexa as a guardrail around code changes:

1. Start with session_context or codexa session-start. This tells the agent whether the index is fresh and what loop to use.

2. Search when the target is unclear. search combines bounded raw search, exact/symbol evidence, Codexa ranking, optional semantic retrieval, likely tests, and known gaps.

3. Ask for a task brief before editing. task_brief / brief returns read-first files, impact expansion, risks, snippets, test recommendations, freshness, and next tool guidance.

4. Save a change plan before non-trivial edits. change_plan with saveSnapshot=true, or CLI change-plan --save-snapshot, records the intended scope and test plan. If you skip this step, the pre-edit hooks save an implicit baseline of the dirty tree on the first edit — the review still gets changed-since-baseline and head-drift accuracy, but only an explicit plan enables unplanned-scope drift detection.

5. Review after editing. post_edit_review / post-edit-review compares the actual dirty tree with the saved snapshot, reports drift, and tells you whether to continue, run tests, inspect, or replan.

6. Finish with a test plan and proof card. test_plan recommends targeted commands and shows what they would cover. proof_card / prove binds the handoff to freshness, a saved plan snapshot, local policies, and reported verification evidence.

Primary MCP loop:

session_context -> search(if target unclear) -> task_brief ->
change_plan(saveSnapshot) -> post_edit_review -> test_plan -> proof_card

What Codexa Builds

Running codexa index /path/to/project writes generated files under the target repo's .codex/codebase/ directory:

.codex/codebase/README.md
.codex/codebase/codex-contract.md
.codex/codebase/repo-map.md
.codex/codebase/relational-packets.md
.codex/codebase/relational-packets.json
.codex/codebase/relational-graph.json
.codex/codebase/packet-summary-prompts.ndjson
.codex/codebase/risk-map.md
.codex/codebase/placeholder-map.md
.codex/codebase/test-map.md
.codex/codebase/conventions.md
.codex/codebase/workflows.md
.codex/codebase/freshness.json
.codex/codebase/index.json
.codex/codebase/facts.ndjson
.codex/codebase/modules/
.codex/codebase/playbooks/

For lay readers, these are the maps and checklists Codex reads. For engineers, the durable machine-readable index is index.json plus facts.ndjson; the Markdown files are compact human/agent-facing projections of the same facts. relational-packets.md is the read-first graph packet view for process traces and module clusters; the JSON companions are bounded machine-readable exports for tools and graph visualizers. packet-summary-prompts.ndjson contains explicit opt-in prompt records only — indexing does not call a model.

Generated cache and working state live under .codex/cache/. Codexa-owned cache writes are allowed; source-file mutation is not exposed through MCP tools.

Main Commands

Most context commands auto-refresh stale or missing Codexa artifacts before answering. Use --no-auto-refresh when you intentionally want to inspect only the stored index.

What It Understands

Codexa indexes git-visible files and skips common generated or dependency directories. The source reader is intentionally small and deterministic.

Native parser lanes:

• TypeScript, TSX, JavaScript, and JSX through Tree-sitter plus TypeScript compiler assist.
• Python through Tree-sitter plus lightweight semantic assist.

Shallow deterministic lanes:

• Rust declarations, imports, methods, calls, and tests.
• Go packages, imports, functions, methods, types, constants, variables, and tests with module-aware import resolution.
• Java packages, imports, classes, interfaces, enums, records, methods, and direct call-like usage.

Lightweight file lanes:

• JSON manifests.
• Markdown, MDX, RST, and text docs.
• Shell scripts.
• Systemd service files.

Facts carry explicit confidence:

• authoritative: syntax or git facts Codexa directly read.
• derived: deterministic links, static assists, report-backed relationships, and likely test relationships.
• heuristic: framework hints, string references, dynamic behavior guesses, or risk hints.
• fallback: low-confidence context used only when nothing better is available.

Codexa should never make heuristic-heavy output look stronger than it is.

Architecture For Engineers

Codexa is a TypeScript package with five main layers.

1. Indexing

Entry point: src/indexer.ts.

Pipeline:

1. Discover git-visible files and dirty state.
2. Parse source files and reuse the content-hash parse cache where possible.
3. Import external static-analysis and symbol-report facts.
4. Apply TypeScript/Python semantic assists.
5. Resolve imports, usage sites, aliases, test edges, and graph links.
6. Rank files/modules with centrality, usage, churn, tests, dirty risk, and bounded outcome signals.
7. Build typed graph edges, workflow traces, functional clusters, and relational packet exports.
8. Record freshness, parser errors, and dirty hashes.
9. Publish artifacts atomically.

The indexer uses a cross-process cache lock so parallel Codexa commands do not stampede artifact writes.

2. Fact Model

Core types live in src/types.ts.

Important fact types:

• RepoSnapshot
• File
• Symbol
• UsageSite
• ImportEdge
• TestEdge
• GraphEdge
• WorkflowTrace
• ModuleCluster
• RiskSignal
• ParserError
• SessionMemoryEntry

Important graph edge kinds:

• DEFINES
• IMPORTS
• CALLS
• REFERENCES
• TESTS
• ROUTE
• JOB
• RISK
• ROUTE_HANDLES
• ROUTE_CALLS_STORE
• STORE_DISPATCHES_ADAPTER
• ADAPTER_REFERENCED_BY_MANIFEST
• UI_CALLS_ENDPOINT
• TEST_COVERS_WORKFLOW
• IMPLEMENTS
• EXTENDS
• EXPORTS
• TYPE_EXPORTS

Relationship claims can include EdgeEvidenceV1, which carries edge kind, source, confidence, reason, path/symbol endpoints, optional range, and stale/degraded flags.

3. Query Layer

Public query exports live in src/queries.ts, intentionally kept as a thin barrel. Implementations live under src/query/.

Key query modules:

• search.ts: repo maps, raw/BM25/exact/symbol/semantic search, target discovery, raw-exact-vs-ranked anchor reporting, and relational process / cluster packet selection.
• context.ts: context_pack, task_brief, focus_brief, and session_context.
• impact.ts: file/symbol blast-radius expansion and verification recipes.
• graph-traversal.ts: callers, callees, and dependency paths.
• workflow.ts: route/job/manifest workflow traces.
• change-plan.ts: pre-edit plans and saved snapshots.
• post-edit.ts: dirty-tree review against saved snapshots.
• test-plan.ts and tests.ts: test recommendations and provenance.
• verification.ts: command coverage, command envelopes, and verification ledger entries.
• session-memory.ts: cache-only working memory queries.

Query sessions (src/query/session.ts) carry the repo root, loaded index, freshness, git state, command budget, warnings, provenance, changed files, and changed symbols. Worktree inspection is allowed to degrade; an empty changed-file set with degradation warnings means "unknown", not "clean".

4. MCP Server

Entry point: src/mcp.ts.

Codexa registers a query-only MCP server. Stdio is the default transport for local Codex use. Streamable HTTP is available only on loopback addresses unless future auth/origin policy is added.

MCP tools:

freshness
repo_map
find_context
search
placeholder_report
symbol_context
impact
diff_impact
test_plan
task_brief
context_pack
focus_brief
session_context
callers
callees
dependency_path
workflow_path
change_plan
post_edit_review
session_memory

MCP resources expose generated .codex/codebase/ artifacts. MCP prompts expose small workflow prompts for impact-before-edit, dirty-diff review, snapshot edit loops, and targeted test planning.

MCP tools may update Codexa-generated artifacts or cache state when auto-refresh, snapshots, or session memory are enabled. They do not expose a source-editing tool.

5. Adapters, Packaging, And Release Tools

Adapters:

• src/cli.ts: Commander-based CLI.
• src/init.ts: repo-local MCP config and hook setup.
• integrations/claude-code/: Claude Code plugin, hooks, and slash commands.
• plugins/codexa/: Codex plugin bundle with manifest, skill, and MCP wrapper.

Operational tools:

• src/doctor.ts: local readiness checks.
• src/github-sync.ts: git/GitHub sync diagnostics.
• src/github-release.ts: release notes, tags, and GitHub Release flow.
• scripts/*.mjs and scripts/*.sh: source hygiene, privacy, package smoke, public snapshot, benchmark, and publish gates.

Optional Lanes

Semantic Retrieval

Semantic retrieval is opt-in and cache-based.

Build the cache:

codexa semantic-index /path/to/project --provider openai
codexa semantic-index /path/to/project --provider local-command --command ./embed-jsonl

After the cache exists, query commands can use it automatically when the snapshot and provider settings match. --semantic forces diagnostics, and --no-semantic disables the lane for one call.

OpenAI uses OPENAI_API_KEY and defaults to text-embedding-3-small. local-command receives JSONL on stdin and returns embedding records. Codexa does not ship a vector database and does not call embedding providers unless the semantic cache/provider path is configured or explicitly forced.

LSP Assist

LSP assist is read-only and bounded. Enable it with --lsp or CODEXA_LSP=1 on supported query commands.

Codexa can query:

• typescript-language-server --stdio
• basedpyright-langserver --stdio
• pyright-langserver --stdio

LSP failures are warnings in the packet, not hard failures. LSP never edits source files.

Static Analysis Reports

Codexa does not vendor Semgrep, CodeQL, ShellCheck, or other scanner engines. The default safe shape is report ingestion:

codexa static-analysis /path/to/project \
  --semgrep-report /tmp/semgrep.json \
  --codeql-report /tmp/codeql.sarif \
  --symbol-report /tmp/codexa-symbols.json \
  --scip-report /tmp/index.scip.json

Codexa also accepts a bounded CodexaSymbolReportV1 JSON document so external language tools can feed symbols and relationships into Codexa with derived confidence. SCIP reports are accepted as JSON exported by scip print --json; Codexa converts them into the same bounded symbol-report lane and does not run or vendor SCIP indexers.

Scanner execution flags such as --run-semgrep, --run-codeql, and --run-shellcheck are explicit opt-ins. They run installed local tools under scrubbed environments and write reports under .codex/static-analysis/.

AutoVerify Hooks

codexa init writes advisory hooks when Codex hooks are available:

• hook-pre-edit saves an implicit pre-edit baseline when no change-plan snapshot exists (and reminds the agent that an explicit change_plan upgrades it with planned scope and tests).
• hook-post-edit runs a bounded post-edit review after edits.

AutoVerify command execution is disabled unless user-owned autonomy is full-access or the environment sets CODEXA_AUTOVERIFY=1 / CODEXA_AUTOVERIFY=true. Even then, AutoVerify is hook-only. MCP post_edit_review never spawns commands.

AutoVerify is not a sandbox. Test code still runs locally with the user's file permissions. Codexa records whether verification mutated source/test/provenance state and treats such reports as non-covering evidence.

Source Map

Safety Boundaries

Codexa is deliberately constrained:

• Local-first by default.
• Query-only MCP surface.
• No source-mutating MCP tools.
• No graph database.
• No vector database.
• No web UI.
• No mandatory embeddings.
• No always-on LSP daemon.
• No hidden scanner execution.
• No broad host-specific planning layer.
• No project-specific private rules in the public setup path.

Context commands can refresh generated .codex/codebase/ artifacts. Snapshot and session-memory tools can write under .codex/cache/. Those are Codexa-owned state paths, not source edits.

Testing And Verification

Common development commands:

npm run typecheck
npm run lint
npm run privacy
npm test
npm run check

npm run check runs typecheck, source hygiene, release-path hygiene, privacy, Claude Code smoke tests, and the Vitest suite.

Release-oriented checks:

npm run smoke:package
npm run benchmark:ci
npm run public:snapshot-check
npm run package:hygiene
npm run security:check

security:check runs the development gate, dependency audit, clean-tree public snapshot verification, package hygiene, and installed-package smoke test. The public snapshot check intentionally refuses a dirty tree so the verified archive matches HEAD.

Public Proof

Codexa has a structured eval harness:

node dist/cli.js index /path/to/project
node dist/cli.js eval /path/to/project --suite all --seed codexa-v1-benchmark

The eval scores structured query data, not prose. It compares Codexa packets against raw rg/git status baselines, tracks recall/precision/test recommendations/context size, and can run ranking experiments without changing production ranking. The claim is deliberately falsifiable: a scenario fails outright if the raw-grep baseline does the job better, and the harness runs in CI on every push (npm run eval:ci in the check workflow, seeded per commit so the synthetic holdouts cannot be overfitted) — "beats grep on its scenarios" is a gate, not a one-off benchmark.

Measured results for v0.3.0 (seed codexa-v030-eval, full suite, archived in reports/benchmarks/v0.3.0-eval.json):

Known imperfections in that run, recorded by the harness itself: 2 false-positive impact files and 1 broad-retrieval failure (synthetic-session-context-seedless) — see the calibrationSummary block in the archived report. The previous run is kept at reports/benchmarks/v0.2.0-eval.json.

Do not update public benchmark claims without rerunning the eval on the current checkout and current target.

GitHub Release Timeline

Use GitHub Releases as the visible source timeline for the current project.

Source sync diagnostic:

codexa github-sync-check /path/to/codexa-checkout
codexa github-sync-check /path/to/codexa-checkout --no-network

GitHub Release dry run and real release:

npm run release:github:dry-run -- --tag v0.2.0
npm run release:github -- --tag v0.2.0

The release command generates a changelog-style summary, changed-area summary, restore commands, branch/worktree continuation commands, and forward-only PR rollback commands. Official releases should come from a clean main after the normal GitHub flow has landed.

Release Automation

Release Please runs after pushes to main. It reads conventional commits, opens or updates a release PR with the package version and changelog changes, and creates the GitHub Release after that release PR is merged.

This does not publish npm on every main merge. Normal feature and fix PRs land on main first, Release Please batches releasable changes into its release PR, and npm publishing stays downstream of the GitHub Release event.

Configure a RELEASE_PLEASE_TOKEN GitHub repository secret with a personal access token that can create pull requests, tags, and releases. Do not use the default GITHUB_TOKEN for Release Please if npm publishing should happen automatically, because releases created by GITHUB_TOKEN do not trigger the separate release: published npm workflow.

npm Package Publishing

The npm package is published by GitHub Actions after the GitHub Release lane publishes a release. The trigger is release: published; pushed tags alone do not publish to npm. The workflow checks the released tag, package identity, repository URL, version availability, and npm run security:check, then runs:

npm publish --registry https://registry.npmjs.org --access public --tag latest --provenance --ignore-scripts

For the first public npm release, configure an NPM_TOKEN GitHub repository secret with publish access. After the package exists and npm trusted publishing is configured, the workflow can remove token-based publishing while keeping the same release gate and --ignore-scripts protection.

Contributing

Read CONTRIBUTING.md before opening a PR.

What usually fits:

• Bug fixes with clear reproduction and regression tests.
• Performance improvements with before/after measurements.
• Documentation fixes.
• Targeted improvements to existing commands or MCP tools.

What usually does not fit:

• New deep language indexers.
• New LLM-based analysis layers.
• Whole-file rewrites for style preference.
• Heavy dependencies where a small deterministic helper is enough.
• New source-mutating agent behavior.

Run this before proposing code changes:

npm run check

License

MIT. See LICENSE.