crushrc: bash-based config

[meowgorithm]

This revision adds imperative, Bash-based configuration to Crush (don't worry, JSON is still supported for now). This makes configuration a lot more flexible than the current JSON format for a few reasons:

• You can add comments
• You can add logic
• You can literally use Bash

Full docs here.

Why did we choose Bash? Because Crush ships with a glorious first class, fully cross platform pure Go bash interpreter, and LLMs are great at Bash. The larger goal, however, is to allow Crush to imperatively change it's config, in realtime, with the same commands used to configure it.

Thus, the configuration itself is done via a series of builtin commands. Here are some things you can do:

# Add Ollama.
provider add ollama --type ollama --base-url "http://localhost:11434/v1"

# Register a model on Ollama.
model add ollama/llama3.3 --name "Llama 3.3" --context-window 128000

# Auto-approve some tools.
permissions allow view edit

# Add an MCP server
mcp add github \
  --type http \
  --url "https://api.githubcopilot.com/mcp/" \
  --header Authorization "Bearer $GITHUB_TOKEN"

It's also composable:

# Load some extra config
source "$XDG_CONFIG_HOME/squid-config.sh"

And can include logic:

# Change config based on the machine you're on.
if [[ $HOSTNAME == "babysquid" ]]; then
    option skill-path "$HOME/squid-skills"
fi

# Get API keys from your password manager.
provider add my-secret-provider \
  --api-key "$(op read my-secret-key)" \
  --type openai-compat \
  --base-url "https://api.example.com/v1" \

The configuration lives globally at ~/.config/crush/crushrc and locallly at ./.crushrc and crushrc.

As mentioned, JSON will still work as normal with one small change:

Note that both crushrc and crush.json are backed with the Bash interpreter and both are trusted code, and should be treated with the same precautions as any other script.

[joestump-agent]

Overall this is a really clean design. Reusing the mvdan/sh interpreter that already powers the bash tool is the right call — source, $(...), conditionals, and ${VAR:?...} all come for free, and the ConfigBuilder-on-context gating is an elegant way to keep the same builtins no-op during normal agent bash execution. The test coverage is thorough (per-builtin unit tests + end-to-end tests through the real config.Load pipeline, including precedence and source).

Leaving inline comments for the issues below; none are blockers, but a few are worth addressing before merge.

Docs (nits, but user-facing):

• docs/config/README.md: $CURSH_VERSION typo defeats the version-gating example entirely.
• README.md security note: "Both crushrc run in a full shell" grammar.
• The notifications option key is marked "(deprecated)" while notification-style is advertised elsewhere as the replacement, but they write different fields (disable_notifications vs notification_style). Worth a sentence clarifying the relationship.
• No mention of the empty-resolved-header drop behavior for --header/--extra-header, which JSON users rely on for env-gated headers.

Design observations (not bugs):

• PLAN.md is a 909-line internal design doc committed to the repo root. It contains stale code sketches (e.g. WithConfigBuilder(context.Background())) that contradict the shipped implementation, plus forward-looking "JSON retirement" and "real-time config" sections. Strongly suggest dropping it from this PR or relocating under docs/config/.
• The same-directory JSON+crushrc warning doesn't name the conflicting keys; for users migrating incrementally, more detail would help.
• There's no $schema/IDE-autocomplete story for crushrc, which JSON users will lose when they migrate.

Code nits:

• isIntOption in internal/shellconfig/options.go is dead (always returns false). Drop or wire up.
• permissions allow + permissions deny interaction (deny wins via disabled_tools) isn't documented or tested. A small test would pin the precedence.
• The shellconfig_*_test.go helpers isolate HOME/XDG_* but don't override CRUSH_GLOBAL_CONFIG/CRUSH_GLOBAL_DATA, so a dev/CI with those set could pull in stray global config. The sibling load_test.go does override them.

Nice work on this — the imperative-map builder that emits one JSON blob per file (rather than per-builtin fragments) is the right evolution, and option reset / provider remove working in-order within a file is a good UX.

[meowgorithm]

Good stuff in here, all of which I would have totally missed. Thank you!