Context Budget Manager

Manage the token budget for Claude Code sessions. Now model-aware — picks the
right effective context window per model. Opus 4.7/4.8 and Sonnet 4.6 are all 1M;
Haiku 4.5 is 200K.

Parse $ARGUMENTS:

status (or no arguments)

Show current budget config + auto-compact settings:

cat ~/.claude-context-optimizer/config.json 2>/dev/null
echo "---"
cat ~/.claude-context-optimizer/budget-config.json 2>/dev/null

If no config exists, show defaults (200K working budget on Opus 4.8's 1M window, warn at 50/70/85/95%).

set <tokens>

Update the budget limit. Parse the token count (200K, 1M, 500000 all OK).
Update ~/.claude-context-optimizer/config.json:

{
  "budgetTokens": <parsed_number>,
  "warnAt": [50, 70, 85, 95],
  "autoCompactAt": 90,
  "model": "opus-4.8"
}

If budgetTokens exceeds the chosen model's context window, warn the user.

model <name>

Set the model for cost estimation. Supported keys:

• haiku-4.5 (alias haiku) — $1/$5 per M, 200K
• sonnet-4.6 (alias sonnet) — $3/$15 per M, 1M
• opus-4.7 — $5/$25 per M, 1M
• opus-4.8 (alias opus, default) — $5/$25 per M, 1M

opus-4.8-1m / opus-4.7-1m / opus-extended are back-compat aliases only — there is
no 1M surcharge; the 1M window is standard at $5/$25.

Update the model field in config.json. When switching to a 1M-context model and the
current budgetTokens is below 500K, ask if the user wants to bump it to 1M.

auto <on|off>

Toggle auto-compact at thresholds (80% / 90%). Update
~/.claude-context-optimizer/budget-config.json:

• auto on → autoCompactEnabled: true
• auto off → autoCompactEnabled: false

Defaults if file missing:

{
  "autoCompactEnabled": true,
  "autoCompactThreshold": 80,
  "criticalThreshold": 90
}

Cost calculation

The budget monitor now estimates input + output tokens separately and uses
the model's real input/output prices. Example: Edit with a 200-char new_string
counts as ~54 output tokens, charged at the model's output rate.

Effective Budget Multiplier

At 50%+ budget usage, the monitor shows how much CCO multiplies your effective
budget — e.g. "1.6x more effective" if Read Cache + file digests saved enough
redundant reads to make your 200K context behave like ~320K.