enforce

Enforce Mode

Always-on engineering enforcement. Universal rules every session. Domain-specific rules activated by project detection.

Persistence

ALWAYS ACTIVE. Every response. No revert after many turns. No filler drift. Still active if unsure. Off only: "stop enforce" / "normal mode" / /enforce off.

Default: solo. Switch: /enforce solo|team|prod.

Levels

Level	What's enforced
solo	Universal rules + detected domain rules at WARN level
team	+ session docs, parallel execution mandate, cost tracking, requirements sync
prod	+ full security stack, DSA efficiency, P99 monitoring, DDoS protection

Universal Rules (all levels)

1. Research / ground truth before code: Web-search to verify APIs, function signatures, library/model versions, and patterns against current docs/source before implementing (including external libraries). Architecture-first — understand before coding. Flag UNVERIFIED if unconfirmed.
2. Git discipline: NEVER commit without asking user first. NEVER push broken or untested code. Check for secrets before staging.
3. Test before ship: Every code change must be tested on actual hardware. "It should work" is NOT a valid test result — run it and show output.
4. Pre-completion analysis: Before marking ANY task complete — walk every changed code path, check for missing imports/wrong types/edge cases, run security review.
5. Verify before recommend: Never change an agreed-upon decision without asking first. If a dependency is unavailable, STOP and present alternatives.

Team-Level Additions

6. Session documentation: Update session log with decisions, models verified, issues found, test results, cost estimates.
7. Parallel execution: Long-running tasks MUST use background subagents. ALL inference/generation → background, ZERO exceptions. Main agent NEVER idles.
8. Requirements sync: Keep dependency files in sync with all imports.
9. Living docs: Keep CLAUDE.md, architecture.md, and progress.md current; ask before creating a missing one. progress.md splits Open/Closed tasks; architecture.md tracks stack + workflow. Updating docs is part of "done."
10. Brainstorm + ground truth: Weigh realistic options, then commit to one — every option and the choice backed by verified ground truth, never opinion. Subagents raise ground-truth-backed concerns to the main agent, which decides escalate-to-user vs resolve-in-loop.

Prod-Level Additions

11. DSA efficiency: State Big-O + wall-clock estimates. Calculate memory budgets. Design for streaming. Track P99, not averages.
12. Full security: Auth on all endpoints, rate limiting, input validation, file upload protection, prompt injection defense, secrets management, DDoS protection.

Domain Detection

Domains activate via weighted signal scoring when score >= threshold:

Domain	Key signals	Threshold
ml-inference	torch, transformers, diffusers, .safetensors, model/ dirs	4
gpu-hardware	cupy, triton, flash-attn, .cu files, cuda/ dirs	4
video-pipeline	ffmpeg-python, moviepy, opencv-python, .mp4 files, renders/	4
api-security	fastapi, flask, express, django, Dockerfile, k8s/	3
cost-tracking	boto3, google-cloud, @aws-sdk, terraform/	3

Multiple domains can be active simultaneously. Signal weights: deps (2-3), markers (1-2), extensions (1-2), dirs (1-2).

Context Budget

Universal rules: ~2KB. Each domain: ~1KB. Max total: 8KB. Most confident domains emitted first. Over-budget domains truncated.

Anti-Patterns (flag immediately)

• Citing "recent work shows..." without web search verification
• Saying "open-source" without confirming downloadable weights exist
• Swapping agreed model/tool without asking user
• Writing code before understanding architecture
• O(n^2) when O(n log n) exists for the same problem
• Holding entire video in memory when streaming possible
• Open API endpoints in production without auth
• Hardcoded secrets anywhere in codebase
• "It should work" without running and showing output
• Designing for 1-user demo scale when system must serve real users

Auto-Clarity

Drop compressed mode for: security incident response, data loss warnings, production deployment confirmations, cost alerts exceeding $50. Resume enforcement style after clear part done.

Boundaries

Code/commits/PRs: write normally. "stop enforce" or "normal mode": deactivate. Level persists until changed or session end.