Skip to main content
Resource management keeps Super Turtle autonomous when usage limits get tight.
The Meta Agent reads Claude Code and Codex quota signals, routes work to the best loop type, and adjusts supervision frequency to preserve headroom. Canonical policy lives in: 
super_turtle/meta/META_SHARED.md
Runtime usage surfaces live in: 
super_turtle/claude-telegram-bot/src/handlers/commands.ts
super_turtle/subturtle/ctl

​
Signals and Data Sources

The Meta Agent uses two quota inputs:

Claude Code usage

getUsageLines() reads OAuth usage from https://api.anthropic.com/api/oauth/usage and formats session + weekly utilization with reset times.

Codex quota usage

getCodexQuotaLines() calls codex app-server via JSON-RPC (account/rateLimits/read) and returns 5-hour + weekly usage windows.
Unified status rendering is handled by formatUnifiedUsage(...) and uses shared badge thresholds: 
✅ < 80%
⚠️ 80-94%
🔴 >= 95%

​
Usage Check Cadence

META_SHARED.md defines when usage should be evaluated:
1

Session start

Check quota state at the beginning of every meta-agent session.
2

During active work

Re-check about every 30 minutes while autonomous work is in progress.
3

Before spawning

If the last reading is stale, refresh usage immediately before spawning new SubTurtles.

​
Decision Matrix

Use this policy matrix for loop-type routing and user messaging:
Claude Code usageCodex usageMeta behavior
<50%<50%Normal operations. Any loop type allowed.
50-80%<50%Prefer yolo-codex. Reduce cron supervision to 10m.
>80%<50%Force yolo-codex only. Minimal check-ins (15m). Keep responses shorter.
Any>80%Switch SubTurtles to yolo (Claude) and warn Codex is constrained.
>80%>80%Alert that both pools are constrained and suggest pausing non-critical work.
Default coding loop type is yolo-codex. Use yolo or slow only when deeper Claude reasoning/review is actually needed.

​
Cron Frequency Adaptation

When Claude usage pressure increases, stretch supervision intervals to reduce meta overhead. Policy rule: 
Claude > 80% => 15m recurring check-ins
Claude <= 80% => return to tighter intervals (typically 10m per matrix)
Operational command: 
./super_turtle/subturtle/ctl reschedule-cron <name> <interval>
Examples: 
# High Claude usage: reduce check-in frequency
./super_turtle/subturtle/ctl reschedule-cron dashboard-api 15m

# Usage recovered: tighten supervision again
./super_turtle/subturtle/ctl reschedule-cron dashboard-api 10m

​
Practical Routing Flow

# 1) Gather usage/quota state
# (meta path uses usage + codex quota helpers)

# 2) Choose loop type
# default: yolo-codex
# codex constrained: yolo
# claude constrained + codex healthy: keep yolo-codex

# 3) Spawn with selected mode
./super_turtle/subturtle/ctl spawn <name> --type <selected> --timeout 1h --state-file -

# 4) Adjust cron interval based on current Claude pressure
./super_turtle/subturtle/ctl reschedule-cron <name> 10m
# or
./super_turtle/subturtle/ctl reschedule-cron <name> 15m

​
Failure and Partial-Data Handling

If one provider’s metrics are missing, formatUnifiedUsage(...) reports partial state (❓) instead of guessing.
Routing should then prefer conservative defaults (yolo-codex) until a fresh measurement is available, unless there is a clear provider-specific limit error.