FORGE-route-least-power

ROUTE LEAST POWER — Lower Machine Entropy

"Can a simpler tool do this?" — Ask this before every tool call. Power is not preference. Power is fitness. Lower entropy = less chaos = fewer incidents.

The Ladder (escalate only if current tier fails)

TIER 1 — READ (zero side effects)
  read, glob, grep, webfetch
  ↓ insufficient? ↓

TIER 2 — OBSERVE (low cost, no mutation)
  arif_observe, forge_registry_status, forge_probe
  ↓ insufficient? ↓

TIER 3 — REASON (advisory only)
  arif_think, sequential-thinking, forge_evaluate
  ↓ insufficient? ↓

TIER 4 — MUTATE (governed, reversible)
  forge_shell, forge_git, forge_filesystem, edit, write
  ↓ insufficient? ↓

TIER 5 — EXTERNAL (network, API calls)
  forge_search, forge_fetch, brave-search, perplexity
  ↓ insufficient? ↓

TIER 6 — IRREVERSIBLE (requires F13)
  arif_seal, forge_execute_sealed, rm -rf, DROP TABLE

Rules

1. Always start at Tier 1. Only escalate when the current tier cannot solve the problem.
2. State your tier before acting. "I'm using Tier 2 (observe) to check organ health."
3. Never skip tiers. Don't jump from read → irreversible.
4. If two tools at the same tier work, prefer the one with fewer side effects.
5. If you're reaching for Tier 5+ for a simple task, you're probably over-engineering. Stop.

Entropy-Reducing Discipline

One Owner Per Task

• Never let 2 agents edit the same file or service concurrently
• If work overlaps, serialize: agent A finishes fully (including verify) before agent B starts
• When in doubt, use forge_lock to claim ownership

Idempotent Actions

• Every agent command should be safe to rerun
• Before running: "What happens if this runs twice?"
• If the answer is "it breaks," refactor for idempotency first
• Prefer forge_filesystem_patch (preview before apply) over blind write

Strict Queues

• Serialize risky work — don't parallelize mutations on the same substrate
• Each phase completes fully (read → plan → execute → verify) before the next begins
• For parallel safe work (independent OBSERVE calls), use forge_parallel

3-Agent Routing

ARCHITECT decides → Tier 3 (reason)
ENGINEER applies  → Tier 4 (mutate)
AUDITOR verifies  → Tier 1-2 (read/observe)

• If you're doing all 3 roles, separate the VERIFY step explicitly
• Never verify with the same tool that mutated

Anti-Patterns

• ❌ Calling forge_shell to read a file (use read)
• ❌ Calling perplexity_research for a local file search (use grep)
• ❌ Calling arif_seal before arif_judge (judge first, seal after)
• ❌ Reaching for task() to spawn a subagent for a single grep
• ❌ Two agents editing the same config file simultaneously
• ❌ A non-idempotent script in a cron job

Decision Heuristic

Is this a question?          → Tier 1-2 (read/observe)
Is this a plan?              → Tier 3 (reason)
Is this an edit?             → Tier 4 (mutate, with F1 AMANAH backup)
Is this external research?   → Tier 5 (fetch/search)
Is this permanent?           → Tier 6 (irreversible, F13 required)
Is there an owner already?   → Wait. Don't race.
Can this run twice safely?   → If no, fix it first.

Seal

This skill prevents tool escalation spirals. It's the agent's cost function. Every unnecessary Tier 5+ call wastes compute and increases blast radius. One owner, idempotent actions, strict queues — lower entropy every cycle.

DITEMPA BUKAN DIBERI.