kodemux v0.4.1

Code-review hardening pass.

• Fixed a router crash on partial tier config overrides (model without efforts)
• AI judge no longer depends on structured-output support — portable across SDK/model versions
• post --run guards against shell-unsafe filenames (closes an injection surface)
• Explainer surfaces the parallel-agent recommendation prominently

84 tests green (5 new). See CHANGELOG.md.

🤖 Generated with Claude Code

kodemux v0.4.0

When the deterministic router isn't confident, kodemux now consults a cheap AI judge instead of only guessing.

• On by default, fires only when confidence is low — the confident majority of prompts pay zero extra cost/latency
• Cheapest model (claude-haiku-4-5) for the judgment
• Reuses existing Anthropic credentials — no separate API key setup
• Never overrides, only raises the deterministic complexity/risk — the core decision stays explainable
• Always fails safe — no credentials/network/timeout/refusal all fall back silently, verified live
• --no-ai / router.aiAssist: false to disable

79 tests green (16 new). See CHANGELOG.md and docs/INTERNALS.md §5.8.

🤖 Generated with Claude Code

codemux v0.3.0

The router now reads the actual diff, not just the prompt string — still deterministic, still no LLM call.

• Automatic git context: codemux route runs git diff to derive real changed-file and diff-line counts (--base, --no-git, and --files/--diff-lines overrides)
• Critical-path detection: a change touching auth/migrations/infra/.env/secrets/payments/.tf raises a critical risk flag and floors the tier — even if the prompt never says so
• Dependency-free glob matcher; router.criticalPaths config
• Explainer gains a 'touches a critical path' toggle, parity-checked against the CLI

63 tests green. See CHANGELOG.md and docs/INTERNALS.md.

🤖 Generated with Claude Code

codemux v0.2.0

Router overhaul: routes by estimated task complexity on a capability ladder (Haiku 4.5 → Sonnet 5 → Opus 4.8 → Fable 5) instead of a fixed keyword→model table.

• Multi-signal complexity score (0–14) + risk flags (security/production)
• Confidence-gated escalation cascade when unsure
• Parallel-agent recommendation (/agents N) for multi-agent work
• Config v2: overridable tiers, thresholds, risk floor
• Interactive explainer rewritten + parity-checked against the compiled router

51 tests green. See CHANGELOG.md.

🤖 Generated with Claude Code

codemux v0.1.0

First release of codemux — repo-native middleware & guardrail engine for AI coding tools.

Highlights

• codemux init — stack detection + .codemux/ scaffold
• codemux route — intent classifier → model/effort/mode directives
• codemux guard / codemux scan — branch-protection & secrets pre-hooks
• codemux post — scoped format/lint/test post-hook

43 unit tests, CI on Node 20 & 22. See CHANGELOG.md.

🤖 Generated with Claude Code