@cabane/companion 0.6.143 → 0.6.144

package/dist/cli.js

@@ -1816,14 +1816,11 @@ var init_turn_event = __esm({
       // claude-code/codex it's the runtime's CUMULATIVE total summed across every model
       // request in the agentic turn, so it grows with the tool-call count and is NOT
       // "how full is the window." `contextTokens` is the distinct CONTEXT-OCCUPANCY
-      // read: the input the model saw on its FINAL request of the turn (uncached +
-      // cache, since cached tokens still occupy the window) — the number the composer
-      // gauge wants. `contextWindow` is the model's true window in tokens when the
-      // runtime reports it (claude-code's SDK does, per model) — a real denominator so
-      // the gauge can show a fraction. Both optional: a runtime that can't source a
-      // clean final-request figure (codex's cumulative-only usage) omits `contextTokens`
-      // and the gauge falls back to the raw count; `contextWindow` falls back to the
-      // model catalog.
+      // read and `contextWindow` the window it is measured against, both as the
+      // harness reported them (claude-code's `getContextUsage`, or its final
+      // request's input and `modelUsage` window when that call gives nothing). A
+      // runtime that can't source them (codex's cumulative-only usage) omits both,
+      // and no surface shows a reading for that turn.
       //
       // CT601: two more optional carry-homes on the terminal result, alongside
       // `usage`. `resolvedModel` is the CONCRETE model the runtime actually ran —
@@ -1846,7 +1843,25 @@ var init_turn_event = __esm({
           cacheReadTokens: z6.number().optional(),
           cacheCreationTokens: z6.number().optional(),
           contextTokens: z6.number().optional(),
-          contextWindow: z6.number().optional()
+          contextWindow: z6.number().optional(),
+          // CT1751: the rest of the harness's own context reading, when it gave
+          // one: where it will compact, its `/context` breakdown, and every
+          // compaction it reported during the turn. Claude Code only.
+          contextCompactAt: z6.number().optional(),
+          contextCategories: z6.array(
+            z6.object({
+              name: z6.string(),
+              tokens: z6.number(),
+              kind: z6.enum(["used", "free", "buffer", "deferred"])
+            })
+          ).optional(),
+          contextCompactions: z6.array(
+            z6.object({
+              trigger: z6.enum(["manual", "auto"]),
+              preTokens: z6.number(),
+              postTokens: z6.number().optional()
+            })
+          ).optional()
         }).optional(),
         resolvedModel: z6.string().optional(),
         resolvedConfig: z6.object({
@@ -2902,6 +2917,7 @@ async function* decodeSdkStream(iter, ctx) {
   let resultReason;
   let sawResult = false;
   let usage;
+  const compactions = [];
   let lastRequestContextTokens;
   let resolvedModel;
   const signals = { sawRejectedLimit: false };
@@ -2927,6 +2943,9 @@ async function* decodeSdkStream(iter, ctx) {
             ...ctx.degraded ? { degraded: true } : {}
           };
         }
+      } else if (msg.type === "system" && msg.subtype === "compact_boundary") {
+        const compaction = readCompaction(msg);
+        if (compaction) compactions.push(compaction);
       } else if (!workspaceProven && (msg.type === "assistant" || msg.type === "user" || msg.type === "result")) {
         if (preInitToolActivity(msg) && ctx.startEvidence) ctx.startEvidence.toolActivity = true;
         const failure = classifyPreInitFrame(msg, signals);
@@ -2971,6 +2990,9 @@ async function* decodeSdkStream(iter, ctx) {
             usage.contextTokens = lastRequestContextTokens;
           const window = readContextWindow(msg, resolvedModel);
           if (window !== void 0) usage.contextWindow = window;
+          if (compactions.length > 0) usage.contextCompactions = compactions;
+          const reading = await readHarnessContext(ctx.readContextUsage);
+          if (reading) Object.assign(usage, reading);
         }
         const isError = msg.is_error === true;
         if (msg.subtype === "success" && !isError) {
@@ -3121,6 +3143,56 @@ function readRequestContextTokens(msg) {
   const num = (v) => typeof v === "number" && Number.isFinite(v) ? v : 0;
   return num(usage.input_tokens) + num(usage.cache_read_input_tokens) + num(usage.cache_creation_input_tokens);
 }
+async function readHarnessContext(read) {
+  if (!read) return void 0;
+  let timer;
+  const timeout = new Promise((resolve2) => {
+    timer = setTimeout(() => resolve2(void 0), CONTEXT_READING_TIMEOUT_MS);
+  });
+  try {
+    return parseContextUsage(await Promise.race([read(), timeout]));
+  } catch {
+    return void 0;
+  } finally {
+    clearTimeout(timer);
+  }
+}
+function parseContextUsage(response) {
+  const r = response;
+  if (!r || typeof r !== "object") return void 0;
+  const count = (v) => typeof v === "number" && Number.isFinite(v) && v >= 0 ? Math.round(v) : void 0;
+  const total = count(r.totalTokens);
+  const window = count(r.maxTokens);
+  if (total === void 0 || window === void 0 || window === 0) return void 0;
+  const categories = [];
+  for (const c of Array.isArray(r.categories) ? r.categories : []) {
+    const row = c;
+    const tokens = count(row?.tokens);
+    if (typeof row?.name === "string" && tokens !== void 0 && typeof row.kind === "string" && CATEGORY_KINDS.has(row.kind)) {
+      categories.push({ name: row.name, tokens, kind: row.kind });
+    }
+  }
+  const buffer = categories.filter((c) => c.kind === "buffer").reduce((sum, c) => sum + c.tokens, 0);
+  return {
+    contextTokens: total,
+    contextWindow: window,
+    ...buffer > 0 && buffer < window ? { contextCompactAt: window - buffer } : {},
+    ...categories.length > 0 ? { contextCategories: categories } : {}
+  };
+}
+function readCompaction(msg) {
+  const meta = msg.compact_metadata;
+  const count = (v) => typeof v === "number" && Number.isFinite(v) && v >= 0 ? Math.round(v) : void 0;
+  const preTokens = count(meta?.pre_tokens);
+  if (!meta || preTokens === void 0) return void 0;
+  if (meta.trigger !== "manual" && meta.trigger !== "auto") return void 0;
+  const postTokens = count(meta.post_tokens);
+  return {
+    trigger: meta.trigger,
+    preTokens,
+    ...postTokens !== void 0 ? { postTokens } : {}
+  };
+}
 function readContextWindow(msg, resolvedModel) {
   const modelUsage = msg.modelUsage;
   if (!modelUsage || typeof modelUsage !== "object") return void 0;
@@ -3136,7 +3208,7 @@ function readContextWindow(msg, resolvedModel) {
   }
   return max;
 }
-var HARNESS_PLACEHOLDERS;
+var HARNESS_PLACEHOLDERS, CONTEXT_READING_TIMEOUT_MS, CATEGORY_KINDS;
 var init_loop = __esm({
   "packages/agent-runtime/src/claude-code/loop.ts"() {
     "use strict";
@@ -3144,11 +3216,13 @@ var init_loop = __esm({
     init_failure();
     init_session();
     HARNESS_PLACEHOLDERS = /* @__PURE__ */ new Set(["No response requested.", "(no content)"]);
+    CONTEXT_READING_TIMEOUT_MS = 2e3;
+    CATEGORY_KINDS = /* @__PURE__ */ new Set(["used", "free", "buffer", "deferred"]);
   }
 });
 
 // packages/agent-runtime/src/claude-code/prompt-input.ts
-function buildQueryPrompt(req) {
+function buildQueryPrompt(req, openUntil) {
   const content = req.content;
   return (async function* () {
     yield {
@@ -3157,6 +3231,7 @@ function buildQueryPrompt(req) {
       message: { role: "user", content },
       parent_tool_use_id: null
     };
+    if (openUntil) await openUntil;
   })();
 }
 var init_prompt_input = __esm({
@@ -3176,19 +3251,34 @@ function attachAbortController(options, signal) {
   else signal.addEventListener("abort", () => sdkAbort.abort(), { once: true });
   options.abortController = sdkAbort;
 }
-function sdkEvents(queryFn, req, signal, built, degraded, startEvidence) {
+async function* sdkEvents(queryFn, req, signal, built, degraded, startEvidence) {
   attachAbortController(built.options, signal);
-  const prompt = buildQueryPrompt(req);
-  const iter = {
-    [Symbol.asyncIterator]: () => queryFn({ prompt, options: built.options })[Symbol.asyncIterator]()
+  let release = () => {
   };
-  return decodeSdkStream(iter, {
-    signal,
-    cwd: req.local.cwd,
-    resumedSessionId: built.resume,
-    degraded,
-    ...startEvidence ? { startEvidence } : {}
+  const held = new Promise((resolve2) => {
+    release = resolve2;
   });
+  const prompt = buildQueryPrompt(req, held);
+  let active;
+  const iter = {
+    [Symbol.asyncIterator]: () => {
+      const q = queryFn({ prompt, options: built.options });
+      active = q;
+      return q[Symbol.asyncIterator]();
+    }
+  };
+  try {
+    yield* decodeSdkStream(iter, {
+      signal,
+      readContextUsage: async () => typeof active?.getContextUsage === "function" ? active.getContextUsage({ detail: "summary" }) : void 0,
+      cwd: req.local.cwd,
+      resumedSessionId: built.resume,
+      degraded,
+      ...startEvidence ? { startEvidence } : {}
+    });
+  } finally {
+    release();
+  }
 }
 async function freshRequest(req, reason, hooks) {
   const recomposed = hooks?.fresh ? await hooks.fresh(reason) : null;

package/dist/runtime.js

@@ -3309,14 +3309,11 @@ var turnEventSchema = z6.discriminatedUnion("type", [
   // claude-code/codex it's the runtime's CUMULATIVE total summed across every model
   // request in the agentic turn, so it grows with the tool-call count and is NOT
   // "how full is the window." `contextTokens` is the distinct CONTEXT-OCCUPANCY
-  // read: the input the model saw on its FINAL request of the turn (uncached +
-  // cache, since cached tokens still occupy the window) — the number the composer
-  // gauge wants. `contextWindow` is the model's true window in tokens when the
-  // runtime reports it (claude-code's SDK does, per model) — a real denominator so
-  // the gauge can show a fraction. Both optional: a runtime that can't source a
-  // clean final-request figure (codex's cumulative-only usage) omits `contextTokens`
-  // and the gauge falls back to the raw count; `contextWindow` falls back to the
-  // model catalog.
+  // read and `contextWindow` the window it is measured against, both as the
+  // harness reported them (claude-code's `getContextUsage`, or its final
+  // request's input and `modelUsage` window when that call gives nothing). A
+  // runtime that can't source them (codex's cumulative-only usage) omits both,
+  // and no surface shows a reading for that turn.
   //
   // CT601: two more optional carry-homes on the terminal result, alongside
   // `usage`. `resolvedModel` is the CONCRETE model the runtime actually ran —
@@ -3339,7 +3336,25 @@ var turnEventSchema = z6.discriminatedUnion("type", [
       cacheReadTokens: z6.number().optional(),
       cacheCreationTokens: z6.number().optional(),
       contextTokens: z6.number().optional(),
-      contextWindow: z6.number().optional()
+      contextWindow: z6.number().optional(),
+      // CT1751: the rest of the harness's own context reading, when it gave
+      // one: where it will compact, its `/context` breakdown, and every
+      // compaction it reported during the turn. Claude Code only.
+      contextCompactAt: z6.number().optional(),
+      contextCategories: z6.array(
+        z6.object({
+          name: z6.string(),
+          tokens: z6.number(),
+          kind: z6.enum(["used", "free", "buffer", "deferred"])
+        })
+      ).optional(),
+      contextCompactions: z6.array(
+        z6.object({
+          trigger: z6.enum(["manual", "auto"]),
+          preTokens: z6.number(),
+          postTokens: z6.number().optional()
+        })
+      ).optional()
     }).optional(),
     resolvedModel: z6.string().optional(),
     resolvedConfig: z6.object({
@@ -4294,6 +4309,7 @@ async function* decodeSdkStream(iter, ctx) {
   let resultReason;
   let sawResult = false;
   let usage;
+  const compactions = [];
   let lastRequestContextTokens;
   let resolvedModel;
   const signals = { sawRejectedLimit: false };
@@ -4319,6 +4335,9 @@ async function* decodeSdkStream(iter, ctx) {
             ...ctx.degraded ? { degraded: true } : {}
           };
         }
+      } else if (msg.type === "system" && msg.subtype === "compact_boundary") {
+        const compaction = readCompaction(msg);
+        if (compaction) compactions.push(compaction);
       } else if (!workspaceProven && (msg.type === "assistant" || msg.type === "user" || msg.type === "result")) {
         if (preInitToolActivity(msg) && ctx.startEvidence) ctx.startEvidence.toolActivity = true;
         const failure = classifyPreInitFrame(msg, signals);
@@ -4363,6 +4382,9 @@ async function* decodeSdkStream(iter, ctx) {
             usage.contextTokens = lastRequestContextTokens;
           const window = readContextWindow(msg, resolvedModel);
           if (window !== void 0) usage.contextWindow = window;
+          if (compactions.length > 0) usage.contextCompactions = compactions;
+          const reading = await readHarnessContext(ctx.readContextUsage);
+          if (reading) Object.assign(usage, reading);
         }
         const isError = msg.is_error === true;
         if (msg.subtype === "success" && !isError) {
@@ -4514,6 +4536,58 @@ function readRequestContextTokens(msg) {
   const num = (v) => typeof v === "number" && Number.isFinite(v) ? v : 0;
   return num(usage.input_tokens) + num(usage.cache_read_input_tokens) + num(usage.cache_creation_input_tokens);
 }
+var CONTEXT_READING_TIMEOUT_MS = 2e3;
+async function readHarnessContext(read) {
+  if (!read) return void 0;
+  let timer;
+  const timeout = new Promise((resolve2) => {
+    timer = setTimeout(() => resolve2(void 0), CONTEXT_READING_TIMEOUT_MS);
+  });
+  try {
+    return parseContextUsage(await Promise.race([read(), timeout]));
+  } catch {
+    return void 0;
+  } finally {
+    clearTimeout(timer);
+  }
+}
+var CATEGORY_KINDS = /* @__PURE__ */ new Set(["used", "free", "buffer", "deferred"]);
+function parseContextUsage(response) {
+  const r = response;
+  if (!r || typeof r !== "object") return void 0;
+  const count = (v) => typeof v === "number" && Number.isFinite(v) && v >= 0 ? Math.round(v) : void 0;
+  const total = count(r.totalTokens);
+  const window = count(r.maxTokens);
+  if (total === void 0 || window === void 0 || window === 0) return void 0;
+  const categories = [];
+  for (const c of Array.isArray(r.categories) ? r.categories : []) {
+    const row = c;
+    const tokens = count(row?.tokens);
+    if (typeof row?.name === "string" && tokens !== void 0 && typeof row.kind === "string" && CATEGORY_KINDS.has(row.kind)) {
+      categories.push({ name: row.name, tokens, kind: row.kind });
+    }
+  }
+  const buffer = categories.filter((c) => c.kind === "buffer").reduce((sum, c) => sum + c.tokens, 0);
+  return {
+    contextTokens: total,
+    contextWindow: window,
+    ...buffer > 0 && buffer < window ? { contextCompactAt: window - buffer } : {},
+    ...categories.length > 0 ? { contextCategories: categories } : {}
+  };
+}
+function readCompaction(msg) {
+  const meta = msg.compact_metadata;
+  const count = (v) => typeof v === "number" && Number.isFinite(v) && v >= 0 ? Math.round(v) : void 0;
+  const preTokens = count(meta?.pre_tokens);
+  if (!meta || preTokens === void 0) return void 0;
+  if (meta.trigger !== "manual" && meta.trigger !== "auto") return void 0;
+  const postTokens = count(meta.post_tokens);
+  return {
+    trigger: meta.trigger,
+    preTokens,
+    ...postTokens !== void 0 ? { postTokens } : {}
+  };
+}
 function readContextWindow(msg, resolvedModel) {
   const modelUsage = msg.modelUsage;
   if (!modelUsage || typeof modelUsage !== "object") return void 0;
@@ -4531,7 +4605,7 @@ function readContextWindow(msg, resolvedModel) {
 }
 
 // packages/agent-runtime/src/claude-code/prompt-input.ts
-function buildQueryPrompt(req) {
+function buildQueryPrompt(req, openUntil) {
   const content = req.content;
   return (async function* () {
     yield {
@@ -4540,6 +4614,7 @@ function buildQueryPrompt(req) {
       message: { role: "user", content },
       parent_tool_use_id: null
     };
+    if (openUntil) await openUntil;
   })();
 }
 
@@ -4551,19 +4626,34 @@ function attachAbortController(options, signal) {
   else signal.addEventListener("abort", () => sdkAbort.abort(), { once: true });
   options.abortController = sdkAbort;
 }
-function sdkEvents(queryFn, req, signal, built, degraded, startEvidence) {
+async function* sdkEvents(queryFn, req, signal, built, degraded, startEvidence) {
   attachAbortController(built.options, signal);
-  const prompt = buildQueryPrompt(req);
-  const iter = {
-    [Symbol.asyncIterator]: () => queryFn({ prompt, options: built.options })[Symbol.asyncIterator]()
+  let release = () => {
   };
-  return decodeSdkStream(iter, {
-    signal,
-    cwd: req.local.cwd,
-    resumedSessionId: built.resume,
-    degraded,
-    ...startEvidence ? { startEvidence } : {}
+  const held = new Promise((resolve2) => {
+    release = resolve2;
   });
+  const prompt = buildQueryPrompt(req, held);
+  let active;
+  const iter = {
+    [Symbol.asyncIterator]: () => {
+      const q = queryFn({ prompt, options: built.options });
+      active = q;
+      return q[Symbol.asyncIterator]();
+    }
+  };
+  try {
+    yield* decodeSdkStream(iter, {
+      signal,
+      readContextUsage: async () => typeof active?.getContextUsage === "function" ? active.getContextUsage({ detail: "summary" }) : void 0,
+      cwd: req.local.cwd,
+      resumedSessionId: built.resume,
+      degraded,
+      ...startEvidence ? { startEvidence } : {}
+    });
+  } finally {
+    release();
+  }
 }
 async function freshRequest(req, reason, hooks) {
   const recomposed = hooks?.fresh ? await hooks.fresh(reason) : null;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cabane/companion",
-  "version": "0.6.143",
+  "version": "0.6.144",
   "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",