@cabane/companion 0.6.73 → 0.6.75

package/dist/cli.js

@@ -3048,6 +3048,12 @@ var turnEventSchema = z5.discriminatedUnion("type", [
   // `progress` row); `terminal: true` is the turn's closing reply (commits as
   // the `final` row).
   z5.object({ type: z5.literal("text"), body: z5.string(), terminal: z5.boolean() }),
+  // Runtime/provider-authored prose surfaced alongside a failed turn. Unlike
+  // `text`, this is not the agent's narration or reply: the pump persists it as
+  // `runtime_notice`, and the transcript renders it in the shared SystemNote
+  // voice. Adapters emit it only from positive runtime evidence; the web never
+  // classifies English strings.
+  z5.object({ type: z5.literal("runtime_notice"), body: z5.string() }),
   // A tool-activity transition. Maps `onToolActivity` — `toolUseId`→`id`,
   // `toolName`→`name` (already prefix-stripped: `cabane_read`, not
   // `mcp__cabane__cabane_read`), `summary` is the short card label. `phase`
@@ -3181,6 +3187,7 @@ var turnDiagnosticsSchema = z6.object({
   eventCounts: z6.object({
     session: z6.number().int().nonnegative(),
     text: z6.number().int().nonnegative(),
+    runtime_notice: z6.number().int().nonnegative(),
     thinking: z6.number().int().nonnegative(),
     tool: z6.number().int().nonnegative(),
     result: z6.number().int().nonnegative()
@@ -3279,6 +3286,17 @@ function classifyErrorText(text) {
   if (BARE_LIMIT.test(t)) return { kind: "usage_capped" };
   return null;
 }
+function classifyRuntimeNoticeText(text) {
+  if (!text) return null;
+  const normalized = text.trim().replace(/\s+/g, " ");
+  if (!normalized) return null;
+  if (/^you(?:'|’)ve hit your (?:session|weekly|usage) limit(?:\s*[·—-]\s*resets?\s+.+)?$/i.test(
+    normalized
+  )) {
+    return { kind: "usage_capped" };
+  }
+  return null;
+}
 var AUTH_PATTERNS = [
   /authentication[_ ]error/,
   /invalid[_ ]?(x-)?api[_ ]?key/,
@@ -3346,6 +3364,8 @@ function isContentBearingEvent(event) {
       return normalizeCommitText(event.body) !== "";
     case "thinking":
       return normalizeCommitText(event.text) !== "";
+    case "runtime_notice":
+      return normalizeCommitText(event.body) !== "";
     case "tool":
       return true;
     case "session":
@@ -3542,6 +3562,12 @@ async function flushHeldText(buffer, emit, terminal, onError) {
   buffer.held = null;
   await safeEmit(emit, { type: "text", body, terminal }, onError);
 }
+async function flushHeldRuntimeNotice(buffer, emit, onError) {
+  if (buffer.held === null) return;
+  const body = buffer.held;
+  buffer.held = null;
+  await safeEmit(emit, { type: "runtime_notice", body }, onError);
+}
 function extractAssistantBlocks(msg) {
   const texts = [];
   const toolUses = [];
@@ -3663,6 +3689,9 @@ function sinkEmitter(sink, _onError) {
       case "thinking":
         if (sink.onThinking) await sink.onThinking({ text: event.text });
         return;
+      case "runtime_notice":
+        if (sink.onRuntimeNotice) await sink.onRuntimeNotice({ text: event.body });
+        return;
       // `session` / `result` are host/pump territory — never produced by the
       // transcript classification, so they have no callback to render onto.
       case "session":
@@ -3716,6 +3745,13 @@ var TurnPump = class {
         const text = truncate(trimmed, THINKING_TEXT_MAX_CHARS);
         const seq = opts.nextSeq();
         await opts.commit.reportThinking({ text, seq });
+      },
+      onRuntimeNotice: async (evt) => {
+        if (opts.signal.aborted) return;
+        const body = normalizeCommitText(evt.text);
+        if (!body) return;
+        const seq = opts.nextSeq();
+        await opts.commit.commitMessage({ body, kind: "runtime_notice", seq });
       }
     };
   }
@@ -4125,7 +4161,13 @@ async function* decodeSdkStream(iter, ctx) {
     ok = false;
     resultReason ??= "session_start_failed";
   }
-  await flushHeldText(buffer, emit, ok);
+  const heldFailure = !ok ? classifyRuntimeNoticeText(buffer.held) : null;
+  const terminalFailure = !ok ? decodeFailureReason(resultReason) : null;
+  if (heldFailure && terminalFailure && heldFailure.kind === terminalFailure.kind) {
+    await flushHeldRuntimeNotice(buffer, emit);
+  } else {
+    await flushHeldText(buffer, emit, ok);
+  }
   yield* drain(out);
   if (!ok && !resultReason && !sawResult) resultReason = "no_result";
   yield {
@@ -4647,22 +4689,92 @@ var BASE_CLAUDE_CODE_CONFORMANCE_FIXTURES = [
   {
     // CT558/CT592/CT731: a subscription cap. The SDK emits a rejected
     // `rate_limit_event` with a named subscription window; the terminal error result
-    // then classifies as the structured `usage_capped` reason. Partial narration
-    // lands as `progress`. The cap event is distinct from a provider 429 throttle.
+    // then classifies as the structured `usage_capped` reason. The held provider
+    // line independently classifies to the same kind, so it lands as a runtime
+    // notice rather than agent-authored progress.
     name: "subscription cap \u2192 usage_capped",
     request: makeRequest(),
     nativeStream: [
       init("s1"),
-      assistantText("Let me work on that."),
+      assistantText("You've hit your session limit \xB7 resets 8:20pm (UTC)"),
+      rateLimitEvent("rejected", void 0, "five_hour"),
+      resultError("error_during_execution")
+    ],
+    expected: [
+      sessionEvent("s1"),
+      {
+        type: "runtime_notice",
+        body: "You've hit your session limit \xB7 resets 8:20pm (UTC)"
+      },
+      { type: "result", ok: false, reason: "usage_capped" }
+    ]
+  },
+  {
+    // CT1336: a failed turn can still contain the agent's own partial narration.
+    // No text classification means no authorship rewrite, even though the
+    // structured result is the same usage cap.
+    name: "partial narration before subscription cap stays progress",
+    request: makeRequest(),
+    nativeStream: [
+      init("s1"),
+      assistantText("Partial work."),
       rateLimitEvent("rejected", void 0, "five_hour"),
       resultError("error_during_execution")
     ],
     expected: [
       sessionEvent("s1"),
-      { type: "text", body: "Let me work on that.", terminal: false },
+      { type: "text", body: "Partial work.", terminal: false },
       { type: "result", ok: false, reason: "usage_capped" }
     ]
   },
+  {
+    // CT1336: even recognized runtime prose stays authored when its kind
+    // disagrees with the structured terminal failure.
+    name: "cap notice before server failure stays progress",
+    request: makeRequest(),
+    nativeStream: [
+      init("s1"),
+      assistantText("You've hit your session limit \xB7 resets 8:20pm (UTC)"),
+      resultErrorFull("error_during_execution", {
+        is_error: true,
+        result: '529 {"type":"error","error":{"type":"overloaded_error"}}'
+      })
+    ],
+    expected: [
+      sessionEvent("s1"),
+      {
+        type: "text",
+        body: "You've hit your session limit \xB7 resets 8:20pm (UTC)",
+        terminal: false
+      },
+      { type: "result", ok: false, reason: "server_error" }
+    ]
+  },
+  {
+    // CT1336 review: `classifyErrorText` intentionally searches arbitrary error
+    // blobs for status tokens. Assistant narration needs the stricter whole-
+    // notice classifier, or an incidental count is de-authored when the turn
+    // later fails with the matching provider kind.
+    name: "incidental 5xx-shaped narration before server failure stays progress",
+    request: makeRequest(),
+    nativeStream: [
+      init("s1"),
+      assistantText("Indexed 500 files. Writing the summary next."),
+      resultErrorFull("error_during_execution", {
+        is_error: true,
+        result: '529 {"type":"error","error":{"type":"overloaded_error"}}'
+      })
+    ],
+    expected: [
+      sessionEvent("s1"),
+      {
+        type: "text",
+        body: "Indexed 500 files. Writing the summary next.",
+        terminal: false
+      },
+      { type: "result", ok: false, reason: "server_error" }
+    ]
+  },
   {
     // CT558/CT592: a cap signalled only by the result's
     // `terminal_reason:'blocking_limit'` (no discrete rate-limit event) still
@@ -8149,7 +8261,7 @@ function initialOutcome() {
     turnResolvedModel: void 0,
     turnResolvedConfig: void 0,
     turnMcpInventory: void 0,
-    eventCounts: { session: 0, text: 0, thinking: 0, tool: 0, result: 0 },
+    eventCounts: { session: 0, text: 0, runtime_notice: 0, thinking: 0, tool: 0, result: 0 },
     runtimeResultKind: null,
     contentBearingEvents: 0,
     latestSessionState: null,

package/dist/runtime.js

@@ -2547,6 +2547,12 @@ var turnEventSchema = z5.discriminatedUnion("type", [
   // `progress` row); `terminal: true` is the turn's closing reply (commits as
   // the `final` row).
   z5.object({ type: z5.literal("text"), body: z5.string(), terminal: z5.boolean() }),
+  // Runtime/provider-authored prose surfaced alongside a failed turn. Unlike
+  // `text`, this is not the agent's narration or reply: the pump persists it as
+  // `runtime_notice`, and the transcript renders it in the shared SystemNote
+  // voice. Adapters emit it only from positive runtime evidence; the web never
+  // classifies English strings.
+  z5.object({ type: z5.literal("runtime_notice"), body: z5.string() }),
   // A tool-activity transition. Maps `onToolActivity` — `toolUseId`→`id`,
   // `toolName`→`name` (already prefix-stripped: `cabane_read`, not
   // `mcp__cabane__cabane_read`), `summary` is the short card label. `phase`
@@ -2680,6 +2686,7 @@ var turnDiagnosticsSchema = z6.object({
   eventCounts: z6.object({
     session: z6.number().int().nonnegative(),
     text: z6.number().int().nonnegative(),
+    runtime_notice: z6.number().int().nonnegative(),
     thinking: z6.number().int().nonnegative(),
     tool: z6.number().int().nonnegative(),
     result: z6.number().int().nonnegative()
@@ -2778,6 +2785,17 @@ function classifyErrorText(text) {
   if (BARE_LIMIT.test(t)) return { kind: "usage_capped" };
   return null;
 }
+function classifyRuntimeNoticeText(text) {
+  if (!text) return null;
+  const normalized = text.trim().replace(/\s+/g, " ");
+  if (!normalized) return null;
+  if (/^you(?:'|’)ve hit your (?:session|weekly|usage) limit(?:\s*[·—-]\s*resets?\s+.+)?$/i.test(
+    normalized
+  )) {
+    return { kind: "usage_capped" };
+  }
+  return null;
+}
 var AUTH_PATTERNS = [
   /authentication[_ ]error/,
   /invalid[_ ]?(x-)?api[_ ]?key/,
@@ -2845,6 +2863,8 @@ function isContentBearingEvent(event) {
       return normalizeCommitText(event.body) !== "";
     case "thinking":
       return normalizeCommitText(event.text) !== "";
+    case "runtime_notice":
+      return normalizeCommitText(event.body) !== "";
     case "tool":
       return true;
     case "session":
@@ -3041,6 +3061,12 @@ async function flushHeldText(buffer, emit, terminal, onError) {
   buffer.held = null;
   await safeEmit(emit, { type: "text", body, terminal }, onError);
 }
+async function flushHeldRuntimeNotice(buffer, emit, onError) {
+  if (buffer.held === null) return;
+  const body = buffer.held;
+  buffer.held = null;
+  await safeEmit(emit, { type: "runtime_notice", body }, onError);
+}
 function extractAssistantBlocks(msg) {
   const texts = [];
   const toolUses = [];
@@ -3162,6 +3188,9 @@ function sinkEmitter(sink, _onError) {
       case "thinking":
         if (sink.onThinking) await sink.onThinking({ text: event.text });
         return;
+      case "runtime_notice":
+        if (sink.onRuntimeNotice) await sink.onRuntimeNotice({ text: event.body });
+        return;
       // `session` / `result` are host/pump territory — never produced by the
       // transcript classification, so they have no callback to render onto.
       case "session":
@@ -3215,6 +3244,13 @@ var TurnPump = class {
         const text = truncate(trimmed, THINKING_TEXT_MAX_CHARS);
         const seq = opts.nextSeq();
         await opts.commit.reportThinking({ text, seq });
+      },
+      onRuntimeNotice: async (evt) => {
+        if (opts.signal.aborted) return;
+        const body = normalizeCommitText(evt.text);
+        if (!body) return;
+        const seq = opts.nextSeq();
+        await opts.commit.commitMessage({ body, kind: "runtime_notice", seq });
       }
     };
   }
@@ -3624,7 +3660,13 @@ async function* decodeSdkStream(iter, ctx) {
     ok = false;
     resultReason ??= "session_start_failed";
   }
-  await flushHeldText(buffer, emit, ok);
+  const heldFailure = !ok ? classifyRuntimeNoticeText(buffer.held) : null;
+  const terminalFailure = !ok ? decodeFailureReason(resultReason) : null;
+  if (heldFailure && terminalFailure && heldFailure.kind === terminalFailure.kind) {
+    await flushHeldRuntimeNotice(buffer, emit);
+  } else {
+    await flushHeldText(buffer, emit, ok);
+  }
   yield* drain(out);
   if (!ok && !resultReason && !sawResult) resultReason = "no_result";
   yield {
@@ -4146,22 +4188,92 @@ var BASE_CLAUDE_CODE_CONFORMANCE_FIXTURES = [
   {
     // CT558/CT592/CT731: a subscription cap. The SDK emits a rejected
     // `rate_limit_event` with a named subscription window; the terminal error result
-    // then classifies as the structured `usage_capped` reason. Partial narration
-    // lands as `progress`. The cap event is distinct from a provider 429 throttle.
+    // then classifies as the structured `usage_capped` reason. The held provider
+    // line independently classifies to the same kind, so it lands as a runtime
+    // notice rather than agent-authored progress.
     name: "subscription cap \u2192 usage_capped",
     request: makeRequest(),
     nativeStream: [
       init("s1"),
-      assistantText("Let me work on that."),
+      assistantText("You've hit your session limit \xB7 resets 8:20pm (UTC)"),
+      rateLimitEvent("rejected", void 0, "five_hour"),
+      resultError("error_during_execution")
+    ],
+    expected: [
+      sessionEvent("s1"),
+      {
+        type: "runtime_notice",
+        body: "You've hit your session limit \xB7 resets 8:20pm (UTC)"
+      },
+      { type: "result", ok: false, reason: "usage_capped" }
+    ]
+  },
+  {
+    // CT1336: a failed turn can still contain the agent's own partial narration.
+    // No text classification means no authorship rewrite, even though the
+    // structured result is the same usage cap.
+    name: "partial narration before subscription cap stays progress",
+    request: makeRequest(),
+    nativeStream: [
+      init("s1"),
+      assistantText("Partial work."),
       rateLimitEvent("rejected", void 0, "five_hour"),
       resultError("error_during_execution")
     ],
     expected: [
       sessionEvent("s1"),
-      { type: "text", body: "Let me work on that.", terminal: false },
+      { type: "text", body: "Partial work.", terminal: false },
       { type: "result", ok: false, reason: "usage_capped" }
     ]
   },
+  {
+    // CT1336: even recognized runtime prose stays authored when its kind
+    // disagrees with the structured terminal failure.
+    name: "cap notice before server failure stays progress",
+    request: makeRequest(),
+    nativeStream: [
+      init("s1"),
+      assistantText("You've hit your session limit \xB7 resets 8:20pm (UTC)"),
+      resultErrorFull("error_during_execution", {
+        is_error: true,
+        result: '529 {"type":"error","error":{"type":"overloaded_error"}}'
+      })
+    ],
+    expected: [
+      sessionEvent("s1"),
+      {
+        type: "text",
+        body: "You've hit your session limit \xB7 resets 8:20pm (UTC)",
+        terminal: false
+      },
+      { type: "result", ok: false, reason: "server_error" }
+    ]
+  },
+  {
+    // CT1336 review: `classifyErrorText` intentionally searches arbitrary error
+    // blobs for status tokens. Assistant narration needs the stricter whole-
+    // notice classifier, or an incidental count is de-authored when the turn
+    // later fails with the matching provider kind.
+    name: "incidental 5xx-shaped narration before server failure stays progress",
+    request: makeRequest(),
+    nativeStream: [
+      init("s1"),
+      assistantText("Indexed 500 files. Writing the summary next."),
+      resultErrorFull("error_during_execution", {
+        is_error: true,
+        result: '529 {"type":"error","error":{"type":"overloaded_error"}}'
+      })
+    ],
+    expected: [
+      sessionEvent("s1"),
+      {
+        type: "text",
+        body: "Indexed 500 files. Writing the summary next.",
+        terminal: false
+      },
+      { type: "result", ok: false, reason: "server_error" }
+    ]
+  },
   {
     // CT558/CT592: a cap signalled only by the result's
     // `terminal_reason:'blocking_limit'` (no discrete rate-limit event) still
@@ -7648,7 +7760,7 @@ function initialOutcome() {
     turnResolvedModel: void 0,
     turnResolvedConfig: void 0,
     turnMcpInventory: void 0,
-    eventCounts: { session: 0, text: 0, thinking: 0, tool: 0, result: 0 },
+    eventCounts: { session: 0, text: 0, runtime_notice: 0, thinking: 0, tool: 0, result: 0 },
     runtimeResultKind: null,
     contentBearingEvents: 0,
     latestSessionState: null,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cabane/companion",
-  "version": "0.6.73",
+  "version": "0.6.75",
   "type": "module",
   "description": "The Cabane Companion (headless): connect a coding agent on your machine to your Cabane workspace as a responder — drive work against your own codebase, files, and MCP servers without putting any of it in Cabane.",
   "license": "UNLICENSED",