@kontextmind/kxm 0.7.152 → 0.7.154

package/.claude-plugin/marketplace.json

@@ -11,7 +11,7 @@
       "name": "kxm",
       "source": "./plugins/kxm",
       "description": "Durable workflows, peer agents, and kxm tui",
-      "version": "0.7.152",
+      "version": "0.7.154",
       "category": "development",
       "tags": ["kxm", "multi-agent", "workflows", "mcp"]
     }

package/CHANGELOG.md

@@ -6,20 +6,6 @@ All notable user-facing changes are documented here. The project follows [Semant
 
 ### Added
 
-- **An opted-in role continues a failed attempt on the next admitted route.**
-  `policy.fallback.onError` lists `rate_limit`, `transport`,
-  `provider_unavailable`, and `context_overflow`. An empty or absent list is
-  unchanged: the attempt fails on the first route. The walk is bounded by
-  `maxSwitches` (default 1), skips a critic-vendor collision when the role
-  requires vendor independence, and records `routing.route_switched` plus a
-  `route_switch` log line. A bounded redacted transcript is appended to the
-  next prompt. `revert: never` keeps the successful route for later steps in
-  the same run. `kxm routing report` and `kxm improve report` show the switch.
-  Cancellation, policy refusal, authentication, an unhosted model, a gate
-  failure, tool policy, and admission errors do not walk. See
-  [role fallback](docs/reference/config-reference.md#kxmrolesroleyaml-kxmrolev2)
-  and [mid-attempt fallback](docs/reference/harness-routing.md#mid-attempt-fallback).
-
 - **A live agent step uses a configurable one-shot timeout, and a cancelling run recovers when its child has already exited.**
   The bound is the step `timeoutMs`, or the project `limits.agentStepTimeoutMs`
   when the step omits it (minimum 60 seconds, default one hour). A wider step
@@ -154,38 +140,30 @@ All notable user-facing changes are documented here. The project follows [Semant
 
 - **Playwright uses Obscura by default.** `resolveBrowserCdpEndpoint()` returns `OBSCURA_CDP_URL` or `http://127.0.0.1:${OBSCURA_PORT:-9222}`. `KXM_BROWSER=steel` still returns the Steel session CDP URL. `node scripts/obscura.mjs` downloads pinned Obscura v0.2.3 and serves it with `--allow-private-network`. `npm run e2e` runs the Playwright smoke test in `test/e2e/` over CDP and does not run `playwright install`. Steel remains the path for human takeover, MFA, and the live session viewer. See [ADR-0005](docs/adr/ADR-0005-obscura-default-playwright.md) and [How do I connect Playwright to Obscura?](docs/kb/how-to-connect-playwright-to-obscura.md).
 
-### Changed
+- **An opted-in role continues a failed attempt on the next admitted route.**
+  `policy.fallback.onError` lists `rate_limit`, `transport`,
+  `provider_unavailable`, and `context_overflow`. An empty or absent list is
+  unchanged: the attempt fails on the first route. The walk is bounded by
+  `maxSwitches` (default 1), skips a critic-vendor collision when the role
+  requires vendor independence, and records `routing.route_switched` plus a
+  `route_switch` log line. A bounded redacted transcript is appended to the
+  next prompt. `revert: never` keeps the successful route for later steps in
+  the same run. `kxm routing report` and `kxm improve report` show the switch.
+  Cancellation, policy refusal, authentication, an unhosted model, a gate
+  failure, tool policy, and admission errors do not walk. See
+  [role fallback](docs/reference/config-reference.md#kxmrolesroleyaml-kxmrolev2)
+  and [mid-attempt fallback](docs/reference/harness-routing.md#mid-attempt-fallback).
 
-- **Pull-request CI runs the unit suite on Linux Node 24, and `CI / required` is the aggregate check.**
-  Docs and plan markdown skip the code jobs. `engine.test.ts` is split by
-  test name. `permission.test.ts` and `runtime.test.ts` run one file at a
-  time, and every other unit file runs in one light lane. Pushes to `main`
-  still run `validate:pr` on Linux and Windows for Node 22.19.0 and Node 24.
-  A pull request that touches path, process, shell, spawn, package, lockfile,
-  or workflow files also runs the unit lanes on Windows. Playwright stays on
-  Obscura. See [CI and release](docs/contributing/ci-and-release.md).
+### Changed
 
-- **Docs match the 2026-09-27 Steel and machine-account infrastructure.**
-  Steel (`steel.kontextmind.com`, alias `steel.theneuro.me`) is reached only
-  through Caddy on `kxmd-proxy` (VM 230) and Authentik forward auth. Direct
-  LAN, tailnet, and host-forward access is blocked. Sessions return
-  `websocketUrl` `wss://steel.kontextmind.com/` (previously
-  `ws://steel-browser/`). The CDP path is `/v1/devtools` with an
-  `Authorization` header. Allowed groups are `steel-users`, `kxmd-users`,
-  `kxmd-admins`, and `kxmd-owners`. `STEEL_API_URL` defaults to
-  `https://steel.kontextmind.com`. `STEEL_API_KEY` is deprecated and is not
-  enforced by Steel or Caddy. Migrate to `STEEL_AUTH_HEADER`, then
-  `STEEL_AUTH_BASIC`, then `STEEL_AUTH_USER` and `STEEL_AUTH_TOKEN`; those
-  override the key. The `svc-steel` credential is 1Password vault
-  `kontextmind`, item `Steel (svc-steel)`, field `basic_auth`, read with
-  `op read` and never written to disk. Steel requires `kxm` 0.7.135 or
-  newer. Playwright stays on Obscura ([ADR-0005](docs/adr/ADR-0005-obscura-default-playwright.md)).
-  The Proxmox boot order and VM names are in
-  [Deploy KXM](docs/operations/deploy.md#boot-the-kxmd-proxmox-host).
-  Machine accounts follow
-  [ADR-0006](docs/adr/ADR-0006-machine-account-names.md). The DOKS deployment
-  record is superseded by
-  [ADR-0007](docs/adr/ADR-0007-steel-caddy-authentik.md).
+- **Windows pull-request unit lanes finish inside the job limit.** The light
+  lane on `windows-latest` was cancelled at 20 minutes on #361, #362, and
+  #363 after it stopped reporting tests. `package-install.test.ts` (171s on
+  that lane, 67s on Linux) now runs with the serial files. The other light
+  files are split into `light-1` and `light-2`. POSIX `#!/bin/sh` shims in
+  the land tests are skipped on Windows. `CI / required` still fails when a
+  lane fails or is cancelled. See
+  [CI and release](docs/contributing/ci-and-release.md).
 
 - **Workforce ids use one convention, and old ids still resolve.**
   Role ids stay `planner`, `writer`, `reviewer-arch`, and `reviewer-cli`.
@@ -511,6 +489,50 @@ All notable user-facing changes are documented here. The project follows [Semant
   persisting a projection write, so a read cannot mutate run state or surface a
   false `run_projection_divergent`.
 
+- **Steel clients authenticate to Authentik with `Authorization: Basic`.**
+  `STEEL_AUTH_BASIC`, or `STEEL_AUTH_USER` and `STEEL_AUTH_TOKEN`, set that
+  header on Steel HTTP requests and on the CDP options from `formatCDPConnect()`.
+  `KXM_BROWSER=steel` passes those headers through `connectBrowserOverCdp()`
+  into `chromium.connectOverCDP`. Obscura stays the default and sends no Steel
+  headers. `STEEL_AUTH_HEADER` overrides the value. The CDP URL omits the credential when
+  those variables are set. A 302 to the identity provider fails closed and does
+  not follow the login redirect. `STEEL_API_KEY` still sends the legacy
+  `x-steel-api-key` header and `apiKey` query parameter for the temporary proxy
+  shim, and warns once. See
+  [Browser automation](docs/guides/browser-automation.md).
+
+- **Pull-request CI runs the unit suite on Linux Node 24, and `CI / required` is the aggregate check.**
+  Docs and plan markdown skip the code jobs. `engine.test.ts` is split by
+  test name. `permission.test.ts`, `runtime.test.ts`, and
+  `package-install.test.ts` run one file at a time, and every other unit file
+  runs in two light shards. Pushes to `main`
+  still run `validate:pr` on Linux and Windows for Node 22.19.0 and Node 24.
+  A pull request that touches path, process, shell, spawn, package, lockfile,
+  or workflow files also runs the unit lanes on Windows. Playwright stays on
+  Obscura. See [CI and release](docs/contributing/ci-and-release.md).
+
+- **Docs match the 2026-09-27 Steel and machine-account infrastructure.**
+  Steel (`steel.kontextmind.com`, alias `steel.theneuro.me`) is reached only
+  through Caddy on `kxmd-proxy` (VM 230) and Authentik forward auth. Direct
+  LAN, tailnet, and host-forward access is blocked. Sessions return
+  `websocketUrl` `wss://steel.kontextmind.com/` (previously
+  `ws://steel-browser/`). The CDP path is `/v1/devtools` with an
+  `Authorization` header. Allowed groups are `steel-users`, `kxmd-users`,
+  `kxmd-admins`, and `kxmd-owners`. `STEEL_API_URL` defaults to
+  `https://steel.kontextmind.com`. `STEEL_API_KEY` is deprecated and is not
+  enforced by Steel or Caddy. Migrate to `STEEL_AUTH_HEADER`, then
+  `STEEL_AUTH_BASIC`, then `STEEL_AUTH_USER` and `STEEL_AUTH_TOKEN`; those
+  override the key. The `svc-steel` credential is 1Password vault
+  `kontextmind`, item `Steel (svc-steel)`, field `basic_auth`, read with
+  `op read` and never written to disk. Steel requires `kxm` 0.7.135 or
+  newer. Playwright stays on Obscura ([ADR-0005](docs/adr/ADR-0005-obscura-default-playwright.md)).
+  The Proxmox boot order and VM names are in
+  [Deploy KXM](docs/operations/deploy.md#boot-the-kxmd-proxmox-host).
+  Machine accounts follow
+  [ADR-0006](docs/adr/ADR-0006-machine-account-names.md). The DOKS deployment
+  record is superseded by
+  [ADR-0007](docs/adr/ADR-0007-steel-caddy-authentik.md).
+
 - **Steel clients authenticate to Authentik with `Authorization: Basic`.**
   `STEEL_AUTH_BASIC`, or `STEEL_AUTH_USER` and `STEEL_AUTH_TOKEN`, set that
   header on Steel HTTP requests and on the CDP options from `formatCDPConnect()`.
@@ -542,9 +564,9 @@ All notable user-facing changes are documented here. The project follows [Semant
 
 ### Fixed
 
-- **Dry-run ship status no longer rewrites `.git/index`.** `git status` refreshes the index under an optional lock. The ship-status read passes `--no-optional-locks`, so a dry run leaves the checkout untouched.
+- **A one-shot outcome is the JSON object that ends the reply.** One forward scan tracks strings, escapes, and brace depth. The newest top-level object, opened at depth 0, must close at the end of the trimmed text (one closing code fence and trailing whitespace allowed). `JSON.parse` runs once on that slice. The `outcome` must be a string in the step's declared values. Braces inside a string do not count as structure, so a summary may contain any number of `{` characters. A bare object, a fenced object, and prose followed by an object that ends the reply settle when the outcome is allowed. Prose after the object, a reply with no object, a disallowed outcome, a stray `"` in the prose, and an inner object at the end of a truncated outer object settle `failed`.
 
-- **A committed checkout counts as authored work, and a one-shot outcome must be a standalone JSON object.** The authoring witness includes `HEAD` with porcelain status and both diffs, so a write that commits its edits is `changed` and can stay `passed`. A `rev-parse` failure keeps that empty head term only when the repository has no commits; any other git failure is `unwitnessed`. A one-shot outcome is accepted when the whole reply is one JSON object, or when that object stands alone on the final line. A closing code fence around the final object is allowed. An object followed by prose, a truncated reply, and an ambiguous tail settle `failed`.
+- **A committed checkout counts as authored work.** The authoring witness includes `HEAD` with porcelain status and both diffs, so a write that commits its edits is `changed` and can stay `passed`. A `rev-parse` failure keeps that empty head term only when the repository has no commits; any other git failure is `unwitnessed`.
 
 - **The test suite no longer passes `--test-timeout`.** Under `node --test` that flag bounds each file, so coverage on CI timed out `test/core/engine.test.ts` at three minutes. The wall clock in `scripts/run-bounded.mjs` still bounds each script.
 
@@ -860,6 +882,10 @@ All notable user-facing changes are documented here. The project follows [Semant
 
 - **`kxm vision` is owned by the browser-verify skill.** `plugins/kxm/skill-suite.json` lists `vision` on `kxm-browser-verify`, the skill that already teaches `kxm vision assert`.
 
+- **Dry-run ship status no longer rewrites `.git/index`.** `git status` refreshes the index under an optional lock. The ship-status read passes `--no-optional-locks`, so a dry run leaves the checkout untouched.
+
+- **A committed checkout counts as authored work, and a one-shot outcome must be a standalone JSON object.** The authoring witness includes `HEAD` with porcelain status and both diffs, so a write that commits its edits is `changed` and can stay `passed`. A `rev-parse` failure keeps that empty head term only when the repository has no commits; any other git failure is `unwitnessed`. A one-shot outcome is accepted when the whole reply is one JSON object, or when that object stands alone on the final line. A closing code fence around the final object is allowed. An object followed by prose, a truncated reply, and an ambiguous tail settle `failed`.
+
 ## 0.7.0 - 2026-09-11
 
 ### Added

package/docs/contracts/lifecycles.md

@@ -105,10 +105,18 @@ created → accepted → dispatched → executing → result_recorded → termin
 | `blocked_uncertain` | A dependent effect cannot be reconciled safely |
 | `terminal` | The logical assignment outcome is final and immutable: passed, failed, or cancelled |
 
-A producer reply becomes a terminal outcome **only** through a declared result: the whole
-reply is one JSON object, or the object stands alone on the final line. A closing code fence
-around that final object is allowed. If the tail is ambiguous, or prose follows the object,
-the outcome is `failed`.
+A producer reply becomes a terminal outcome **only** through a declared result. A one-shot
+reply declares that result when one forward scan finds a JSON object that closes at the
+end of the trimmed text. The scan tracks strings, escapes, and brace depth together. A
+`{` outside a string at depth 0 opens a top-level object, and the newest object that
+returns to depth 0 is the candidate. One closing code fence may follow the object, with
+whitespace. The object may be the whole reply, a fenced reply, or the end of surrounding
+prose, on one line or several. `JSON.parse` runs once on that slice. The value must be a
+plain object whose `outcome` string is one of the step's declared values. Braces inside
+strings do not change depth. A stray `"` leaves the scan inside a string, so a following
+object is not top-level and settles `failed`. Prose after the object, a reply with no
+such object, a disallowed outcome, and an inner object at the end of a truncated outer
+object settle `failed`.
 
 Naming an outcome *word* anywhere in a reply is not a result. `"the gate did not pass, so I
 would not call this passed"` must not advance a step, and an empty or unstructured reply is

package/docs/contributing/ci-and-release.md

@@ -39,8 +39,8 @@ together.
 | Before you push | Local | `npm run verify` |
 | Pull request, push to `main`, or manual | `ci.yml` (`required`) | Aggregates the lanes below into one pass/fail check named `CI / required` |
 | Pull request and push | `ci.yml` (Docs lint) | `lint:docs` and `check:versions`, always |
-| Code pull request and code push | `ci.yml` (Unit engine and Unit, Linux Node 24) | `engine.test.ts` split by test name; `permission.test.ts` and `runtime.test.ts` one file at a time; every other unit file together. Typecheck and `check-generated` run on the light lane |
-| Platform-sensitive pull request | `ci.yml` (Unit, Windows Node 24) | The same four unit lanes on `windows-latest` |
+| Code pull request and code push | `ci.yml` (Unit engine and Unit, Linux Node 24) | `engine.test.ts` split by test name; `permission.test.ts`, `runtime.test.ts`, and `package-install.test.ts` one file at a time; every other unit file in two light shards. Typecheck and `check-generated` run on `light-1` |
+| Platform-sensitive pull request | `ci.yml` (Unit, Windows Node 24) | The same unit lanes on `windows-latest` |
 | Push to `main`, or manual | `ci.yml` (Validate matrix) | `validate:pr` on Linux and Windows for Node 22.19.0 and Node 24; not on a pull request |
 | Code pull request and code push | `ci.yml` (Plugin validation) | `claude plugin validate --strict` on the marketplace and the plugin |
 | Daily at 04:00 UTC, or manual | `nightly.yml` | `test:coverage:complete`, `check`, `check:generated`, `npm pack --dry-run` |
@@ -54,7 +54,7 @@ The npm scripts behind those rows:
 | Script | Composition |
 |---|---|
 | `verify` | `npm test` (core and package unit tests), `check`, `check:generated` |
-| `test:ci-shard` | One unit lane: build, then `engine <index> <total>`, `serial`, or `light` over `test/core/*.test.ts` and `packages/core/*/tests/unit/*.test.ts` |
+| `test:ci-shard` | One unit lane: build, then `engine <index> <total>`, `serial`, `light`, or `light-<index>` over `test/core/*.test.ts` and `packages/core/*/tests/unit/*.test.ts` |
 | `validate:pr` | `build`, `typecheck`, a compact contract and smoke set of nine `test/core` files, `check:versions`, and the generated-`dist` check |
 | `validate:ci` | `test:coverage` (core and package tests, 91/80/92 floors), `check`, `npm pack --dry-run`; the Release workflow runs it, and it stays available locally |
 | `test:coverage:complete` | Core, simulation and package tests with 93/80/93 floors |
@@ -81,15 +81,20 @@ blocker: it is the only place the simulation suite and coverage floors run.
 ### Lanes and the required check
 
 Code pull requests run Docs lint, two Linux Node 24 engine shards, a serial
-lane (`permission.test.ts` then `runtime.test.ts`), a light lane for every
-other unit file, and Plugin validation. The light lane typechecks and checks
-generated bundles. `engine.test.ts` is split by test name because that file
-alone was 174 seconds; the serial lane keeps the next two longest files off
-the light pool, which was 268 seconds when every non-engine file shared one
-job. Linux jobs use the npm cache from
-`actions/setup-node`. Restoring a `node_modules` tarball was slower than
-`npm ci` on the Linux runners (about 24s versus 17s on 2026-09-24), so that
-cache stays on the Windows jobs, where `npm ci` is the slow step.
+lane (`permission.test.ts`, then `runtime.test.ts`, then
+`package-install.test.ts`), two light shards for every other unit file, and
+Plugin validation. `light-1` typechecks and checks generated bundles.
+`engine.test.ts` is split by test name because that file alone was 174
+seconds; the serial lane keeps the next longest files off the light pool,
+which was 268 seconds when every non-engine file shared one job.
+`package-install.test.ts` is serial because on Windows it took 171 seconds
+and the light job stopped reporting tests after that file finished (cancelled
+at 20 minutes on pull requests #361, #362, and #363). The other light files
+are round-robin split so each Windows job is about half of that pool. Linux
+jobs use the npm cache from `actions/setup-node`. Restoring a
+`node_modules` tarball was slower than `npm ci` on the Linux runners (about
+24s versus 17s on 2026-09-24), so that cache stays on the Windows jobs,
+where `npm ci` is the slow step.
 
 The job `required` always runs. Its check name is `CI / required`. It fails
 when a lane fails or is cancelled, and it passes when a lane was skipped

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kontextmind/kxm",
-  "version": "0.7.152",
+  "version": "0.7.154",
   "description": "KXM local-first multi-agent orchestration and operator dashboard",
   "type": "module",
   "author": "KontextMind",

package/plugins/kxm/.claude-plugin/plugin.json

@@ -2,7 +2,7 @@
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "kxm",
   "displayName": "KXM",
-  "version": "0.7.152",
+  "version": "0.7.154",
   "description": "Headless multi-agent orchestration, durable workflows, and a live operator dashboard for Pi and Claude Code",
   "author": {
     "name": "KontextMind",

package/plugins/kxm/dist/mcp-server.js

@@ -17313,7 +17313,7 @@ function sessionTokenFixHint(policy) {
 }
 
 // plugins/kxm/src/mcp-server.ts
-var VERSION = "0.7.152";
+var VERSION = "0.7.154";
 var CONFIGURE_PLUGIN = "/plugin configure kxm@kxm";
 var inbox = /* @__PURE__ */ new Map();
 var notifiedInbox = /* @__PURE__ */ new Set();

package/plugins/kxm/dist/runtime-supervisor.js

@@ -30773,15 +30773,29 @@ function recoverKxmRun(context, runId, request) {
 }
 
 // plugins/kxm/src/oneshot-producer.ts
-function lastBalancedJsonObject(text) {
+function isClosingFence(line) {
+  return /^(`{3,}|~{3,})\s*$/.test(line.trim());
+}
+function stripOneClosingFence(trimmed) {
+  const lineBreak = Math.max(trimmed.lastIndexOf("\n"), trimmed.lastIndexOf("\r"));
+  const lastLine = lineBreak < 0 ? trimmed : trimmed.slice(lineBreak + 1);
+  if (!isClosingFence(lastLine)) return trimmed;
+  if (lineBreak < 0) return "";
+  return trimmed.slice(0, lineBreak).trimEnd();
+}
+function determineOutcome(text, allowedOutcomes) {
+  const trimmed = text.trim();
+  if (!trimmed) return "failed";
+  const body = stripOneClosingFence(trimmed);
+  if (!body) return "failed";
   let depth = 0;
   let inString = false;
   let escape2 = false;
-  let start = -1;
-  let lastStart = -1;
-  let lastEnd = -1;
-  for (let i = 0; i < text.length; i++) {
-    const ch = text[i];
+  let currentStart = -1;
+  let objectStart = -1;
+  let objectEnd = -1;
+  for (let i = 0; i < body.length; i++) {
+    const ch = body[i];
     if (inString) {
       if (escape2) {
         escape2 = false;
@@ -30799,7 +30813,7 @@ function lastBalancedJsonObject(text) {
       continue;
     }
     if (ch === "{") {
-      if (depth === 0) start = i;
+      if (depth === 0) currentStart = i;
       depth++;
       continue;
     }
@@ -30807,50 +30821,22 @@ function lastBalancedJsonObject(text) {
       if (depth === 0) continue;
       depth--;
       if (depth === 0) {
-        lastStart = start;
-        lastEnd = i;
+        objectStart = currentStart;
+        objectEnd = i;
       }
     }
   }
-  if (lastStart < 0) return void 0;
-  return text.slice(lastStart, lastEnd + 1);
-}
-function standaloneObjectText(text) {
-  const trimmed = text.trim();
-  if (!trimmed.startsWith("{") || !trimmed.endsWith("}")) return void 0;
-  const slice = lastBalancedJsonObject(trimmed);
-  return slice === trimmed ? slice : void 0;
-}
-function outcomeOfObjectText(text) {
-  const slice = standaloneObjectText(text);
-  if (slice === void 0) return void 0;
+  if (objectStart < 0 || objectEnd !== body.length - 1) return "failed";
+  let result;
   try {
-    const result = JSON.parse(slice);
-    if (!result || typeof result !== "object" || Array.isArray(result)) return void 0;
-    const outcome = result.outcome;
-    return typeof outcome === "string" ? outcome : void 0;
+    result = JSON.parse(body.slice(objectStart, objectEnd + 1));
   } catch {
-    return void 0;
+    return "failed";
   }
-}
-function isClosingFence(line) {
-  return /^(`{3,}|~{3,})\s*$/.test(line.trim());
-}
-function determineOutcome(text, allowedOutcomes) {
-  const trimmed = text.trim();
-  if (!trimmed) return "failed";
-  const whole = outcomeOfObjectText(trimmed);
-  if (whole !== void 0) return allowedOutcomes.includes(whole) ? whole : "failed";
-  const lines = trimmed.split(/\r?\n/);
-  while (lines.length > 0 && lines[lines.length - 1].trim() === "") lines.pop();
-  if (lines.length > 0 && isClosingFence(lines[lines.length - 1])) {
-    lines.pop();
-    while (lines.length > 0 && lines[lines.length - 1].trim() === "") lines.pop();
-  }
-  if (lines.length === 0) return "failed";
-  const declared = outcomeOfObjectText(lines[lines.length - 1]);
-  if (declared === void 0 || !allowedOutcomes.includes(declared)) return "failed";
-  return declared;
+  if (!result || typeof result !== "object" || Array.isArray(result)) return "failed";
+  const outcome = result.outcome;
+  if (typeof outcome !== "string" || !allowedOutcomes.includes(outcome)) return "failed";
+  return outcome;
 }
 function createKxmOneShotProducer(options = {}) {
   const running = /* @__PURE__ */ new Map();

package/plugins/kxm/dist/runtime.js

@@ -31521,15 +31521,29 @@ function recoverKxmRun(context, runId, request) {
 }
 
 // plugins/kxm/src/oneshot-producer.ts
-function lastBalancedJsonObject(text) {
+function isClosingFence(line) {
+  return /^(`{3,}|~{3,})\s*$/.test(line.trim());
+}
+function stripOneClosingFence(trimmed) {
+  const lineBreak = Math.max(trimmed.lastIndexOf("\n"), trimmed.lastIndexOf("\r"));
+  const lastLine = lineBreak < 0 ? trimmed : trimmed.slice(lineBreak + 1);
+  if (!isClosingFence(lastLine)) return trimmed;
+  if (lineBreak < 0) return "";
+  return trimmed.slice(0, lineBreak).trimEnd();
+}
+function determineOutcome(text, allowedOutcomes) {
+  const trimmed = text.trim();
+  if (!trimmed) return "failed";
+  const body = stripOneClosingFence(trimmed);
+  if (!body) return "failed";
   let depth = 0;
   let inString = false;
   let escape2 = false;
-  let start = -1;
-  let lastStart = -1;
-  let lastEnd = -1;
-  for (let i = 0; i < text.length; i++) {
-    const ch = text[i];
+  let currentStart = -1;
+  let objectStart = -1;
+  let objectEnd = -1;
+  for (let i = 0; i < body.length; i++) {
+    const ch = body[i];
     if (inString) {
       if (escape2) {
         escape2 = false;
@@ -31547,7 +31561,7 @@ function lastBalancedJsonObject(text) {
       continue;
     }
     if (ch === "{") {
-      if (depth === 0) start = i;
+      if (depth === 0) currentStart = i;
       depth++;
       continue;
     }
@@ -31555,50 +31569,22 @@ function lastBalancedJsonObject(text) {
       if (depth === 0) continue;
       depth--;
       if (depth === 0) {
-        lastStart = start;
-        lastEnd = i;
+        objectStart = currentStart;
+        objectEnd = i;
       }
     }
   }
-  if (lastStart < 0) return void 0;
-  return text.slice(lastStart, lastEnd + 1);
-}
-function standaloneObjectText(text) {
-  const trimmed = text.trim();
-  if (!trimmed.startsWith("{") || !trimmed.endsWith("}")) return void 0;
-  const slice = lastBalancedJsonObject(trimmed);
-  return slice === trimmed ? slice : void 0;
-}
-function outcomeOfObjectText(text) {
-  const slice = standaloneObjectText(text);
-  if (slice === void 0) return void 0;
+  if (objectStart < 0 || objectEnd !== body.length - 1) return "failed";
+  let result;
   try {
-    const result = JSON.parse(slice);
-    if (!result || typeof result !== "object" || Array.isArray(result)) return void 0;
-    const outcome = result.outcome;
-    return typeof outcome === "string" ? outcome : void 0;
+    result = JSON.parse(body.slice(objectStart, objectEnd + 1));
   } catch {
-    return void 0;
+    return "failed";
   }
-}
-function isClosingFence(line) {
-  return /^(`{3,}|~{3,})\s*$/.test(line.trim());
-}
-function determineOutcome(text, allowedOutcomes) {
-  const trimmed = text.trim();
-  if (!trimmed) return "failed";
-  const whole = outcomeOfObjectText(trimmed);
-  if (whole !== void 0) return allowedOutcomes.includes(whole) ? whole : "failed";
-  const lines = trimmed.split(/\r?\n/);
-  while (lines.length > 0 && lines[lines.length - 1].trim() === "") lines.pop();
-  if (lines.length > 0 && isClosingFence(lines[lines.length - 1])) {
-    lines.pop();
-    while (lines.length > 0 && lines[lines.length - 1].trim() === "") lines.pop();
-  }
-  if (lines.length === 0) return "failed";
-  const declared = outcomeOfObjectText(lines[lines.length - 1]);
-  if (declared === void 0 || !allowedOutcomes.includes(declared)) return "failed";
-  return declared;
+  if (!result || typeof result !== "object" || Array.isArray(result)) return "failed";
+  const outcome = result.outcome;
+  if (typeof outcome !== "string" || !allowedOutcomes.includes(outcome)) return "failed";
+  return outcome;
 }
 function createKxmOneShotProducer(options = {}) {
   const running = /* @__PURE__ */ new Map();

package/plugins/kxm/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-plugin",
-  "version": "0.7.152",
+  "version": "0.7.154",
   "private": true,
   "type": "module",
   "engines": {

package/plugins/kxm/src/mcp-server.ts

@@ -11,7 +11,7 @@ import { deliverInboxNotification } from "./inbox.ts";
 import type { HubEvent, MessageRecord } from "./protocol.ts";
 import { sessionTokenFixHint } from "./session-token-hint.ts";
 
-const VERSION = "0.7.152";
+const VERSION = "0.7.154";
 const CONFIGURE_PLUGIN = "/plugin configure kxm@kxm";
 const inbox = new Map<string, MessageRecord>();
 const notifiedInbox = new Set<string>();

package/plugins/kxm/src/oneshot-producer.ts

@@ -51,21 +51,47 @@ export interface KxmOneShotProducer extends KxmProducer {
   close(): Promise<void>;
 }
 
-// The last complete top-level `{...}`. One forward pass tracks strings,
-// escapes, and brace depth. A `{` that lifts depth from 0 to 1 opens a
-// candidate; the matching `}` that returns depth from 1 to 0 closes it.
-// The last such span wins. Callers accept the span only when it is the
-// entire trimmed reply or the entire final line; an inner object is not
-// an outcome.
-function lastBalancedJsonObject(text: string): string | undefined {
+function isClosingFence(line: string): boolean {
+  return /^(`{3,}|~{3,})\s*$/.test(line.trim());
+}
+
+// Drop one closing code fence when it is the last line. The caller has already
+// trimmed the reply. Any other tail stays, so prose after the object still fails.
+function stripOneClosingFence(trimmed: string): string {
+  const lineBreak = Math.max(trimmed.lastIndexOf("\n"), trimmed.lastIndexOf("\r"));
+  const lastLine = lineBreak < 0 ? trimmed : trimmed.slice(lineBreak + 1);
+  if (!isClosingFence(lastLine)) return trimmed;
+  if (lineBreak < 0) return "";
+  return trimmed.slice(0, lineBreak).trimEnd();
+}
+
+// The outcome object is the newest complete top-level JSON object in the
+// trimmed reply, and that object must close on the reply's last character.
+// One closing code fence may follow it; the fence is removed before the scan.
+// One forward pass tracks string state and brace depth together. A `"` toggles
+// a string, and `\` escapes only inside a string. A `{` outside a string at
+// depth 0 opens a top-level object. When depth returns to 0, that object is
+// the newest complete top-level object. Braces inside strings do not change
+// depth, so the same scan shows that the object started at depth 0 and that
+// quotes in the prefix are closed. After the scan, if that object closes at
+// the end of the body, JSON.parse runs once on the slice. The value must be
+// a plain object whose outcome is a string in allowedOutcomes. Anything else
+// is failed: prose after the object, no object, a disallowed outcome, an
+// inner object at the end of a truncated outer object, a stray quote that
+// leaves a string open, or a prefix whose braces never return to depth 0.
+function determineOutcome(text: string, allowedOutcomes: readonly string[]): string {
+  const trimmed = text.trim();
+  if (!trimmed) return "failed";
+  const body = stripOneClosingFence(trimmed);
+  if (!body) return "failed";
   let depth = 0;
   let inString = false;
   let escape = false;
-  let start = -1;
-  let lastStart = -1;
-  let lastEnd = -1;
-  for (let i = 0; i < text.length; i++) {
-    const ch = text[i];
+  let currentStart = -1;
+  let objectStart = -1;
+  let objectEnd = -1;
+  for (let i = 0; i < body.length; i++) {
+    const ch = body[i];
     if (inString) {
       if (escape) {
         escape = false;
@@ -83,7 +109,7 @@ function lastBalancedJsonObject(text: string): string | undefined {
       continue;
     }
     if (ch === "{") {
-      if (depth === 0) start = i;
+      if (depth === 0) currentStart = i;
       depth++;
       continue;
     }
@@ -91,59 +117,22 @@ function lastBalancedJsonObject(text: string): string | undefined {
       if (depth === 0) continue;
       depth--;
       if (depth === 0) {
-        lastStart = start;
-        lastEnd = i;
+        objectStart = currentStart;
+        objectEnd = i;
       }
     }
   }
-  if (lastStart < 0) return undefined;
-  return text.slice(lastStart, lastEnd + 1);
-}
-
-// A standalone object is one balanced `{...}` covering the entire trimmed text.
-function standaloneObjectText(text: string): string | undefined {
-  const trimmed = text.trim();
-  if (!trimmed.startsWith("{") || !trimmed.endsWith("}")) return undefined;
-  const slice = lastBalancedJsonObject(trimmed);
-  return slice === trimmed ? slice : undefined;
-}
-
-function outcomeOfObjectText(text: string): string | undefined {
-  const slice = standaloneObjectText(text);
-  if (slice === undefined) return undefined;
+  if (objectStart < 0 || objectEnd !== body.length - 1) return "failed";
+  let result: unknown;
   try {
-    const result: unknown = JSON.parse(slice);
-    if (!result || typeof result !== "object" || Array.isArray(result)) return undefined;
-    const outcome = (result as Record<string, unknown>).outcome;
-    return typeof outcome === "string" ? outcome : undefined;
+    result = JSON.parse(body.slice(objectStart, objectEnd + 1));
   } catch {
-    return undefined;
-  }
-}
-
-function isClosingFence(line: string): boolean {
-  return /^(`{3,}|~{3,})\s*$/.test(line.trim());
-}
-
-// The whole reply is one JSON object, or the object stands alone on the final
-// line. A closing code fence after that line is allowed. Prose after the
-// object, an ambiguous tail, a truncated reply, and a missing or disallowed
-// outcome settle failed.
-function determineOutcome(text: string, allowedOutcomes: readonly string[]): string {
-  const trimmed = text.trim();
-  if (!trimmed) return "failed";
-  const whole = outcomeOfObjectText(trimmed);
-  if (whole !== undefined) return allowedOutcomes.includes(whole) ? whole : "failed";
-  const lines = trimmed.split(/\r?\n/);
-  while (lines.length > 0 && lines[lines.length - 1]!.trim() === "") lines.pop();
-  if (lines.length > 0 && isClosingFence(lines[lines.length - 1]!)) {
-    lines.pop();
-    while (lines.length > 0 && lines[lines.length - 1]!.trim() === "") lines.pop();
+    return "failed";
   }
-  if (lines.length === 0) return "failed";
-  const declared = outcomeOfObjectText(lines[lines.length - 1]!);
-  if (declared === undefined || !allowedOutcomes.includes(declared)) return "failed";
-  return declared;
+  if (!result || typeof result !== "object" || Array.isArray(result)) return "failed";
+  const outcome = (result as Record<string, unknown>).outcome;
+  if (typeof outcome !== "string" || !allowedOutcomes.includes(outcome)) return "failed";
+  return outcome;
 }
 
 export function createKxmOneShotProducer(options: KxmOneShotProducerOptions = {}): KxmOneShotProducer {

package/scripts/ci-unit-shard.mjs

@@ -7,24 +7,33 @@
 // sequentially, and running them in one process pool stretched the rest of
 // the suite to 268s. engine.test.ts is split by test name across two jobs.
 // permission and runtime run one file at a time in a serial job so they do
-// not steal cores from each other. Every other unit file runs in a light
-// job at concurrency 4.
+// not steal cores from each other.
+//
+// package-install.test.ts is serial as well. On the Windows light lane for
+// PRs #361, #362, and #363 it took 171s (67s on Linux) and was the file that
+// finished immediately before that lane stopped reporting tests. The job
+// then sat until the 20-minute cancel. npm pack/install stays off the light
+// pool. The other light files are round-robin split into two shards so a
+// Windows job is not one 120-file process.
 //
 // Dynamic `test(\`...\${...}\`)` names stay together as one pattern so a
 // loop is not dropped. Run via `npm run test:ci-shard` so npm_execpath is
-// set for the packed-install test: `engine <index> <total>`, `serial`, or
-// `light`.
+// set for the packed-install test: `engine <index> <total>`, `serial`,
+// `light`, or `light-<index>` (two shards).
 
 import { spawn, spawnSync } from "node:child_process";
 import { globSync, readFileSync } from "node:fs";
 import { join } from "node:path";
 
 export const SHARD_TOTAL = 2;
+export const LIGHT_SHARD_TOTAL = 2;
 export const ENGINE_FILE = "test/core/engine.test.ts";
-// Solo Node 24 timings: permission 74s, runtime 53s. One file at a time.
+// Solo Node 24 timings: permission 74s, runtime 53s. package-install is 171s
+// on Windows CI and 67s on Linux. One file at a time.
 export const SERIAL_FILES = [
   "test/core/permission.test.ts",
   "test/core/runtime.test.ts",
+  "test/core/package-install.test.ts",
 ];
 const LIGHT_CONCURRENCY = 4;
 
@@ -125,6 +134,15 @@ export function planLight(root) {
   return light;
 }
 
+export function planLightShard(root, index, total = LIGHT_SHARD_TOTAL) {
+  if (!Number.isInteger(index) || !Number.isInteger(total) || index < 1 || index > total) {
+    throw new Error(`light shard ${index}/${total} is outside 1..${total}`);
+  }
+  const shard = planLight(root).filter((_, fileIndex) => fileIndex % total === index - 1);
+  if (shard.length === 0) throw new Error(`light shard ${index}/${total} assigned no unit files`);
+  return shard;
+}
+
 export function coverageOfShards(root, total = SHARD_TOTAL) {
   const heavy = new Map();
   for (let index = 1; index <= total; index += 1) {
@@ -138,7 +156,15 @@ export function coverageOfShards(root, total = SHARD_TOTAL) {
 
 function runNode(args, label, children) {
   return new Promise((resolve) => {
-    const child = spawn(process.execPath, args, { env: process.env });
+    const child = spawn(process.execPath, args, {
+      env: {
+        ...process.env,
+        // Git Credential Manager on windows-latest can wait on a prompt
+        // until the job is cancelled. Refuse that wait for every lane.
+        GIT_TERMINAL_PROMPT: process.env.GIT_TERMINAL_PROMPT || "0",
+        GCM_INTERACTIVE: process.env.GCM_INTERACTIVE || "Never",
+      },
+    });
     children.add(child);
     let stdout = "";
     let stderr = "";
@@ -220,10 +246,14 @@ async function main() {
       (children) => runNode(testArgs(1, files), "serial", children),
     ], 1));
   }
-  if (mode === "light") {
-    const files = planLight(root);
+  const lightShard = /^light-(\d+)$/.exec(mode);
+  if (mode === "light" || lightShard) {
+    const files = lightShard
+      ? planLightShard(root, Number(lightShard[1]), LIGHT_SHARD_TOTAL)
+      : planLight(root);
+    const label = lightShard ? mode : "light";
     process.exit(await runPool([
-      (children) => runNode(testArgs(LIGHT_CONCURRENCY, files), "light", children),
+      (children) => runNode(testArgs(LIGHT_CONCURRENCY, files), label, children),
     ], 1));
   }
   if (mode === "engine") {
@@ -234,7 +264,7 @@ async function main() {
       (children) => runNode(testArgs(1, [plan.file], plan.patterns), "engine", children),
     ], 1));
   }
-  process.stderr.write("usage: ci-unit-shard.mjs <serial|light|engine> [index total]\n");
+  process.stderr.write("usage: ci-unit-shard.mjs <serial|light|light-<index>|engine> [index total]\n");
   process.exit(2);
 }