Configuration

Config is JSON at ~/.config/tokenmaxxing/config.json. Every field is optional; defaults apply per-field.

{
  "thresholds": { "session": 95, "weekly": 98 },
  "hardThresholds": { "session": 100, "weekly": 100 },
  "claudeBin": "/path/to/real/claude",
  "codexBin": "/path/to/real/codex",
  "policy": {
    "projectionMargin": 0,
    "greedySessionFloor": 50,
    "switchModels": ["fable"],
    "usagePollTtlMs": 90000,
    "maxWaitMs": 3600000
  }
}

• thresholds.session / thresholds.weekly: the screening bars described in How switching decides.
• hardThresholds.session / hardThresholds.weekly: the wall bars (default 100, each account's real server limit). When every account is over its screening bar, the automatic Claude decision squeezes each one up to its wall before parking - see The wall. Set equal to thresholds to disable; Codex ignores this setting.
• claudeBin / codexBin: pins to the REAL binaries the supervisors spawn (written by init / init --codex; the recursion-guard diagnostic tells you to fix claudeBin here when it stops pointing at the real Claude binary).
• policy.projectionMargin: subtracted from both the screening bars and the wall bars to form the effective bars.
• policy.greedySessionFloor: session-used percent at which the greedy convergence engages.
• policy.switchModels: model families whose per-model weekly cap gates a switch (entries are lowercased on load).
• policy.usagePollTtlMs: how long a /usage per-model poll stays fresh.
• policy.maxWaitMs: the depleted-pool auto-wait bound.

tokenmaxxing config shows the effective config with sources; config set/unset edit dotted keys; config tidy prunes unknown keys.