model-config

name	model-config
version	1.0.0
description	This skill should be used when the user asks to "change model", "set default model", "switch model", "use 1M context", "enable 1M context window", "change effort level", "set thinking depth", "configure effort", "adjust auto compact", "change compact percentage", "set auto compact threshold", "configure auto compaction", "configure model settings", or wants to modify the model, effortLevel, or CLAUDE_AUTOCOMPACT_PCT_OVERRIDE settings in Claude Code.

Configure the default model, thinking effort level, and auto-compact threshold in ~/.claude/settings.json.

Background

Three settings in ~/.claude/settings.json control core Claude Code behavior:

• model (top-level key) — the default model ID used for all sessions
• effortLevel (top-level key) — thinking depth for models that support extended thinking
• CLAUDE_AUTOCOMPACT_PCT_OVERRIDE (inside the env block, as a string) — the percentage of context capacity at which Claude Code automatically compacts conversation history

Supported models and their valid effort levels:

Model	Valid effort levels	Default effort	1M context
claude-opus-4-7	low, medium, high, xhigh, max	xhigh	yes
claude-opus-4-6	low, medium, high, max	high	yes
claude-sonnet-4-6	low, medium, high, max	high	yes
claude-haiku-4-5-20251001	not supported	—	no

1M Context Window

Opus 4.7, Opus 4.6, and Sonnet 4.6 support a 1 million token context window. To enable it, append [1m] to either a model alias or a full model name:

• Aliases: opus[1m], sonnet[1m]
• Full names: claude-opus-4-7[1m], claude-opus-4-6[1m], claude-sonnet-4-6[1m]

Claude Code strips the [1m] suffix before sending the model ID to the API — it only controls the local context window behavior. On Max, Team, and Enterprise plans, Opus is automatically upgraded to 1M context even without the suffix, but using it makes the setting explicit.

For auto-compact: the default behavior (no setting) triggers compaction at ~95% context capacity. The valid range is 1–83; values above 83 are silently capped internally.

Step 1: Fetch Current Model Information

Use WebFetch to read https://code.claude.com/docs/en/model-config and extract the current list of supported models, their effort levels, and 1M context support. Parse the page content to identify:

• Available model IDs
• Which models support extended thinking and their valid effort levels
• Which models support the [1m] suffix for 1M context window

If WebFetch fails or the page format is unrecognizable, fall back to the hardcoded table in the Background section above. Inform the user: "Could not fetch latest model info; using built-in defaults."

Step 2: Read Current Settings

Read ~/.claude/settings.json to check the current configuration.

Step 3: Display Current Values

Extract and display all three current values:

• Model: the value of the top-level model key, or "not set (Claude Code default)" if absent
• Effort level: the value of the top-level effortLevel key, or "not set (model default)" if absent
• Auto-compact threshold: the value of env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE, or "not set (default: ~95%)" if absent

Step 4: Ask Which Settings to Change

Use AskUserQuestion to ask:

"Which setting(s) would you like to change? (model / effort / compact — or name multiple)"

Parse the response and collect a list of the selected settings. Proceed through the applicable steps below in order, skipping any not selected.

Step 5: Update Model (if selected)

Step 5a: Select Base Model

Use AskUserQuestion to present the following choices:

1. claude-opus-4-7
2. claude-opus-4-6
3. claude-sonnet-4-6
4. claude-haiku-4-5-20251001

Include in the question text: "Or type a custom model ID directly."

Parse the user's response:

• If 1–4, use the corresponding model ID.
• If the user enters a model name matching one of options 1–4, use the corresponding model ID.
• If the user enters a different model ID string, use that value exactly as a custom model ID.
• If the input is empty or unrecognized, inform the user and ask again.

Step 5b: Enable 1M Context Window

If the selected base model supports 1M context (check the model table in Background, or the data fetched in Step 1), use AskUserQuestion to ask:

"Enable 1M context window? (yes / no)"

• If yes, append [1m] to the model ID (e.g., claude-opus-4-6 → claude-opus-4-6[1m]).
• If no, use the base model ID as-is.
• Inform the user: "On Max, Team, and Enterprise plans, Opus models auto-upgrade to 1M context even without the suffix."

If the selected model does not support 1M context (e.g., Haiku), skip this sub-step.

For custom model IDs, inform the user they can manually append [1m] to any supported model for 1M context.

Step 5c: Validate Selection

• If the final model ID (with or without [1m]) equals the current setting, note "Already set to that model. No change needed." and skip the write for this field.

Record the chosen model ID as the target model for Step 6. When determining the base model for effort level validation in Step 6, strip the [1m] suffix (e.g., claude-opus-4-6[1m] → validate against claude-opus-4-6).

Step 6: Update Effort Level (if selected)

Determine which model to validate against:

• If the user changed the model in Step 5, use the newly selected model as the reference. Strip the [1m] suffix before validation (e.g., claude-opus-4-6[1m] → validate against claude-opus-4-6).
• Otherwise, use the current model value from settings.json (also stripping any [1m] suffix). If no model key exists, treat as unknown.

If the reference model is claude-haiku-4-5-20251001, inform the user: "Haiku 4.5 does not support effortLevel. Skipping effort configuration." If an effortLevel key currently exists in settings.json, ask the user if they want to remove it. Proceed to Step 7.

Use AskUserQuestion to present the valid effort levels for the reference model:

• For claude-opus-4-7: low, medium, high, xhigh, max (default: xhigh)
• For claude-opus-4-6 or claude-sonnet-4-6: low, medium, high, max (default: high)
• For unknown or custom model: low, medium, high, xhigh, max — note that validity depends on the model

Parse the response:

• If the value is not in the valid list for the reference model, inform the user of the valid options and ask again.
• If the value equals the current setting, note "Already set to that level. No change needed." and skip the write for this field.
• Normalize the value to lowercase before writing.

Known bug with max: If the user selects max, warn them:

"Note: max effort has a known bug where it does not persist correctly in settings.json. Preferred workaround: add "CLAUDE_CODE_EFFORT_LEVEL": "max" to the env block in settings.json. Alternatively, add export CLAUDE_CODE_EFFORT_LEVEL=max to your shell profile. Would you like to set it via the env block, or skip?"

If the user chooses the env block workaround, write CLAUDE_CODE_EFFORT_LEVEL with value "max" to the env block in Step 8 instead of writing the effortLevel key. If an effortLevel key currently exists, remove it to avoid conflicts. If the user chooses to skip, do not write the effortLevel key.

Step 7: Update Auto-Compact Threshold (if selected)

Use AskUserQuestion to ask:

"Enter desired auto-compact percentage (1–83). Recommended: 60–75 for most work. Values above 83 are capped internally."

Parse the user's response as an integer:

• If not a valid integer, inform the user and ask again.
• If the value is less than 1, inform the user the minimum is 1 and ask again.
• If the value is greater than 83, inform the user: "The maximum effective value is 83 (Claude Code caps anything higher internally). Setting to 83." Use 83.
• If the value equals the current setting, note "Already set to N%. No change needed." and skip the write for this field.

Step 8: Write All Changes

Read ~/.claude/settings.json again (to get the latest state), then apply all collected changes in a single surgical update:

• If ~/.claude/settings.json does not exist, use the Write tool to create it with only the keys being set.
• If the file exists, use the Edit tool for surgical updates.
• For the model key: add or replace the top-level model key. Preserve all other top-level keys.
• For the effortLevel key: add or replace the top-level effortLevel key. Preserve all other top-level keys. If the user chose the max env block workaround in Step 6, remove the effortLevel key instead.
• For CLAUDE_CODE_EFFORT_LEVEL (max workaround): if the user chose the env block workaround in Step 6, add "CLAUDE_CODE_EFFORT_LEVEL": "max" to the env block.
• For CLAUDE_AUTOCOMPACT_PCT_OVERRIDE: if no env block exists, add one. Add or replace only the CLAUDE_AUTOCOMPACT_PCT_OVERRIDE key within env. Preserve all other keys in both the env block and the rest of the file.

All values are JSON strings: model, effortLevel, and env var values must be quoted.

Example partial result after setting all three:

{
  "model": "claude-opus-4-7[1m]",
  "effortLevel": "xhigh",
  "env": {
    "EXISTING_KEY": "preserved",
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "50"
  }
}

Only write keys that the user selected and that differ from the current value.

Step 9: Confirm

For each setting that was changed, report the new value. Then inform the user: "Restart Claude Code for changes to take effect."

If nothing was changed, confirm: "No changes were made."

Uninstall

To revert to defaults, remove the relevant keys from ~/.claude/settings.json:

• Remove the top-level model key to restore Claude Code's default model selection.
• Remove the top-level effortLevel key to restore the model's default thinking depth.
• Remove the CLAUDE_AUTOCOMPACT_PCT_OVERRIDE key from the env block to restore the ~95% default compaction threshold.
• Remove the CLAUDE_CODE_EFFORT_LEVEL key from the env block if present (set via the max effort workaround).
• If the env block becomes empty after removal, remove the env block as well.

Preserve all other keys in the file.