@ggui-ai/ui-gen 0.20.0 → 0.22.0

package/dist/harness/llm-trace-sink.d.ts

@@ -34,7 +34,13 @@ import '@ggui-ai/protocol';
  */
 
 /** Provider tag that produced the call. Mirrors {@link AgentConfig.provider}. */
-type LlmTraceProvider = 'anthropic' | 'openai' | 'google' | 'openrouter';
+/**
+ * `claude-code-login` is the Anthropic model reached through the machine's
+ * own Claude Code login (ggui#1185) rather than a key — a trace reader can
+ * tell the two apart, which the routing vocabulary (`provider: 'anthropic'`)
+ * deliberately does not.
+ */
+type LlmTraceProvider = 'anthropic' | 'claude-code-login' | 'openai' | 'google' | 'openrouter';
 /** Which `LLMAgent` method initiated the call. */
 type LlmTraceKind = 'callText' | 'callTools' | 'callWithTools' | 'callStructured' | 'callVision';
 /**

package/dist/harness/llm-trace-sink.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/harness/llm-trace-sink.ts"],"names":[],"mappings":";AAwGA,IAAI,UAAA,GAAkC,IAAA;AAM/B,SAAS,gBAAgB,IAAA,EAAiC;AAC/D,EAAA,UAAA,GAAa,IAAA;AACf;AAGO,SAAS,eAAA,GAAuC;AACrD,EAAA,OAAO,UAAA;AACT;AAOO,SAAS,kBAAkB,KAAA,EAA4B;AAC5D,EAAA,MAAM,IAAA,GAAO,UAAA;AACb,EAAA,IAAI,CAAC,IAAA,EAAM;AACX,EAAA,IAAI;AACF,IAAA,IAAA,CAAK,KAAK,KAAK,CAAA;AAAA,EACjB,CAAA,CAAA,MAAQ;AAAA,EAER;AACF;AAGO,SAAS,eACd,KAAA,EACwE;AACxE,EAAA,OAAO,KAAA,CAAM,GAAA,CAAI,CAAC,CAAA,MAAO,EAAE,IAAA,EAAM,CAAA,CAAE,IAAA,EAAM,WAAA,EAAa,CAAA,CAAE,WAAA,EAAY,CAAE,CAAA;AACxE;AAMO,SAAS,aAAA,GAAwB;AACtC,EAAA,OACE,IAAA,CAAK,GAAA,EAAI,CAAE,QAAA,CAAS,EAAE,CAAA,GACtB,IAAA,CAAK,MAAA,EAAO,CAAE,QAAA,CAAS,EAAE,CAAA,CAAE,KAAA,CAAM,GAAG,EAAE,CAAA;AAE1C","file":"llm-trace-sink.js","sourcesContent":["/**\n * Live LLM trace sink — devtools introspection of every LLM call the\n * generation harness makes.\n *\n * **Distinct from {@link TelemetrySink} and {@link AuditSink}.**\n *   - **Telemetry** (mcp-server-core) = ops signals (`pair.completed`,\n *     `render.duration_ms`). Sink may be lossy and drops on backpressure.\n *     Attributes are flat scalars (`string | number | boolean`).\n *   - **Audit** (mcp-server-core) = compliance log of privileged\n *     mutations. Durable, async, may throw.\n *   - **LLM trace** (this) = devtools-only, hostable in-memory ring\n *     buffer of every LLM call's full payload (system prompt, user\n *     prompt, tool defs, completion, token counts, cache hits, error).\n *     Lossy by default. A hosted closed runtime may swap in a durable\n *     sink for internal debugging; OSS ships an in-memory sink the\n *     `/devtools/llm-trace` page reads.\n *\n * **Why module-level registry instead of constructor injection.** The\n * harness constructs LLM agents on every generation via\n * {@link createAgent}. Threading a sink through every call site\n * (createAgent → run-coding-turn → llm-evaluator → file-agent → …)\n * touches ~5 packages for a devtools-only surface. The OSS ggui server\n * is a single process per CLI invocation — global state has no\n * confusion-cost there. A hosted deployment isolates per request via a\n * process pool, so a global per-pool is also safe. If we ever serve\n * several apps inside one process we'll thread it then.\n *\n * **Default = no sink.** When unset, the router emits nothing and\n * spends no CPU formatting events. Passing `null` removes a previously\n * registered sink.\n */\nimport type { LLMToolDef } from '../llm.js';\n\n/** Provider tag that produced the call. Mirrors {@link AgentConfig.provider}. */\nexport type LlmTraceProvider = 'anthropic' | 'openai' | 'google' | 'openrouter';\n\n/** Which `LLMAgent` method initiated the call. */\nexport type LlmTraceKind =\n  | 'callText'\n  | 'callTools'\n  | 'callWithTools'\n  | 'callStructured'\n  | 'callVision';\n\n/**\n * One LLM call. Emitted **after** the call completes (success or error)\n * — single event, not start/end split. Devtools UI is timeline-friendly\n * because each event carries `at` (start) + `endedAt` (end).\n */\nexport interface LlmTraceEvent {\n  /** Random per-event ID. */\n  readonly id: string;\n  /** Epoch ms when the call was issued. */\n  readonly at: number;\n  /** Epoch ms when the call completed (success or error). */\n  readonly endedAt: number;\n  /** `endedAt - at` — convenience. */\n  readonly durationMs: number;\n  readonly provider: LlmTraceProvider;\n  readonly model: string;\n  readonly kind: LlmTraceKind;\n  readonly systemPrompt?: string;\n  readonly userPrompt?: string;\n  /**\n   * Tool defs sent on the call (if any). Stripped to `name + description`\n   * — full JSON schema would balloon the trace and is rarely useful in\n   * the operator UI.\n   */\n  readonly tools?: ReadonlyArray<{\n    readonly name: string;\n    readonly description: string;\n  }>;\n  /** Set on successful completion. */\n  readonly result?: {\n    readonly inputTokens: number;\n    readonly outputTokens: number;\n    /** Anthropic only. Bytes counted toward the cache write. */\n    readonly cacheCreated?: number;\n    /** Anthropic only. Bytes served from cache (90% discounted). */\n    readonly cacheRead?: number;\n    /** Set when the call returned text content. */\n    readonly text?: string;\n    /** Set when the call returned tool_use blocks. */\n    readonly toolCalls?: ReadonlyArray<{\n      readonly name: string;\n      readonly input: unknown;\n    }>;\n    /** Set for `callWithTools` — number of agentic turns consumed. */\n    readonly turnsUsed?: number;\n  };\n  /** Set on failure. `result` is undefined. */\n  readonly error?: { readonly message: string };\n}\n\n/**\n * Sink that receives one event per LLM call. Implementations MUST be\n * sync + non-throwing — the harness fires events on the hot path and\n * cannot tolerate backpressure or rejected promises. Buffer + drop or\n * fan out to a queue inside the implementation.\n */\nexport interface LlmTraceSink {\n  emit(event: LlmTraceEvent): void;\n}\n\nlet activeSink: LlmTraceSink | null = null;\n\n/**\n * Register the active sink. Pass `null` to remove. Subsequent\n * {@link emitLlmTraceEvent} calls dispatch to this sink.\n */\nexport function setLlmTraceSink(sink: LlmTraceSink | null): void {\n  activeSink = sink;\n}\n\n/** Read the active sink. Mostly for tests. */\nexport function getLlmTraceSink(): LlmTraceSink | null {\n  return activeSink;\n}\n\n/**\n * Internal — used by `LLMAgent` adapters. No-op when no sink is\n * registered. Swallows sink-thrown errors (a broken devtools sink must\n * not break generation).\n */\nexport function emitLlmTraceEvent(event: LlmTraceEvent): void {\n  const sink = activeSink;\n  if (!sink) return;\n  try {\n    sink.emit(event);\n  } catch {\n    // Devtools sink is allowed to be buggy — generation must not die.\n  }\n}\n\n/** Strip `LLMToolDef[]` to the trace-friendly subset. */\nexport function summarizeTools(\n  tools: ReadonlyArray<LLMToolDef>,\n): ReadonlyArray<{ readonly name: string; readonly description: string }> {\n  return tools.map((t) => ({ name: t.name, description: t.description }));\n}\n\n/**\n * Crockford-style random ID. Crypto.randomUUID() would do, but we want\n * to keep this dep-free for non-Node runtimes the harness might run in.\n */\nexport function newLlmTraceId(): string {\n  return (\n    Date.now().toString(36) +\n    Math.random().toString(36).slice(2, 10)\n  );\n}\n"]}
+{"version":3,"sources":["../../src/harness/llm-trace-sink.ts"],"names":[],"mappings":";AA8GA,IAAI,UAAA,GAAkC,IAAA;AAM/B,SAAS,gBAAgB,IAAA,EAAiC;AAC/D,EAAA,UAAA,GAAa,IAAA;AACf;AAGO,SAAS,eAAA,GAAuC;AACrD,EAAA,OAAO,UAAA;AACT;AAOO,SAAS,kBAAkB,KAAA,EAA4B;AAC5D,EAAA,MAAM,IAAA,GAAO,UAAA;AACb,EAAA,IAAI,CAAC,IAAA,EAAM;AACX,EAAA,IAAI;AACF,IAAA,IAAA,CAAK,KAAK,KAAK,CAAA;AAAA,EACjB,CAAA,CAAA,MAAQ;AAAA,EAER;AACF;AAGO,SAAS,eACd,KAAA,EACwE;AACxE,EAAA,OAAO,KAAA,CAAM,GAAA,CAAI,CAAC,CAAA,MAAO,EAAE,IAAA,EAAM,CAAA,CAAE,IAAA,EAAM,WAAA,EAAa,CAAA,CAAE,WAAA,EAAY,CAAE,CAAA;AACxE;AAMO,SAAS,aAAA,GAAwB;AACtC,EAAA,OACE,IAAA,CAAK,GAAA,EAAI,CAAE,QAAA,CAAS,EAAE,CAAA,GACtB,IAAA,CAAK,MAAA,EAAO,CAAE,QAAA,CAAS,EAAE,CAAA,CAAE,KAAA,CAAM,GAAG,EAAE,CAAA;AAE1C","file":"llm-trace-sink.js","sourcesContent":["/**\n * Live LLM trace sink — devtools introspection of every LLM call the\n * generation harness makes.\n *\n * **Distinct from {@link TelemetrySink} and {@link AuditSink}.**\n *   - **Telemetry** (mcp-server-core) = ops signals (`pair.completed`,\n *     `render.duration_ms`). Sink may be lossy and drops on backpressure.\n *     Attributes are flat scalars (`string | number | boolean`).\n *   - **Audit** (mcp-server-core) = compliance log of privileged\n *     mutations. Durable, async, may throw.\n *   - **LLM trace** (this) = devtools-only, hostable in-memory ring\n *     buffer of every LLM call's full payload (system prompt, user\n *     prompt, tool defs, completion, token counts, cache hits, error).\n *     Lossy by default. A hosted closed runtime may swap in a durable\n *     sink for internal debugging; OSS ships an in-memory sink the\n *     `/devtools/llm-trace` page reads.\n *\n * **Why module-level registry instead of constructor injection.** The\n * harness constructs LLM agents on every generation via\n * {@link createAgent}. Threading a sink through every call site\n * (createAgent → run-coding-turn → llm-evaluator → file-agent → …)\n * touches ~5 packages for a devtools-only surface. The OSS ggui server\n * is a single process per CLI invocation — global state has no\n * confusion-cost there. A hosted deployment isolates per request via a\n * process pool, so a global per-pool is also safe. If we ever serve\n * several apps inside one process we'll thread it then.\n *\n * **Default = no sink.** When unset, the router emits nothing and\n * spends no CPU formatting events. Passing `null` removes a previously\n * registered sink.\n */\nimport type { LLMToolDef } from '../llm.js';\n\n/** Provider tag that produced the call. Mirrors {@link AgentConfig.provider}. */\n/**\n * `claude-code-login` is the Anthropic model reached through the machine's\n * own Claude Code login (ggui#1185) rather than a key — a trace reader can\n * tell the two apart, which the routing vocabulary (`provider: 'anthropic'`)\n * deliberately does not.\n */\nexport type LlmTraceProvider = 'anthropic' | 'claude-code-login' | 'openai' | 'google' | 'openrouter';\n\n/** Which `LLMAgent` method initiated the call. */\nexport type LlmTraceKind =\n  | 'callText'\n  | 'callTools'\n  | 'callWithTools'\n  | 'callStructured'\n  | 'callVision';\n\n/**\n * One LLM call. Emitted **after** the call completes (success or error)\n * — single event, not start/end split. Devtools UI is timeline-friendly\n * because each event carries `at` (start) + `endedAt` (end).\n */\nexport interface LlmTraceEvent {\n  /** Random per-event ID. */\n  readonly id: string;\n  /** Epoch ms when the call was issued. */\n  readonly at: number;\n  /** Epoch ms when the call completed (success or error). */\n  readonly endedAt: number;\n  /** `endedAt - at` — convenience. */\n  readonly durationMs: number;\n  readonly provider: LlmTraceProvider;\n  readonly model: string;\n  readonly kind: LlmTraceKind;\n  readonly systemPrompt?: string;\n  readonly userPrompt?: string;\n  /**\n   * Tool defs sent on the call (if any). Stripped to `name + description`\n   * — full JSON schema would balloon the trace and is rarely useful in\n   * the operator UI.\n   */\n  readonly tools?: ReadonlyArray<{\n    readonly name: string;\n    readonly description: string;\n  }>;\n  /** Set on successful completion. */\n  readonly result?: {\n    readonly inputTokens: number;\n    readonly outputTokens: number;\n    /** Anthropic only. Bytes counted toward the cache write. */\n    readonly cacheCreated?: number;\n    /** Anthropic only. Bytes served from cache (90% discounted). */\n    readonly cacheRead?: number;\n    /** Set when the call returned text content. */\n    readonly text?: string;\n    /** Set when the call returned tool_use blocks. */\n    readonly toolCalls?: ReadonlyArray<{\n      readonly name: string;\n      readonly input: unknown;\n    }>;\n    /** Set for `callWithTools` — number of agentic turns consumed. */\n    readonly turnsUsed?: number;\n  };\n  /** Set on failure. `result` is undefined. */\n  readonly error?: { readonly message: string };\n}\n\n/**\n * Sink that receives one event per LLM call. Implementations MUST be\n * sync + non-throwing — the harness fires events on the hot path and\n * cannot tolerate backpressure or rejected promises. Buffer + drop or\n * fan out to a queue inside the implementation.\n */\nexport interface LlmTraceSink {\n  emit(event: LlmTraceEvent): void;\n}\n\nlet activeSink: LlmTraceSink | null = null;\n\n/**\n * Register the active sink. Pass `null` to remove. Subsequent\n * {@link emitLlmTraceEvent} calls dispatch to this sink.\n */\nexport function setLlmTraceSink(sink: LlmTraceSink | null): void {\n  activeSink = sink;\n}\n\n/** Read the active sink. Mostly for tests. */\nexport function getLlmTraceSink(): LlmTraceSink | null {\n  return activeSink;\n}\n\n/**\n * Internal — used by `LLMAgent` adapters. No-op when no sink is\n * registered. Swallows sink-thrown errors (a broken devtools sink must\n * not break generation).\n */\nexport function emitLlmTraceEvent(event: LlmTraceEvent): void {\n  const sink = activeSink;\n  if (!sink) return;\n  try {\n    sink.emit(event);\n  } catch {\n    // Devtools sink is allowed to be buggy — generation must not die.\n  }\n}\n\n/** Strip `LLMToolDef[]` to the trace-friendly subset. */\nexport function summarizeTools(\n  tools: ReadonlyArray<LLMToolDef>,\n): ReadonlyArray<{ readonly name: string; readonly description: string }> {\n  return tools.map((t) => ({ name: t.name, description: t.description }));\n}\n\n/**\n * Crockford-style random ID. Crypto.randomUUID() would do, but we want\n * to keep this dep-free for non-Node runtimes the harness might run in.\n */\nexport function newLlmTraceId(): string {\n  return (\n    Date.now().toString(36) +\n    Math.random().toString(36).slice(2, 10)\n  );\n}\n"]}

package/dist/harness/prompts.js

@@ -122,7 +122,7 @@ var SHELL_HINTS = {
 - Padding: use var(--ggui-spacing-2) for inner elements, var(--ggui-spacing-4) for sections. Tight.
 - Scrolling: parent scrolls (chat feed) \u2014 component should NOT have internal scroll
 - Layout: single column, compact. No sidebar. Stack everything vertically.
-- Width: width: 100% (fills parent bubble). Do NOT set max-width.`,
+- Width: the root is width: 100% (fills the bubble) \u2014 never cap the ROOT. A prose column inside it carries the reading cap (<Container maxWidth="sm">), which is inert at bubble width.`,
   fullscreen: `**Fullscreen Shell \u2014 Takes over entire viewport**
 - Container: the whole viewport, edge to edge \u2014 the frame stretches your root to its full height (you never set a viewport height yourself)
 - Chrome: NONE \u2014 your component IS the entire UI. You own all visual chrome.