@gleapai/kai-bridge 0.10.1 → 0.10.2

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "@gleapai/kai-bridge",
-  "version": "0.10.1",
+  "version": "0.10.2",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@gleapai/kai-bridge",
-      "version": "0.10.1",
+      "version": "0.10.2",
       "hasInstallScript": true,
       "license": "MIT",
       "dependencies": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gleapai/kai-bridge",
-  "version": "0.10.1",
+  "version": "0.10.2",
   "description": "Kai Code Bridge runs Kai Code on your computer or server with your own coding subscriptions and local previews.",
   "type": "module",
   "license": "MIT",

package/runner/lib/acp/harnesses.mjs

@@ -302,7 +302,12 @@ export const HARNESSES = {
      */
     supportsPromptSuggestions: (ctx) => !ctx.isPlanMode && !ctx.isArtifactWriter && claudeAdapterForwardsSuggestions(ctx),
     /** ACP session modes to try, in order (`session/set_mode`) — the adapter's own ids. */
-    sessionModePreference: (ctx) => (ctx.isPlanMode ? ["plan"] : ctx.isArtifactWriter ? ["dontAsk", "plan"] : ["bypassPermissions", "acceptEdits", "default"]),
+    // Artifact writers: `dontAsk` where the adapter still offers it, else
+    // `default` (the runner's permissionPolicy answers each request: `.kai/**`
+    // writes allowed, repo mutations denied). Never `plan` — it declines the
+    // `.kai/` write too, and claude-agent-acp ≥0.71 no longer offers
+    // `dontAsk` (mirrors gleap_code_analyzer#70).
+    sessionModePreference: (ctx) => (ctx.isPlanMode ? ["plan"] : ctx.isArtifactWriter ? ["dontAsk", "default", "plan"] : ["bypassPermissions", "acceptEdits", "default"]),
     /** A prior turn's transcript on disk is what makes `resume` viable. */
     hasResumableSession: (ctx) => !!findClaudeTranscript(ctx.configDir, ctx.workDir, ctx.resumeSessionId),
     /**

package/runner/lib/acp/transcripts.mjs

@@ -194,9 +194,18 @@ export function readClaudePriorCostUsd({ configDir, cwd, sessionId }) {
 /**
  * Codex rollout `token_count` events report `last_token_usage` (the
  * most recent request) and a running `total_token_usage`. The per-
- * request rows are the `last_token_usage` deltas; `cached_input_tokens`
- * is INCLUDED in `input_tokens` (OpenAI convention — the tracker treats
- * `inputTokens` as cache-inclusive too, so pass it through unchanged).
+ * request rows are the `last_token_usage` deltas.
+ *
+ * OpenAI conventions (verified on a live GPT-6 rollout 2026-09-23:
+ * `total_tokens === input_tokens + output_tokens` on every request):
+ *  - `cached_input_tokens` is a SUBSET of `input_tokens`; the tracker's
+ *    `inputTokens` is FRESH input (the Anthropic shape), so the row carries
+ *    `input_tokens - cached_input_tokens`. Passing the inclusive figure
+ *    counted every cached token twice — reported input ~1.9x the real
+ *    volume, cache rate ~46% instead of ~85%, and the double-counted
+ *    cached volume billed at the FULL input rate.
+ *  - `reasoning_output_tokens` is a SUBSET of `output_tokens`; adding it
+ *    double-billed reasoning.
  */
 export function parseCodexRollout(text, { sinceTs, model } = {}) {
   const since = sinceTs ? new Date(sinceTs).getTime() : 0;
@@ -226,13 +235,14 @@ export function parseCodexRollout(text, { sinceTs, model } = {}) {
     }
     const input = num(last.input_tokens);
     if (input <= 0 && num(last.output_tokens) <= 0) continue;
+    const cached = Math.min(num(last.cached_input_tokens), input);
     rows.push({
       id: `${ts || rows.length}`,
       model: activeModel,
-      inputTokens: input,
-      cachedInputTokens: num(last.cached_input_tokens),
+      inputTokens: input - cached,
+      cachedInputTokens: cached,
       cacheWriteInputTokens: 0,
-      outputTokens: num(last.output_tokens) + num(last.reasoning_output_tokens),
+      outputTokens: num(last.output_tokens),
       sidechain: false,
       timestamp: ts || null,
     });