Skip to content

rules.md

Latest commit

 

History

History
181 lines (143 loc) · 7.01 KB

rules.md

File metadata and controls

181 lines (143 loc) · 7.01 KB
Raw

Rules

All constraints, operational rules, convergence logic, and pattern selection.

Operational Rules

These override all defaults. Violating any is a mistake — log it to .claude/mistakes.md.

1. Memory Before Action

BEFORE any task:
  1. Search memory for matching patterns/solutions
  2. Score relevance (0.0 - 1.0)
  3. If score > 0.7 → reuse the pattern
  4. If score 0.4-0.7 → use as starting point
  5. If score < 0.4 → build from scratch

2. Orchestrator, Not Executor

Claude coordinates. Agents do the work. If you're writing/researching/reviewing directly, stop and dispatch.

3. Enhanced Swarm for Production

Task analyzer → dynamic routing → GRPO sampling → persona evolution → RLM synthesis.

Supervisor (1)
  ├── Researcher (1-2)
  ├── Writer (1)
  ├── Reviewer (1)
  ├── Fact Checker (0-1)
  ├── SEO Optimizer (0-1)
  ├── Code Expert (0-1)
  └── Audience Adapter (0-1)

Min 3, max 8 agents. Other patterns are for experimentation only.

4. Learn After Success

When review_score >= 8.0: extract pattern (agents, routing, strengths) → store in patterns namespace.

5. 3-Tier Routing

Tier 1: CACHED       → identical/similar task solved before
  ↓ (miss)
Tier 2: LIGHTWEIGHT  → check [AGENT_BOOSTER_AVAILABLE]
  ↓ (miss)
Tier 3: FULL AGENT   → complete LLM call

6. Commands Return Instantly

Commands create records only. Never wait. Supervisor monitors completion.

Convergence Rules

Pattern Pass Condition Max Iterations Fallback
Hierarchical review_score >= 8.0 AND accuracy >= 9.5 3 Accept best draft
Peer Debate Score improvement < threshold AND accuracy >= 9.5 patience counter Accept current draft
Dynamic Swarm Task queue empty AND accuracy >= 9.5 3 re-analysis rounds Accept current state
Enhanced Swarm Adaptive threshold AND accuracy >= 9.5 Experience-aware Rollback to best

Constraints

  • Never instantiate ChatOpenAI directly — always use get_llm() from shared/agents.py
  • Never return full AgentState from an agent — only the fields that changed
  • Never mutate topic after initialization — it's the immutable input
  • shared/ modules must not import from patterns/ — dependency flows one way
  • Agents must be stateless functions — no instance variables, no side effects
  • Review scores are floats 0-10. Passing threshold is 8.0. Max iterations is 3
  • Output files go to outputs/ as markdown

Pattern Selection

Use Hierarchical When

  • Task workflow is known upfront
  • You need a single point of control and auditability
  • Speed matters more than output quality

Use Peer Debate When

  • Output quality is the top priority
  • Task is subjective (writing, design, strategy)
  • You can afford higher token costs and latency

Use Dynamic Swarm When

  • Task complexity is unknown upfront
  • Requirements may emerge during execution

Use Enhanced Swarm When

  • Production workload with high quality requirements
  • You want agents that improve over time

Performance

Metric Hierarchical Peer Debate Dynamic Swarm Enhanced Swarm
LLM calls/run 4-10 10-25 6-15 15-40
Token usage Low High Medium Very High
Output quality Good Excellent Good Excellent

Intent Routing

All intents recognized by command_router.py:

Intent Example Triggers
CREATE_TASKS Multi-line list, add task, !! urgent, ! high
SHOW_TASKS "show tasks", "my todo", "what do I have today"
COMPLETE_TASK "done: X", "mark X done", "complete: X"
REMOVE_TASK "remove: X", "delete: X"
WEEKLY_PLAN "plan", "weekly plan", "carry forward"
CALENDAR "calendar", "what's today", "events"
CREATE_EVENT "remind me at 3pm", "set event", "schedule"
GMAIL "check emails", "inbox", "any recruiter emails"
GITHUB "commits", "what did I push"
TRENDING "trending", "hot repos"
BRIEFING "briefing", "morning update"
ARXIV "arxiv", "ai papers", "latest papers"
LOG_SPEND "spent 15 on lunch", "$20 for groceries"
LOG_INCOME "earned 500 freelance", "got paid 2000"
LOG_SAVINGS "saved 100", "invest 50"
SET_BUDGET "set budget groceries 50", "limit transport to 30"
SHOW_BUDGET "budget", "weekly spending", "summary"
UNDO_BUDGET "undo", "undo last transaction"
RECURRING_BUDGET "recurring: 10 on spotify monthly", "show recurring"
WEEKLY_REPORT "weekly report", "this week summary"
EXPORT "export", "backup"
CONVERSATION Any unmatched text → free-form LLM chat
REMOTE_SHELL "run: ls", "$ git status"
GIT_OPS "git status", "commit: msg", "push"
FILE_OPS "show: file.py", "logs", "errors", "more"
SYSTEM_STATUS "status", "daemon check"
CLEAR_CHAT "clear chat", "new conversation"
SCAN_JOBS "scan jobs", "find jobs", "run autopilot"
SHOW_JOBS "jobs", "show jobs", "pending jobs"
APPROVE_JOBS "apply 1,3,5", "apply all"
REJECT_JOB "reject 2", "skip 3"
JOB_DETAIL "job 3", "details 5"
JOB_STATS "job stats", "application stats"
SEARCH_CONFIG "search: add title X"
PAUSE_JOBS "pause jobs", "stop autopilot"
RESUME_JOBS "resume jobs", "start autopilot"

Job Autopilot Rules

Rate Limits (March 2026 — research-backed)

  • Total daily cap: 25 applications across all platforms
  • LinkedIn: 15/day, session break of 30min every 5 apps (ML detection risk)
  • Indeed: 8/day (aggressive IP banning, permanent account suspension)
  • Workday: 5/day (behavioral analysis + 3rd-party bot detection)
  • All platforms: 20-45s random delay between submissions, 10min break every 5 apps

Anti-Detection

  • All adapters use headed mode (not headless) with --disable-blink-features=AutomationControlled
  • LinkedIn uses persistent browser profile with human-like typing (50-150ms/char)
  • Thread mutex prevents concurrent apply_job() calls (TOCTOU race protection)
  • Pipeline lock prevents concurrent run_scan_window() runs (cron vs Telegram)
  • Application recorded BEFORE submission (prevents silent limit bypass on error)
  • UTC timezone for daily cap tracking (prevents midnight drift)
  • Verification Wall Learning: detect + record + correlate + adapt. 17 signals per session. Statistical engine (free) + LLM (every 5th block). 2hr→4hr→48hr cooldown. Human-like interaction on all Playwright scanners.

Input Modes

Mode Source Processing
Text Telegram, Slack, Discord Rule-based → LLM fallback classification
Voice Telegram voice messages Whisper transcription → text classification
Webhook External HTTP POST Payload extraction → dispatcher

Migration Path

Hierarchical → Dynamic Swarm → Enhanced Swarm
     │
     └──→ Peer Debate (if quality > speed)