localpi 0.4.0 → 0.5.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -1,10 +1,10 @@
1
1
  # localpi
2
2
 
3
- Localpi is a local Pi launcher for open-weight models.
3
+ Localpi is a Swiss army knife for running Pi with local inference engines.
4
4
 
5
5
  By default, Localpi discovers available local providers, lets you choose when more than one model is loaded, points Pi at the selected model, and writes Pi config for the other discovered models so `/model` can switch among them during the session.
6
6
 
7
- Localpi supports LM Studio, vLLM, custom OpenAI-compatible servers, and an optional managed `llama-server` fallback.
7
+ Localpi is meant to be the practical bridge from Pi to local inference stacks such as llama.cpp/`llama-server`, vLLM, SGLang, LM Studio, Ollama, and custom provider endpoints.
8
8
 
9
9
  Localpi is intentionally generic. It does not contain classifier prompts, dataset workflows, GitHub routing logic, or final-schema output machinery. Structured classifier runs belong in caller tools such as `localpager-agent`.
10
10
 
@@ -18,6 +18,12 @@ See:
18
18
  npm install -g localpi
19
19
  ```
20
20
 
21
+ Or the latest GitHub release directly:
22
+
23
+ ```bash
24
+ npm install -g https://github.com/osolmaz/localpi/releases/latest/download/localpi.tgz
25
+ ```
26
+
21
27
  During development:
22
28
 
23
29
  ```bash
@@ -74,6 +80,14 @@ Localpi launches Pi with:
74
80
 
75
81
  The approval gate makes failed or denied tool calls explicit to the model so the model does not claim that a blocked command ran.
76
82
 
83
+ ## Diffusion Canvas Visualizer
84
+
85
+ The live diffusion canvas visualizer (watching DiffusionGemma denoise its
86
+ answer in the TUI) lives in its own project now:
87
+ [diffusionpi](https://github.com/osolmaz/diffusionpi). It is a pi-factory app
88
+ bundle plus a standalone Pi widget package; the widget also installs into any
89
+ Pi session via `pi install`.
90
+
77
91
  ## LM Studio Alternative
78
92
 
79
93
  LM Studio exposes an OpenAI-compatible endpoint, usually:
@@ -115,7 +129,7 @@ Run an endless TUI demo:
115
129
  localpi --demo --model gemma-e4b
116
130
  ```
117
131
 
118
- Demo mode requires an explicit model, opens the normal Pi TUI, and keeps one live Pi session so followup prompts continue from the first prompt while Pi owns streaming, tok/s status, slash commands, and exit behavior.
132
+ Demo mode requires an explicit model, opens the normal Pi TUI, and keeps one live Pi session so followup prompts continue from the first prompt while Pi owns streaming, tok/s status, slash commands, and exit behavior. Under the hood it loads the shared [pi-demo-mode](https://github.com/osolmaz/pi-demo-mode) extension, configured through the `--demo-*` flags below.
119
133
 
120
134
  Override the demo prompts:
121
135
 
@@ -174,6 +188,18 @@ Stop the managed `llama-server` runtime:
174
188
  localpi --stop
175
189
  ```
176
190
 
191
+ ## Demo Grid and Recording
192
+
193
+ The `localpi grid` and `localpi record` subcommands moved to
194
+ [demowall](https://github.com/osolmaz/demowall), a standalone tool that runs
195
+ N copies of any command in a tiled tmux wall and records tmux sessions to
196
+ video. A wall of localpi demo sessions is:
197
+
198
+ ```bash
199
+ demowall grid --concurrency 4 --start -- localpi --demo --model gemma-e4b
200
+ demowall record --session demowall-<timestamp> --out demo.mp4 --seconds 60
201
+ ```
202
+
177
203
  ## Options
178
204
 
179
205
  - `--runtime <auto|llama-server|lmstudio|vllm|openai-compatible>`: runtime backend. Default: `auto`
@@ -1,4 +1,4 @@
1
- import { runPiApp } from "@dutifuldev/pi-factory";
1
+ import { runPiApp } from "@osolmaz/pi-factory";
2
2
  import { errorMessage, fail, ok } from "../common/result.js";
3
3
  import { parseLocalpiArgs, usage } from "../localpi/options.js";
4
4
  import { aliasListOutput, connectionStatus, resolveRuntime, statusOutput, stopRuntime } from "../localpi/runtime.js";
@@ -22,7 +22,9 @@ export async function run(args) {
22
22
  });
23
23
  const connection = await resolveRuntime(options);
24
24
  const selectorOptions = startupModelSelectorOptions(options, connection);
25
- const extensions = await writeDefaultExtensions(options, selectorOptions === undefined ? {} : { startupModelSelector: selectorOptions });
25
+ const extensions = await writeDefaultExtensions(options, {
26
+ ...(selectorOptions === undefined ? {} : { startupModelSelector: selectorOptions })
27
+ });
26
28
  const app = createLocalpiAppDefinition(options, connection, extensions);
27
29
  return await launchResolvedRuntime(app, connection);
28
30
  }
File without changes
@@ -4,6 +4,18 @@ const localpiAppIdentity = {
4
4
  name: "localpi",
5
5
  version: localpiVersion
6
6
  };
7
+ const demoLaunchOverrideBooleanFlags = new Set([
8
+ "--no-tools",
9
+ "-nt",
10
+ "--no-builtin-tools",
11
+ "-nbt",
12
+ "--approve",
13
+ "-a",
14
+ "--no-approve",
15
+ "-na"
16
+ ]);
17
+ const demoLaunchOverrideValueFlags = new Set(["--tools", "-t", "--exclude-tools", "-xt"]);
18
+ const demoLaunchOverrideEqualsFlags = ["--tools=", "--exclude-tools="];
7
19
  export function createLocalpiAppDefinition(options, connection, extensions) {
8
20
  return {
9
21
  ...localpiAppIdentity,
@@ -22,9 +34,37 @@ function appDirectories(options) {
22
34
  function piCommand(options) {
23
35
  return {
24
36
  piCommand: options.piCommand,
25
- forwardedArgs: options.forwardedArgs
37
+ forwardedArgs: options.demo ? demoForwardedArgs(options.forwardedArgs) : options.forwardedArgs
26
38
  };
27
39
  }
40
+ function demoForwardedArgs(args) {
41
+ return [...withoutConflictingDemoFlags(args), "--no-tools", "--no-approve"];
42
+ }
43
+ function withoutConflictingDemoFlags(args) {
44
+ const filtered = [];
45
+ for (let index = 0; index < args.length; index += 1) {
46
+ const arg = args[index];
47
+ if (arg === undefined) {
48
+ continue;
49
+ }
50
+ if (isDemoLaunchOverrideFlag(arg)) {
51
+ if (isValueFlag(arg) && !arg.includes("=")) {
52
+ index += 1;
53
+ }
54
+ continue;
55
+ }
56
+ filtered.push(arg);
57
+ }
58
+ return filtered;
59
+ }
60
+ function isDemoLaunchOverrideFlag(arg) {
61
+ return (demoLaunchOverrideBooleanFlags.has(arg) ||
62
+ isValueFlag(arg) ||
63
+ demoLaunchOverrideEqualsFlags.some((flag) => arg.startsWith(flag)));
64
+ }
65
+ function isValueFlag(arg) {
66
+ return demoLaunchOverrideValueFlags.has(arg);
67
+ }
28
68
  function runtimeSelection(options, connection) {
29
69
  const selection = {
30
70
  providers: providersForConnection(options, connection),
@@ -103,6 +143,7 @@ function extensionConfig(extensions) {
103
143
  }
104
144
  return {
105
145
  extensions: extensions.paths.map((extensionPath) => ({ path: extensionPath })),
106
- appendSystemPrompts: [extensions.systemPrompt]
146
+ appendSystemPrompts: [extensions.systemPrompt],
147
+ ...(Object.keys(extensions.env).length === 0 ? {} : { env: extensions.env })
107
148
  };
108
149
  }
@@ -1,6 +1,9 @@
1
- export function tokenStatusExtensionSource() {
1
+ export function tokenStatusExtensionSource(options = {}) {
2
+ const includeContext = options.includeContext ?? true;
2
3
  return `import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
3
4
 
5
+ const includeContext = ${JSON.stringify(includeContext)};
6
+
4
7
  type Usage = {
5
8
  input?: number;
6
9
  output?: number;
@@ -11,6 +14,7 @@ type Usage = {
11
14
 
12
15
  type TurnState = {
13
16
  startedAt: number;
17
+ firstOutputAt?: number;
14
18
  outputText: string;
15
19
  estimatedOutputTokens: number;
16
20
  lastStatusAt: number;
@@ -33,6 +37,7 @@ export default function localpiTokenStatus(pi: ExtensionAPI): void {
33
37
  if (!ctx.hasUI || state === undefined) {
34
38
  return;
35
39
  }
40
+ const now = Date.now();
36
41
  const update = textUpdateFromUnknown(event.assistantMessageEvent ?? event.message ?? event);
37
42
  if (update.kind === "delta") {
38
43
  state.outputText += update.text;
@@ -40,11 +45,14 @@ export default function localpiTokenStatus(pi: ExtensionAPI): void {
40
45
  state.outputText = update.text;
41
46
  }
42
47
  state.estimatedOutputTokens = Math.ceil(state.outputText.length / 4);
43
- if (Date.now() - state.lastStatusAt < 250) {
48
+ if (state.firstOutputAt === undefined && state.outputText.length > 0) {
49
+ state.firstOutputAt = now;
50
+ }
51
+ if (now - state.lastStatusAt < 250) {
44
52
  return;
45
53
  }
46
- state.lastStatusAt = Date.now();
47
- ctx.ui.setStatus("localpi-perf", ctx.ui.theme.fg("dim", statusText(state)));
54
+ state.lastStatusAt = now;
55
+ ctx.ui.setStatus("localpi-perf", ctx.ui.theme.fg("dim", statusText(state, now)));
48
56
  });
49
57
 
50
58
  pi.on("turn_end", (event, ctx) => {
@@ -65,10 +73,14 @@ export default function localpiTokenStatus(pi: ExtensionAPI): void {
65
73
  const input = usage?.input ?? 0;
66
74
  const cacheRead = usage?.cacheRead ?? 0;
67
75
  const cacheWrite = usage?.cacheWrite ?? 0;
68
- const elapsedSeconds = elapsed(state);
69
- const context = ctx.getContextUsage();
70
- const contextText =
71
- context && context.percent !== null
76
+ const now = Date.now();
77
+ const elapsedSeconds = elapsed(state, now);
78
+ const decodeSeconds = generationElapsed(state, now);
79
+ const prefillText = prefillStatusText(state, input, cacheWrite);
80
+ const context = includeContext ? ctx.getContextUsage() : undefined;
81
+ const contextText = !includeContext
82
+ ? undefined
83
+ : context && context.percent !== null
72
84
  ? \`ctx \${Math.round(context.percent)}%/\${Math.round(context.contextWindow / 1000)}k\`
73
85
  : "ctx ?";
74
86
 
@@ -77,7 +89,8 @@ export default function localpiTokenStatus(pi: ExtensionAPI): void {
77
89
  ctx.ui.theme.fg(
78
90
  "dim",
79
91
  [
80
- \`\${(output / elapsedSeconds).toFixed(1)} tok/s\`,
92
+ \`gen \${(output / decodeSeconds).toFixed(1)} tok/s\`,
93
+ prefillText,
81
94
  \`out \${output}\`,
82
95
  \`in \${input}\`,
83
96
  cacheRead > 0 ? \`cache \${cacheRead}\` : undefined,
@@ -98,13 +111,48 @@ export default function localpiTokenStatus(pi: ExtensionAPI): void {
98
111
  });
99
112
  }
100
113
 
101
- function statusText(state: TurnState): string {
102
- const elapsedSeconds = elapsed(state);
103
- return \`\${(state.estimatedOutputTokens / elapsedSeconds).toFixed(1)} tok/s | out ~\${state.estimatedOutputTokens} | \${elapsedSeconds.toFixed(1)}s\`;
114
+ function statusText(state: TurnState, now: number): string {
115
+ const elapsedSeconds = elapsed(state, now);
116
+ if (state.firstOutputAt === undefined) {
117
+ return \`prefill \${elapsedSeconds.toFixed(1)}s | out ~\${state.estimatedOutputTokens}\`;
118
+ }
119
+ const decodeSeconds = generationElapsed(state, now);
120
+ const prefillSeconds = secondsBetween(state.startedAt, state.firstOutputAt);
121
+ return [
122
+ \`gen \${(state.estimatedOutputTokens / decodeSeconds).toFixed(1)} tok/s\`,
123
+ \`out ~\${state.estimatedOutputTokens}\`,
124
+ \`prefill \${prefillSeconds.toFixed(1)}s\`,
125
+ \`total \${elapsedSeconds.toFixed(1)}s\`
126
+ ].join(" | ");
127
+ }
128
+
129
+ function prefillStatusText(
130
+ state: TurnState,
131
+ input: number,
132
+ cacheWrite: number
133
+ ): string | undefined {
134
+ const tokens = prefillTokenCount(input, cacheWrite);
135
+ if (state.firstOutputAt === undefined || tokens <= 0) {
136
+ return undefined;
137
+ }
138
+ const seconds = secondsBetween(state.startedAt, state.firstOutputAt);
139
+ return \`prefill \${(tokens / seconds).toFixed(1)} tok/s\`;
140
+ }
141
+
142
+ function prefillTokenCount(input: number, cacheWrite: number): number {
143
+ return Math.max(input + cacheWrite, 0);
144
+ }
145
+
146
+ function generationElapsed(state: TurnState, now: number): number {
147
+ return secondsBetween(state.firstOutputAt ?? state.startedAt, now);
148
+ }
149
+
150
+ function elapsed(state: TurnState, now: number): number {
151
+ return secondsBetween(state.startedAt, now);
104
152
  }
105
153
 
106
- function elapsed(state: TurnState): number {
107
- return Math.max((Date.now() - state.startedAt) / 1000, 0.001);
154
+ function secondsBetween(start: number, end: number): number {
155
+ return Math.max((end - start) / 1000, 0.001);
108
156
  }
109
157
 
110
158
  type TextUpdate = {
@@ -1,8 +1,8 @@
1
1
  import { mkdir, writeFile } from "node:fs/promises";
2
+ import { createRequire } from "node:module";
2
3
  import path from "node:path";
3
4
  import { localpiSettingsPath } from "../localpi/settings-state.js";
4
5
  import { resolveDemoPrompts } from "./demo.js";
5
- import { demoModeExtensionSource } from "./extension-sources/demo-mode.js";
6
6
  import { startupModelSelectorExtensionSource } from "./extension-sources/startup-model-selector.js";
7
7
  import { thinkingControlExtensionSource } from "./extension-sources/thinking-control.js";
8
8
  import { tokenStatusExtensionSource } from "./extension-sources/token-status.js";
@@ -11,24 +11,40 @@ export async function writeDefaultExtensions(options, extensionOptions = {}) {
11
11
  const extensionDir = path.join(options.stateDir, "pi-extensions");
12
12
  await mkdir(extensionDir, { recursive: true });
13
13
  const paths = [];
14
+ let env = {};
14
15
  if (extensionOptions.startupModelSelector !== undefined) {
15
16
  paths.push(await writeExtension(extensionDir, "startup-model-selector.ts", startupModelSelectorExtensionSource(extensionOptions.startupModelSelector)));
16
17
  }
17
18
  if (options.demo) {
18
- paths.push(await writeExtension(extensionDir, "demo-mode.ts", demoModeExtensionSource(await resolveDemoPrompts(options))));
19
+ paths.push(demoModeExtensionPath());
20
+ const prompts = await resolveDemoPrompts(options);
21
+ env = {
22
+ ...env,
23
+ PI_DEMO_MODE: "1",
24
+ PI_DEMO_INITIAL_PROMPT: prompts.initial,
25
+ PI_DEMO_FOLLOWUP_PROMPT: prompts.followup
26
+ };
19
27
  }
20
28
  paths.push(await writeExtension(extensionDir, "thinking-control.ts", thinkingControlExtensionSource(localpiSettingsPath(options))));
21
29
  if (options.approval) {
22
30
  paths.push(await writeExtension(extensionDir, "tool-approval.ts", approvalExtensionSource()));
23
31
  }
24
32
  if (options.tokenStatus) {
25
- paths.push(await writeExtension(extensionDir, "token-status.ts", tokenStatusExtensionSource()));
33
+ paths.push(await writeExtension(extensionDir, "token-status.ts", tokenStatusExtensionSource({ includeContext: !options.demo })));
26
34
  }
27
35
  return {
28
36
  paths,
37
+ env,
29
38
  systemPrompt: localpiSystemPrompt(options.approval)
30
39
  };
31
40
  }
41
+ // Demo mode is the shared pi-demo-mode package (a git dependency), loaded
42
+ // straight from node_modules and configured through PI_DEMO_* env vars.
43
+ export function demoModeExtensionPath() {
44
+ const require = createRequire(import.meta.url);
45
+ const packageJson = require.resolve("pi-demo-mode/package.json");
46
+ return path.join(path.dirname(packageJson), "extensions", "demo-mode.ts");
47
+ }
32
48
  async function writeExtension(extensionDir, name, source) {
33
49
  const extensionPath = path.join(extensionDir, name);
34
50
  await writeFile(extensionPath, source, "utf8");
@@ -150,7 +150,7 @@ session dir: ~/.local/state/localpi/sessions
150
150
  Localpi installs two default extensions:
151
151
 
152
152
  - tool approval gate: ask before each tool call, and tell the model clearly when a tool call was blocked
153
- - token status: show live output token estimate while streaming and final exact token stats when usage data is available
153
+ - token status: show live generation speed while streaming, then final prefill and generation rates when usage data is available
154
154
 
155
155
  ## System Prompt
156
156
 
package/package.json CHANGED
@@ -1,17 +1,17 @@
1
1
  {
2
2
  "name": "localpi",
3
- "version": "0.4.0",
4
- "description": "Pi-compatible local model launcher with managed llama-server support.",
3
+ "version": "0.5.2",
4
+ "description": "Swiss army knife for running Pi with local inference engines.",
5
5
  "type": "module",
6
6
  "license": "MIT",
7
7
  "repository": {
8
8
  "type": "git",
9
- "url": "git+https://github.com/dutifuldev/localpi.git"
9
+ "url": "git+https://github.com/osolmaz/localpi.git"
10
10
  },
11
11
  "bugs": {
12
- "url": "https://github.com/dutifuldev/localpi/issues"
12
+ "url": "https://github.com/osolmaz/localpi/issues"
13
13
  },
14
- "homepage": "https://github.com/dutifuldev/localpi#readme",
14
+ "homepage": "https://github.com/osolmaz/localpi#readme",
15
15
  "publishConfig": {
16
16
  "access": "public"
17
17
  },
@@ -26,7 +26,8 @@
26
26
  "LICENSE"
27
27
  ],
28
28
  "scripts": {
29
- "build": "tsc -p tsconfig.json",
29
+ "build": "tsc -p tsconfig.json && chmod +x dist/src/cli/main.js",
30
+ "prepare": "npm run build",
30
31
  "format": "prettier --check .",
31
32
  "lint": "eslint .",
32
33
  "typecheck": "tsc --noEmit",
@@ -48,13 +49,14 @@
48
49
  "@vitest/coverage-v8": "^3.0.0",
49
50
  "eslint": "^9.0.0",
50
51
  "prettier": "^3.0.0",
51
- "slophammer-ts": "0.4.0",
52
+ "slophammer-ts": "0.4.1",
52
53
  "tsx": "^4.19.4",
53
54
  "typescript": "^5.0.0",
54
55
  "typescript-eslint": "^8.0.0",
55
56
  "vitest": "^3.0.0"
56
57
  },
57
58
  "dependencies": {
58
- "@dutifuldev/pi-factory": "^0.1.1"
59
+ "@osolmaz/pi-factory": "^0.1.2",
60
+ "pi-demo-mode": "github:osolmaz/pi-demo-mode#v0.1.0"
59
61
  }
60
62
  }
@@ -1,52 +0,0 @@
1
- export function demoModeExtensionSource(prompts) {
2
- const initialPromptSource = JSON.stringify(prompts.initial);
3
- const followupPromptSource = JSON.stringify(prompts.followup);
4
- return `import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
5
-
6
- const initialPrompt = ${initialPromptSource};
7
- const followupPrompt = ${followupPromptSource};
8
-
9
- export default function localpiDemoMode(pi: ExtensionAPI): void {
10
- let started = false;
11
- let stopped = false;
12
-
13
- pi.on("session_start", (event, ctx) => {
14
- if (started || stopped || event.reason !== "startup" || ctx.mode !== "tui") {
15
- return;
16
- }
17
- started = true;
18
- queueMicrotask(() => {
19
- if (!stopped) {
20
- pi.sendUserMessage(initialPrompt);
21
- }
22
- });
23
- });
24
-
25
- pi.on("turn_end", (event, ctx) => {
26
- if (!started || stopped || ctx.mode !== "tui") {
27
- return;
28
- }
29
- if (event.message.role !== "assistant") {
30
- return;
31
- }
32
- switch (event.message.stopReason) {
33
- case "aborted":
34
- case "error":
35
- stopped = true;
36
- return;
37
- case "toolUse":
38
- return;
39
- }
40
- queueMicrotask(() => {
41
- if (!stopped) {
42
- pi.sendUserMessage(followupPrompt, { deliverAs: "followUp" });
43
- }
44
- });
45
- });
46
-
47
- pi.on("session_shutdown", () => {
48
- stopped = true;
49
- });
50
- }
51
- `;
52
- }