Visual

Shell 1 command (grep -rn "friendlyLabel\|nounFor" ./tests ./src --include="*.test.ts" 2>/dev/null | head -10) ctrl+o to expand
Write 1 file (README.md) +148 -129 ctrl+o to expand
Shell 1 command & Read 1 file ctrl+o to expand
⠋ Read file

Installation

Pi package manager:

pi install npm:@fahmiirsyadk/pi-minimal-toolcall

Git Repository:

pi install git:github.com/fahmiirsyadk/pi-minimal-toolcall

Local clone (for development or pinning to a fork):

git clone https://github.com/fahmiirsyadk/pi-minimal-toolcall.git ~/pi-minimal-toolcall

Then add to your ~/.pi/agent/settings.json:

{
  "extensions": [
    "~/pi-minimal-toolcall/index.ts"
  ]
}

Quick start for customizing

Edit ~/.pi/agent/extensions/pi-minimal-toolcall/config.json (or $PI_CODING_AGENT_DIR/extensions/pi-minimal-toolcall/config.json):

{
  "toolsExpandedByDefault": true,
  "hiddenThinkingLabel": "",
  "registerToolOverrides": { "bash": false },
  "groupingMode": "proximity",
  "showArgOnSummary": "always"
}

Then /reload to apply.

Three starter presets ship in config/presets/ — copy the one you want to your config file:

• calm.json — the defaults (one row per group, collapsed tools, thinking label).
• verbose.json — expanded previews, larger body cap, args on multi-tool rows.
• minimal.json — bare frames, no args, no diff, no ✗, empty thinking label.

What you get (Configuration Option)

Behavior	Default	How to change
Tool rows	collapsed	toolsExpandedByDefault: true
Thinking	hidden behind thinking label	hiddenThinkingLabel: "..."
Grouping mode	proximity (any tool call joins until text/thinking)	groupingMode: "consecutive" | "none"
Latest arg on summary	single-tool groups only	showArgOnSummary: "always" | "never"
Aggregated +N -M	on	showDiffSuffix: false
Per-tool ✗	on	showErrorMark: false
Write expand	full content, syntax-highlighted	writeExpandMode: "summary" | "both"
Expanded body cap	200 lines	expandedBodyMaxLines: <n>
Spinner	⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏ @ 80ms	spinnerFrames: ["a","b"], spinnerIntervalMs: <n>
Per-tool ownership	all 7 built-ins	registerToolOverrides: { "read": false, ... }
Batch tools (read_files, edit_files, grep_files, find_files)	on	batchToolsEnabled: false
Debug log	off	debug: true (writes to <agent-dir>/.../debug/debug.log)
customToolOverrides	loaded + normalized; not yet rendered	(awaiting SDK support — see below)

customToolOverrides (not yet wired)

The customToolOverrides config field is parsed and normalized (built-in tool names are rejected, _comment* keys are dropped, booleans are coerced, invalid outputMode values fall back to "summary", entries are capped at 256 to guard against accidental JSON imports). The SDK's public pi.getAllTools() API currently returns ToolInfo (name + description + parameters) only — not the full ToolDefinition (no execute / renderCall / renderResult) — so the package cannot wrap a non-builtin tool's execute from the extension layer. Your config is preserved on disk and will activate the moment the SDK exposes the full definition.

Doctor command

/minimal-toolcall-doctor is a read-only diagnostic command. It prints the resolved config (loaded from disk or the shipped defaults) and flags likely footguns:

• spinnerIntervalMs < 50 (CPU peg)
• expandedBodyMaxLines ≤ 0 (collapses every expanded view)
• All 7 built-in tools disabled (the package becomes a no-op)
• batchToolsEnabled: false (the batch tools won't be registered)
• Non-empty customToolOverrides (the field is parsed but doesn't yet affect rendering)

The command replaces the dropped 0.2.0 /minimal-toolcall runtime. It does not mutate state — reloads and config changes go through the file + /reload.

How grouping works

A group is a run of tool calls with no text or thinking between them, regardless of tool name. Frozen when a text or thinking block appears (message_update with text_start / thinking_start), or when a new agent loop starts (agent_start — separates prompts).

read, read, read                  → Read 3 files
read, read, bash, read            → one row (proximity)
text or thinking
bash, bash                        → Shell 2 commands (new group)

groupingMode: "consecutive" restores the old "different tool name = new group" behavior. "none" makes every call its own row.

Batch tools

The package registers four batch tools the model can call instead of repeating built-ins:

Tool	Input
read_files	{ paths, offset?, limit? }
edit_files	{ edits: [{ path, oldText, newText }] }
grep_files	{ queries: [{ pattern, path? }] }
find_files	{ queries: [{ pattern, path? }] }

Each renders as one row; ctrl+o expands to per-item ✓/✗ status plus aggregated output. Partial failures surface per item.

Reload

/reload re-reads config.json and re-registers. Tool ownership, the per-tool grouping session, and the debug-log path all take effect on the next session_start.

Compatibility

• Pi 0.79.0+. Peer dependencies: @earendil-works/pi-coding-agent, pi-ai, pi-tui (all * range — pi bundles these).
• This package only changes tool-call rendering. We touch two pieces of global UI state on session_start (and only these two — we don't touch the working indicator, the prompt box, or anything else outside our scope):
  • ctx.ui.setToolsExpanded(config.toolsExpandedByDefault) — the resting state of tool output (collapsed vs expanded). ctrl+o toggles it during a session.
  • ctx.ui.setHiddenThinkingLabel(config.hiddenThinkingLabel) — the label on the thinking block.
• Tools registered by other extensions render with their own renderer and break proximity groups.

Development

npm install
npm run check        # typecheck + test
npm run pack:dry     # preview the published tarball

Uses tsc for typecheck and tsx --test for tests. No bundler, no formatter, no linter — keep the dep tree small.

License

MIT