CodeDecay v0.4.1

CodeDecay v0.4.1

CodeDecay v0.4.1 is the final Codex India Hackathon 2026 submission release.
It keeps the project local-first and makes the AI-assisted PR safety workflow
easier to reproduce, inspect, and hand back to a user-owned coding agent.

Patch from v0.4.0

This patch does not expand the frozen feature scope. It establishes one final
public release identity across npm, the CLI, documentation, Judge Lab evidence,
and the captioned demo after npm publishing access became available. The
existing v0.4.0 tag and release remain unchanged as immutable history.

Shipped

• A recommended codedecay ai workflow that packages preflight, deterministic
  PR evidence, repair tasks, configured checks, and current-tree revalidation.
• Packed CLI and GitHub Action acceptance that exercises the same public
  command surface without workspace-only dependencies.
• Behavior-specific edge-case scenarios ranked by changed routes, symbols,
  user flows, requirements, memory, proof type, and confidence, with
  machine-readable overflow retained.
• Agent task bundles whose limitations accurately reflect whether configured
  commands or explicit model investigation ran.
• A bundled MCP CLI boundary that keeps its server dependency graph out of
  fresh npm consumer installs.
• A public, credential-free Judge Lab with fixed safe scenarios, reproducible
  analyzer evidence, benchmark disclosure, and an embedded captioned Codex
  repair demo.
• A genuine Codex repair record with a sanitized transcript, exact fixture
  commits and commands, red-before-green endpoint proof, and current-tree
  CodeDecay revalidation.
• Parser-grounded JavaScript security matchers that require real command sinks
  and data flow for path traversal and hardcoded-secret evidence.
• Output-aware analysis that prevents a generated JSON, Markdown, or SARIF
  report from becoming evidence in the next run while preserving intentional
  CodeDecay config, memory, and arbitrary user files.
• A parser-backed copied-implementation gate that suppresses type shapes,
  type-only imports, and declarative fixtures while preserving executable
  copied-oracle detection.

Reproducible evaluation

The deterministic fixture corpus reports 23 of 23 planted issues recalled and
a 2.22% false-positive rate at zero model cost. These are finite planted-fixture
results, not production accuracy or a guarantee that a merge is safe.

The final release validation includes the full repository suite, package build
and dry-run, child-repository acceptance, PR-safety evaluation, deterministic
benchmark, fresh tarball consumer, published-package smoke, and Judge Lab
server/browser checks.

Safety and privacy

• No required CodeDecay API key.
• No hidden telemetry.
• No hidden model call or private-code upload.
• Commands run only when explicitly configured and allowed.
• Deterministic/tool evidence remains distinct from agent suggestions.
• The public Judge Lab accepts fixed scenario IDs only; it does not accept a
  repository URL, prompt, upload, model credential, or arbitrary command.

Known limitations

• The controlled paired real-agent outcome benchmark in
  #683 is not complete.
• Independent three-participant published-workflow UAT in
  #692 is not complete.
• The public benchmark is planted and finite, and deterministic findings can
  produce false positives.
• Framework/language coverage is incomplete; runtime confidence depends on the
  repository's configured checks and available tools.
• Copied-implementation detection remains a conservative three-line heuristic;
  unsupported syntax, shorter copies, or structurally rewritten logic may not
  be detected.
• The demo fixture deliberately does not model token expiry, malformed JWTs,
  or authorization scope, and its repaired report remains High without runtime
  coverage evidence.

CodeDecay helps find what a coding agent missed before merge. It does not
guarantee perfect safety.

CodeDecay v0.4.0

CodeDecay v0.4.0

CodeDecay v0.4.0 is the frozen Codex India Hackathon 2026 submission release.
It keeps the project local-first and makes the AI-assisted PR safety workflow
easier to reproduce, inspect, and hand back to a user-owned coding agent.

Shipped

• A recommended codedecay ai workflow that packages preflight, deterministic
  PR evidence, repair tasks, configured checks, and current-tree revalidation.
• Packed CLI and GitHub Action acceptance that exercises the same public
  command surface without workspace-only dependencies.
• Behavior-specific edge-case scenarios ranked by changed routes, symbols,
  user flows, requirements, memory, proof type, and confidence, with
  machine-readable overflow retained.
• Agent task bundles whose limitations accurately reflect whether configured
  commands or explicit model investigation ran.
• A bundled MCP CLI boundary that keeps its server dependency graph out of
  fresh npm consumer installs.
• A public, credential-free Judge Lab with fixed safe scenarios, reproducible
  analyzer evidence, benchmark disclosure, and an embedded captioned Codex
  repair demo.
• A genuine Codex repair record with a sanitized transcript, exact fixture
  commits and commands, red-before-green endpoint proof, and current-tree
  CodeDecay revalidation.
• Parser-grounded JavaScript security matchers that require real command sinks
  and data flow for path traversal and hardcoded-secret evidence.
• Output-aware analysis that prevents a generated JSON, Markdown, or SARIF
  report from becoming evidence in the next run while preserving intentional
  CodeDecay config, memory, and arbitrary user files.
• A parser-backed copied-implementation gate that suppresses type shapes,
  type-only imports, and declarative fixtures while preserving executable
  copied-oracle detection.

Reproducible evaluation

The deterministic fixture corpus reports 23 of 23 planted issues recalled and
a 2.22% false-positive rate at zero model cost. These are finite planted-fixture
results, not production accuracy or a guarantee that a merge is safe.

The final release validation includes the full repository suite, package build
and dry-run, child-repository acceptance, PR-safety evaluation, deterministic
benchmark, fresh tarball consumer, published-package smoke, and Judge Lab
server/browser checks.

Safety and privacy

• No required CodeDecay API key.
• No hidden telemetry.
• No hidden model call or private-code upload.
• Commands run only when explicitly configured and allowed.
• Deterministic/tool evidence remains distinct from agent suggestions.
• The public Judge Lab accepts fixed scenario IDs only; it does not accept a
  repository URL, prompt, upload, model credential, or arbitrary command.

Known limitations

• The controlled paired real-agent outcome benchmark in
  #683 is not complete.
• Independent three-participant published-workflow UAT in
  #692 is not complete.
• The public benchmark is planted and finite, and deterministic findings can
  produce false positives.
• Framework/language coverage is incomplete; runtime confidence depends on the
  repository's configured checks and available tools.
• Copied-implementation detection remains a conservative three-line heuristic;
  unsupported syntax, shorter copies, or structurally rewritten logic may not
  be detected.
• The demo fixture deliberately does not model token expiry, malformed JWTs,
  or authorization scope, and its repaired report remains High without runtime
  coverage evidence.

CodeDecay helps find what a coding agent missed before merge. It does not
guarantee perfect safety.

CodeDecay v0.3.5

CodeDecay v0.3.5

This patch release packages the recent PR safety improvements while keeping the public release line on 0.3.x.

Highlights

• Persistent JS/TS symbol impact graph for richer changed-code reachability.
• Changed path test proof map to show whether production paths are actually exercised.
• Base/head differential runtime probe evidence.
• Architecture boundary and ownership checks.
• OpenAPI contract breaking-change detection.
• Reviewable memory learning proposals from PR, CI, and incident signals.
• Agent preflight guidance before code generation.
• Incremental analyzer cache for large-repo scan viability.

Safety boundaries

• No hidden model calls.
• No telemetry.
• No CodeDecayCloud dependency.
• Command execution remains explicit and safety-gated.

Install

npm install -D @submuxhq/codedecay@0.3.5
npx codedecay --help

Validation

• CI passed on PR #653.
• @submuxhq/codedecay@0.3.5 published to npm with latest dist-tag.
• Published-package smoke passed for @submuxhq/codedecay@0.3.5.

CodeDecay v0.3.4

CodeDecay v0.3.4

This release ships the new codedecay loop command: a local-first closed-loop controller that can drive a user-owned coding agent through fix, re-verify, and repeat rounds before merge.

Highlights

• Adds codedecay loop for autonomous plan/fix/recheck workflows.
• Runs in plan-only mode when no agent command is configured.
• Supports explicit user-owned agent commands via --agent-cmd, with prompts passed on stdin.
• Re-runs deterministic CodeDecay analysis and configured checks after each agent action.
• Stops honestly with merge-safe, unverified, plan-only, stuck, needs-human, or agent-error.
• Never auto-commits or auto-pushes; changes stay in the working tree for review.

Safety boundaries

• No hidden LLM/model calls.
• No telemetry.
• No CodeDecayCloud dependency.
• No required API keys.
• Agent output is treated as untrusted until deterministic checks verify it.
• Command execution still goes through CodeDecay's configured execution safety controls.

Install

npm install -D @submuxhq/codedecay@0.3.4
npx codedecay loop --help

CodeDecay v0.3.3

CodeDecay v0.3.3

This release sharpens CodeDecay as a local-first PR red-team orchestrator for AI-assisted development.

Highlights

• Adds the reproducible codedecay benchmark proof command with deterministic recall, false-positive, cost, LLM, and telemetry metrics.
• Adds sticky GitHub Action PR comments so CodeDecay can update one pull request report instead of spamming new comments.
• Adds benchmark-backed launch/README positioning generated from real benchmark output.
• Adds optional memory provider configuration plus safe Mem0 and Supermemory provider adapters.
• Adds codedecay memory setup with dry-run defaults and explicit --apply behavior.
• Wires configured memory provider context into redteam/agent workflows while keeping external context untrusted.
• Adds JWT auth edge-case knowledge pack groundwork and OSS tool recommendations in doctor output.

Safety guarantees

• Local-first by default.
• No telemetry.
• No required API keys.
• No hidden LLM/model calls.
• Optional external memory providers only run when explicitly configured.

Package

• npm: @submuxhq/codedecay@0.3.3
• CLI: codedecay

CodeDecay v0.3.2

CodeDecay v0.3.2 ships the next local-first PR safety harness update.\n\nHighlights:\n- Adds a deterministic security matcher registry.\n- Adds a language/parser support boundary.\n- Adds optional grounded redteam investigation.\n- Adds a resumable local audit data model.\n- Adds deterministic revalidation and memory preview.\n- Adds an explicit plugin registry.\n- Adds a unified harness planted-issue benchmark corpus.\n\nCodeDecay remains local-first and OSS-first: no telemetry, no required API keys, no required LLM/model calls, and no CodeDecayCloud dependency.

CodeDecay v0.3.1

CodeDecay v0.3.1 is a release of the local-first PR safety harness and CLI.

Highlights:

• Adds Semgrep static-analysis tool adapter evidence.
• Adds coverage artifact/tool adapter support.
• Adds Agent Process harness support for user-owned local agents such as Codex, Claude Code, OpenCode, Pi, desktop agents, or custom CLIs.
• Keeps local execution gated behind explicit config and safety.allowCommands.
• Records agent output as untrusted agent-suggestion evidence.
• Fixes CodeQL-reported sanitization and regex security alerts.
• Expands OSS-first agent guidance and requires squash merges/modular package structure for future PRs.

Validation:

• pnpm install
• pnpm run lint
• pnpm typecheck
• pnpm test (341 tests)
• pnpm build
• pnpm --filter @submuxhq/codedecay pack --dry-run
• Installed npm package smoke test
• Code scanning open alerts: 0

Package:

• npm: @submuxhq/codedecay@0.3.1
• CLI: codedecay

CodeDecay remains local-first, deterministic by default, no telemetry, no required API keys, and no hidden model calls.

CodeDecay v0.3.0

CodeDecay v0.3.0

This release focuses on making CodeDecay safer to adopt in real PR workflows: lower-noise scoring, clearer release gating, stronger CLI utilities, memory learning, and package smoke coverage.

Highlights:

• Recalibrated low-signal changes so docs/assets/metadata-only updates do not create false high-risk reports.
• Added a focused runtime config plus database/schema risk lift for production-sensitive changes.
• Added memory learning flows for CI, PR, and CodeDecay report signals.
• Added/verified CLI utility surfaces for help, man, update, uninstall, version, and unknown-command suggestions.
• Added StrykerJS mutation report ingestion support and stronger published-package smoke coverage.

Install:

npm install -D @submuxhq/codedecay@0.3.0

Verification completed:

• GitHub CI, docs deploy, and CodeQL passed on the release commit.
• Tarball smoke passed before publish.
• npm registry package smoke passed after publish.

CodeDecay v0.2.0

CodeDecay v0.2.0

This release migrates the public npm package to the GitHub organization-aligned scope:

npm install -D @submuxhq/codedecay

The installed CLI binary remains unchanged:

npx codedecay --help

What Changed

• Renamed the primary npm package from @submux/codedecay to @submuxhq/codedecay.
• Kept the CLI binary name as codedecay.
• Updated README, docs, release docs, GitHub Action metadata, CLI help text, tests, and workflow/package references.
• Kept npmjs as the default public install path.
• Kept GitHub Packages as an optional mirror using the same package name.

Validation

• pnpm install
• pnpm run lint
• pnpm typecheck
• pnpm test - 239 passing
• pnpm build
• pnpm --filter @submuxhq/codedecay pack --dry-run
• Fresh install from npm: npm install @submuxhq/codedecay@0.2.0
• Binary check: node_modules/.bin/codedecay version -> 0.2.0

CodeDecay remains local-first, deterministic by default, and has no telemetry, no required API keys, and no required LLM/model calls.

CodeDecay v0.1.6

CodeDecay v0.1.6

Patch release for release hygiene, security scanning cleanup, and CLI polish.

Changes

• Fixes CodeQL polynomial regex alerts by replacing risky regex parsing with linear scanners.
• Resolves dependency alerts for Vite and esbuild through workspace overrides.
• Adds stronger end-user demo coverage for MCP/client and GitHub Action runtime smoke paths.
• Improves CLI discovery by suggesting similar commands and flags.
• Keeps CodeDecay local-first, deterministic, no telemetry, no API keys, and no required LLM/model calls.

Package

• npm: npm install -D @submux/codedecay
• version: @submux/codedecay@0.1.6
• binary: codedecay

Validation

• pnpm install
• pnpm run lint
• pnpm typecheck
• pnpm test
• pnpm build
• pnpm --filter @submux/codedecay pack --dry-run
• fresh npm install and codedecay --help