mixdog 1.2.1 → 1.2.3

package/src/agents/reviewer/agent.json

@@ -1,6 +1,6 @@
 {
   "id": "reviewer",
   "name": "Reviewer",
-  "description": "Use once at the final check of non-trivial work: after the implementation is written and verified, before reporting completion, to independently verify correctness, regressions, risks, and acceptance criteria. Skip for low-difficulty work.",
+  "description": "Use only at the end, as the final check of non-trivial work: after the implementation is written and verified, before reporting completion, to independently verify correctness, regressions, risks, and acceptance criteria. Skip for low-difficulty work.",
   "entry": "AGENT.md"
 }

package/src/defaults/shipped-history.json

@@ -1,6 +1,7 @@
 {
   "agents": {
     "advisor": [
+      "6bd0318995f70547c162efda92795e79",
       "d3b4a7926c9484833fe8ddba70260f9d"
     ],
     "debugger": [
@@ -26,6 +27,7 @@
     "reviewer": [
       "07643b62440d3eba29b7360b8a33f09f",
       "25b44b46269490e26ca31c1852925171",
+      "a60ea68f9287d55489e92e5450432f62",
       "a61cb4557def4e7ed647d4bdacbb72e4",
       "de860d4111b61985f3b59eb1011fa421"
     ],
@@ -61,6 +63,7 @@
       "56355187d05762ef50579d6285636c71",
       "599d6b09954a3afb163091878d3b7040",
       "5fa8ddf4731b39cec5fa269d632cfcc4",
+      "627f69923f1b895e3ac5f95f9b2d766b",
       "660cc103947dd30b94c70bcd1f22cefe",
       "6bbd8bb9e7657bc642f7be02f36dfc6d",
       "7976531213723a66a7a53b9a1cf0f372",
@@ -98,6 +101,7 @@
   "outputStyles": {
     "common.md": [
       "4e6134d60ee5c54c2888e874653bcdad",
+      "4ecd89b0c07ad8a81aa1f84882cf1b00",
       "899529696b7d6d3373ae83a76976ab48",
       "8d343d0a406fd5817b7fd98a3ac47010",
       "a1b73ce404125d3ce0a5762e4a1e1555",
@@ -335,9 +339,11 @@
       "70c54f802c77b630863b9e17a16d8224",
       "81e979463e0066e7a45cf511a31c5d68",
       "8476c811b2f02da3a434bcb23a1c3d12",
+      "8af2b614e2a07b58dad6c0f1318d22e2",
       "9f90d3872b1ecd312977ce661b770b31",
       "ade615dd0f6ae6ed1601023e5b9cb692",
       "b4f9c79dcb08151e6275b5ee35c2bca4",
+      "c6db973f51166fc42814808ae0001797",
       "c8826387a65631b63e8023f55ef807a3",
       "cb6f66d14ba75b11e7470eb3fd5ec94c",
       "de9727a48e52673258bf5f9670564d56",

package/src/defaults/skills/setup/SKILL.md

@@ -69,7 +69,8 @@ that a call returned without transport error.
   injection stays on regardless. `set_recap_enabled` changes only the
   background cycles. Core Memory entries
   are read and changed with the `memory` tool, not setup.
-- Built-ins have installation state separate from enabled state. Browser Use,
+- Built-ins have installation state separate from enabled state; Visualize
+  is toggle-only (on by default, no install step). Browser Use,
   Computer Use actions require this conversation open in the local Desktop
   window; voice runs on the runtime and needs no Desktop.
 - Hook rules are before-tool policy, separate from MCP and plugins. Deleting a
@@ -94,6 +95,10 @@ that a call returned without transport error.
 - For an MCP failure, inspect the reported transport error before changing
   command, arguments, directory, environment, URL, headers, or enabled state.
 - MCP reads intentionally omit raw connection values that may contain secrets.
+- An HTTP/SSE MCP server that needs OAuth waits for sign-in instead of
+  failing; setup rows then show it enabled, not connected, with no `error`.
+  Setup cannot sign in, and `reconnect_mcp` does not log in: hand off to
+  `open mcp` (Connect/Reconnect) or the TUI `/mcp login`.
 - For an unavailable built-in, distinguish not installed, installed but
   disabled, and bridge inactive.
 - Desktop requests require a single live claimant and a confirmed receipt.

package/src/defaults/skills/setup/references/actions.md

@@ -4,11 +4,17 @@ Read this when choosing a `setup` mutation.
 
 ## Read and navigate
 
-- `status` reads one domain: summary, model, agents, workflow, websearch,
-  output-style, profile, autoclear, compaction, memory, features, local-provider, shell,
-  providers, mcp, skills, plugins, update, onboarding, worktree, capabilities, developer,
-  desktop, appearance, projects, connection, schedules, webhooks, hooks, or
-  media.
+- `status` reads one domain from the schema enum; pick the narrowest.
+- Results are compact JSON. `status providers` lists configured providers
+  only (`id`, `kind`, `authenticated`, `status`, `accounts` count); add
+  `detail: true` for every provider with key URLs, expiry and account ids.
+- `list_models` narrows in steps: no `catalog` returns the provider index
+  (`provider`, `count`, `defaultModel?`); `catalog.provider` returns compact
+  rows (`id`, `display`, `context`, `efforts`, `fast?`, `vision?`) with
+  `total`/`page`/`pages`/`nextPage` (`page`, `limit` ≤ 100, default 25);
+  `catalog.model` (plus `provider` if several offer the id) returns that
+  model's full options. `webSearch` limits to native-search models; `refresh`
+  reloads the catalog.
 - `open` navigates to a supported UI surface. Read `surfaces.md` for targets and
   UI-only settings.
 
@@ -28,7 +34,7 @@ Read this when choosing a `setup` mutation.
 | Background memory cycles | `set_recap_enabled` |
 | Web Search tool exposure | `set_web_search_enabled` |
 | Which user instruction files (AGENTS.md / CLAUDE.md) new sessions load: all, agents, claude, or off | `set_instruction_files` |
-| Git, Office, Code Tidy, Local Provider, Auto Effort, Browser, Computer, or voice enabled state | `set_builtin_enabled` |
+| Git, Office, Code Tidy, Local Provider, Auto Effort, Browser, Computer, voice, or Visualize enabled state | `set_builtin_enabled` |
 | Git, Memory, Office, Code Tidy, Local Provider, Auto Effort, Browser, Computer, or voice preparation | `install_builtin` |
 | Managed local model installation | `install_local_model` |
 | Start or resume a background runtime/model download | `start_local_installation` |
@@ -44,7 +50,7 @@ Read this when choosing a `setup` mutation.
 | Automatic updates | `set_auto_update` |
 | Worktree defaults: new tasks on the current branch or a new worktree, branch prefix, large-repository warning size, automatic cleanup | `set_worktree_settings` |
 | Remove stored provider authentication (one OAuth account at a time) | `forget_provider_auth` |
-| Select, rename, reorder, or auto-switch OAuth accounts, or allow paid credits | `set_provider_account` |
+| Select, rename, or reorder OAuth accounts, set the switching policy, or allow paid credits | `set_provider_account` |
 | Developer options such as risk-gated OAuth providers | `set_developer_option` |
 | Add an MCP server | `add_mcp_server` |
 | Edit or rename an MCP server | `save_mcp_server` |
@@ -58,7 +64,7 @@ Read this when choosing a `setup` mutation.
 | Enable or disable a plugin | `set_plugin_enabled` |
 | Remove a plugin | `remove_plugin` |
 | Turn back on and reconnect every MCP server a plugin declares | `enable_plugin_mcp` |
-| Available hosted models and their options; optionally one provider, native-search capable only, or a refreshed catalog | `list_models` |
+| Hosted model catalog: provider index, one provider's paged model rows, or one model's full options | `list_models` |
 | Agent orchestration, independent of workflow instructions | `set_orchestration_mode` |
 | Safe MCP edit metadata, excluding raw connection/credential values | `get_mcp_server` |
 | Read a workflow, agent, or skill definition | `read_definition` |
@@ -69,7 +75,7 @@ Read this when choosing a `setup` mutation.
 | Delete an existing schedule/webhook | `delete_automation` |
 | Enable/disable a schedule/webhook | `set_automation_enabled` |
 | Webhook listener enabled state, port, or domain | `set_webhook_config` |
-| Desktop keep-awake, run in background after window close, turn-completion notifications, notification volume (Windows), usage pin, or Computer observe-only state | `set_desktop_settings` |
+| Desktop keep-awake, run in background after window close, turn-completion notifications, notification volume (Windows), usage pin, usage rail short + long bars, or Computer observe-only state | `set_desktop_settings` |
 | Desktop theme, display language, side panels, markdown default view, link previews, or tool activity expansion | `set_appearance` |
 | Add a before-tool hook rule | `add_hook_rule` |
 | Enable or disable a hook rule by its `status hooks` index | `set_hook_rule_enabled` |
@@ -88,8 +94,14 @@ workflow` before changing the corresponding domain.
 
 - Main and agent routes also accept `modelParameters`, `contextPercent`, and
   `contextTokens` (an exact window in tokens that overrides `contextPercent`);
-  agent routes accept `disabled`.
-  Use the model catalog for actual available options, not remembered model ids.
+  agent routes accept `disabled`. Route `autoEffort` (Main and agent only)
+  turns Auto reasoning on/off for that route; the Auto Effort built-in toggle
+  is the machine-wide switch and must also be on and `ready`.
+  Before `set_route`, read `list_models` with `catalog.model` (and `provider`)
+  for the model's `effortOptions`, `defaultEffort`, `fastCapable`,
+  `serviceTiers`, `modelParameterOptions` and `parameterVariants`; take
+  `effort`, `fast` and `modelParameters` values only from it, never from
+  remembered model ids.
 - A Web Search route must support native web search; tool exposure is changed
   separately.
 - Definition actions use `definitionKind` (workflow, agent, skill).
@@ -102,12 +114,13 @@ workflow` before changing the corresponding domain.
 
 ## Providers and authentication
 
-Use `status providers` for authenticated state and source.
+Use `status providers` for authenticated state; `detail: true` adds each API
+key's source.
 
 - API-key and OAuth providers: `open providers`, then the user completes
   authentication in the visible UI.
 - Custom providers are registered, tested, and removed in the Providers UI;
-  `status providers` lists them as API rows with `custom: true`.
+  `status providers` with `detail: true` lists them as API rows with `custom: true`.
 - Local Provider is installed and enabled as a Built-in. Its endpoint is
   Mixdog-managed and is never configured by the user.
   Use the `local-provider` skill for hardware checks, model selection, downloads,
@@ -115,10 +128,14 @@ Use `status providers` for authenticated state and source.
   the detail dialog lists installed models rather than the download catalog.
 - Forgetting authentication is destructive and requires explicit approval.
   It removes one OAuth account; a provider with several accounts requires
-  `accountId` from the account list in `status providers`.
-- `set_provider_account` changes the selected account, automatic switching,
-  paid-credit use after plan limits (`useCredits`), account order (every account id exactly once), or an account label. Sign-in
-  for a new account still happens in the Providers UI.
+  `accountId` from the account list in `status providers` with `detail: true`.
+- `set_provider_account` changes the selected account, paid-credit use after
+  plan limits (`useCredits`), account order (every account id exactly once),
+  an account label, or the switching `policy`: `manual` stays on the selected
+  account, `limit` switches when it hits a limit, `headroom` starts each new
+  session on the account with the most headroom. `status providers` with
+  `detail: true` returns the current `policy` when one is set. Sign-in for a new account still
+  happens in the Providers UI.
 - `status developer` lists developer options. An option with a `warning` is
   enabled only after the user explicitly accepts that warning, passed as
   `riskAccepted:true`; never assume acceptance.
@@ -159,7 +176,12 @@ markers are read with `status desktop`.
   reasoning effort with a local judge model downloaded in the background.
   Enabling or installing it fails on unsupported machines; `status features`
   reports `supported`, `installing` (judge download running) and `ready`.
+- Visualize is toggle-only through `set_builtin_enabled` (on by default);
+  `install_builtin` does not accept it.
 - Terminal follows the live Desktop terminal bridge and has no toggle.
+- Environment feature overrides (`MIXDOG_FEATURE_*`) win over these toggles;
+  they are diagnostics for headless or benchmark environments, not user
+  settings.
 - Browser Use and Computer Use installation/toggles use the same
   `install_builtin` and `set_builtin_enabled` actions through a live Desktop
   receipt. Computer Use runs on Windows, macOS, and Linux desktops. These
@@ -177,16 +199,12 @@ markers are read with `status desktop`.
   only as `hasArgumentEdit`; their contents are never returned. Indexes shift
   after a delete;
   re-read status before addressing another rule.
-- `add_hook_rule` takes `action` (allow, deny, block, modify, rewrite, ask)
-  and optional `tool`, `match`, `cwd`, `reason`. modify/rewrite need exactly
-  one of `patch` (merged into the arguments) or `args` (replaces them).
+- `add_hook_rule` with modify/rewrite needs exactly one of `patch` or `args`.
 - `delete_hook_rule` is destructive and requires explicit approval for the
   named rule; disabling with `set_hook_rule_enabled` preserves it.
 - `status media` shows the remembered lane/model per kind. `set_media_default`
   accepts only a signed-in lane that offers the kind (and the model, when
   given) in the live media catalog; the media tool's `list` reads that catalog.
-- Environment feature overrides are diagnostics for headless or benchmark
-  environments, not ordinary user settings.
 
 ## Skills, MCP, and plugins
 
@@ -201,6 +219,10 @@ markers are read with `status desktop`.
   transport. Raw credential fields are not accepted; `env_vars`,
   `bearer_token_env_var`, and `env_http_headers` reference existing environment
   variables without exposing their values.
+- `startupTimeoutSec` (connect/handshake, default 30) and `toolTimeoutSec`
+  (per tool call, default 120) are positive seconds. Raise them for slow
+  servers instead of reconnecting repeatedly; a timed-out call is never
+  replayed.
 - Environment/header maps are partial patches. Omitted keys are preserved;
   `null` explicitly removes one key, including a stored credential.
 - Plugin enablement moves its contributed skills and MCP integrations together.

package/src/defaults/skills/setup/references/surfaces.md

@@ -18,7 +18,7 @@ or requests a setting with no mutation action.
 | `profile` | User profile |
 | `autoclear` | Session lifecycle controls |
 | `memory` | Memory capability surface |
-| `mcp` | Extensions → Skill → MCP |
+| `mcp` | Extensions → Skill → MCP, including OAuth Connect/Reconnect/Disconnect |
 | `skills` | Extensions → Skill → Skills |
 | `plugins` | Extensions → Plugin |
 | `update` | Update surface |
@@ -56,15 +56,15 @@ the user to Desktop.
 
 ## Desktop-backed settings
 
-Setup can read and change Desktop appearance (including markdown default
-view, link previews, and tool activity expansion), keep-awake, run in
-background, turn-completion notifications, notification volume (Windows),
-usage pin, activity rail pins, Computer observe-only, Browser/Computer
-installation and toggles, Projects, and linked-device revocation through the
-attached Desktop. A receipt identifies `scope:desktop-host`.
+`set_desktop_settings`, `set_appearance`, `set_activity_rail_pins`,
+`save_project`, `remove_project`, `revoke_linked_device`, and Browser/Computer
+install or toggles run through the attached Desktop; the receipt identifies
+`scope:desktop-host`.
 
 Computer observe-only and background Memory cycles (`set_recap_enabled`)
-have no settings control in either app; setup is their only supported route. Usage pin is the pin toggle on the activity rail usage card.
+have no settings control in either app; setup is their only supported route.
+Usage pin and the short + long rail bars live in the usage card's flyout menu
+on the activity rail.
 
 Schedules/webhooks and workflow/agent/skill definitions use their runtime
 actions directly. These do not require clicking their Desktop rails.
@@ -74,6 +74,8 @@ actions directly. These do not require clicking their Desktop rails.
 The setup tool does not mutate:
 
 - API keys, OAuth, usage sign-in, and GitHub CLI connection;
+- MCP server OAuth sign-in and sign-out (Desktop MCP list/editor, TUI
+  `/mcp login` and `/mcp logout`; remote windows paste the redirect address);
 - notification permission and subscription on the receiving device;
 - web-app pairing credentials and OS permission dialogs;
 - TUI-local theme palettes (use `/theme`; Desktop themes are separate).

package/src/headless-exec.mjs

@@ -437,19 +437,26 @@ async function createHeadlessSessionRuntime(boundary, { provider, model, cwd, we
 // hundred bytes and the write is atomic, so the cost is negligible and a
 // live run stays readable from outside.
 function headlessAskOptions({ stats, lifecycle, observedModels, writeUsage }) {
+  const recordUsage = (delta) => {
+    const observedModel = clean(delta?.model);
+    if (observedModel) observedModels.add(observedModel);
+    applyUsageDelta(stats, delta);
+    try {
+      writeUsage();
+    } catch {
+      // Telemetry must never break the session; the exit path reports.
+    }
+  };
   return {
     onTextReset: () => true,
     onUsageDelta: (delta) => {
-      const observedModel = clean(delta?.model);
-      if (observedModel) observedModels.add(observedModel);
-      applyUsageDelta(stats, delta);
+      recordUsage(delta);
       lifecycle?.onUsageDelta(delta);
-      try {
-        writeUsage();
-      } catch {
-        // Telemetry must never break the session; the exit path reports.
-      }
     },
+    // Native web search runs its own provider send; fold its spend into the
+    // run's totals. It is not a main-stream request, so the lifecycle's
+    // model-request events stay untouched.
+    onNativeWebSearchUsage: recordUsage,
     ...(lifecycle
       ? {
           onProviderSendStarted: () => lifecycle.onProviderSendStarted(),

package/src/lib/keychain-cjs.cjs

@@ -282,7 +282,8 @@ function darwinSet(account, value) {
 
 function darwinDelete(account) {
   const r = darwinRun(['delete-generic-password', '-a', account, '-s', SERVICE]);
-  if (r.status !== 0) throw new Error(`[keychain] security delete failed: ${r.stderr || r.stdout}`);
+  // Exit 44 is errSecItemNotFound: nothing stored, so nothing to delete.
+  if (r.status !== 0 && r.status !== 44) throw new Error(`[keychain] security delete failed: ${r.stderr || r.stdout}`);
 }
 
 // ---------------------------------------------------------------------------
@@ -983,9 +984,11 @@ function win32Set(account, value) {
 }
 
 function win32Delete(account) {
-  const file = dpApiFile(account);
-  if (!fs.existsSync(file)) throw new Error(`[keychain] secret not found: ${account}`);
-  fs.unlinkSync(file);
+  try {
+    fs.unlinkSync(dpApiFile(account));
+  } catch (error) {
+    if (error?.code !== 'ENOENT') throw error;
+  }
 }
 
 // ---------------------------------------------------------------------------

package/src/output-styles/common.md

@@ -7,13 +7,14 @@ partial: true
 
 ## Shared Output Format
 
-- Lead with the answer or action. Use natural, complete sentences and the
-  reader's vocabulary; state each point once.
-- Choose structure for clarity, not a fixed template. Use GitHub-flavored
-  Markdown, `$$…$$` for display math, and tables or code when alignment helps.
-- Distinguish facts, inference, and uncertainty. Preserve exact technical
-  literals, useful links, and material risks, limitations, and scope.
-- Follow explicit requests for depth; never pad. Omit tangents, unsolicited
-  advice or offers, raw dumps, and process recaps.
-- Apply the selected depth to substantive answers, not preambles or progress
-  updates.
+- Lead with the answer; no restated request or closing recap. Natural,
+  complete sentences in the reader's vocabulary, each point once; no arrow
+  chains or unexplained shorthand.
+- Plain prose for simple questions; headings, lists, or tables only for real
+  structure. GitHub-flavored Markdown, `$$…$$` for display math.
+- Distinguish facts, inference, and uncertainty; keep exact literals and
+  material risks. Caveats only when they change the reader's next step.
+- Never pad: no tangents, unsolicited offers, raw dumps, or process recaps.
+  Explicit depth requests win; depth applies to substantive answers only.
+- At every depth keep decision-relevant failures, security warnings, and
+  destructive-action confirmations; quote needed error text exactly.

package/src/repl/slash-commands.mjs

@@ -3,6 +3,7 @@
  * to `ctx.out`; only `/exit` and `/quit` return 'exit'.
  */
 import { dim, green, red, yellow, colorEnabled } from '../runtime/shared/statusline/ansi.mjs';
+import { switchesSessionModel } from '../session-runtime/model-route/model-switch-target.mjs';
 import { createSessionStats } from '../ui/session-stats.mjs';
 import { statuslineFor } from './turn-output.mjs';
 
@@ -65,7 +66,21 @@ async function setModel(ctx, arg) {
     return;
   }
   const runtime = await ctx.ensureRuntime();
-  await runtime.setRoute({ model: arg });
+  const request = { model: arg };
+  // The user's model choice switches this conversation in place; with
+  // auto-compact on model change (default on) the history is compacted on the
+  // current model first. A failed compaction never blocks the switch.
+  if (switchesSessionModel(runtime, request) && runtime.getCompactionSettings?.()?.onModelChange !== false) {
+    // An invalid or unavailable selector is rejected before anything compacts.
+    await runtime.setRoute(request, { validateOnly: true });
+    out.write(`${dim('compacting conversation before the model change…')}\n`);
+    try {
+      await runtime.compact();
+    } catch {
+      /* next-turn auto-compaction fits the new window */
+    }
+  }
+  await runtime.setRoute(request, { switchEstablished: true });
   out.write(`${green(`✓ model → ${runtime.provider}/${runtime.model}`)}\n`);
 }
 

package/src/rules/agent/AGENT.md

@@ -6,8 +6,8 @@
   unapproved destructive action, or a blocker.
 - Communicate in English. Call tools immediately, with no preamble or
   progress text.
-- Reporting: report once, when the work is done, as the handoff — outcome;
-  changed files with the key functions or symbols; the tests and checks you ran for your
-  own changes and their result; material risks, blockers, or unresolved tool
-  failures. Never raw logs, process narration, or a restatement of the brief.
-  Lead owns the independent review.
+- Reporting: report once, when done, per your role's output contract;
+  otherwise done, partial, or blocked with the outcome, then only applicable
+  items: changed files with key symbols, checks run (passes in one line,
+  failures with decisive error text), material risks or blockers. No raw
+  logs, narration, or restated brief. Lead owns the independent review.

package/src/rules/lead/LEAD.md

@@ -17,11 +17,11 @@
 <!-- tools: agent -->
 - Briefing an agent: it sees only the brief, never this conversation. The
   brief is written in English whatever language the user writes in, and
-  holds exactly the task in the user's terms, the scope it owns (files or
-  modules), the exact changes located by path plus function or symbol name
-  or a quoted line, not line numbers, and the completion criteria — nothing
-  else: no rationale, measurements, history, or rules the
-  agent already has. A fresh or respawned agent gets the whole brief; a
+  holds only the task in the user's terms and its use; owned scope and
+  exclusions; changes located by path plus symbol or quoted line, not line
+  numbers; completion criteria; any return need beyond the role. No
+  background, rediscoverable detail, or known rules; large evidence by path.
+  A fresh or respawned agent gets the whole brief; a
   continuing agent gets only what changed. Never assign the same work twice;
   interrupt only to cancel.
 <!-- tools: agent -->

package/src/runtime/agent/orchestrator/agent-runtime/agent-progress-watchdog.mjs

@@ -59,6 +59,25 @@ function assistantMessageText(content) {
     .join('\n');
 }
 
+// Visible assistant text inside a providerReplay envelope. Item shapes by
+// provider: anthropic `{type:'text',text}`; openai/xai responses
+// `{type:'message',content:[{type:'output_text',text}]}`; gemini/antigravity
+// parts `{text}` (reasoning is `{thought:true,text}`). Reasoning/thinking and
+// tool items are skipped.
+function replayItemText(item) {
+  if (!item || typeof item !== 'object') return '';
+  if (item.type === 'message') return assistantMessageText(item.content);
+  if (item.type === 'text' || item.type === 'output_text' || (item.type === undefined && item.thought !== true)) {
+    return typeof item.text === 'string' ? item.text : '';
+  }
+  return '';
+}
+
+function providerReplayText(replay) {
+  const items = Array.isArray(replay?.items) ? replay.items : [];
+  return items.map(replayItemText).filter(Boolean).join('\n');
+}
+
 /** Message index at askSession start — salvage only assistant rows appended this run. */
 export function resolveHandoffMessageStartIndex(session) {
   const messages = Array.isArray(session?.messages) ? session.messages : [];
@@ -72,7 +91,8 @@ function collectSessionAssistantHandoffText(session, messageStartIndex = 0) {
   for (let i = start; i < messages.length; i += 1) {
     const m = messages[i];
     if (m?.role !== 'assistant') continue;
-    const t = assistantMessageText(m.content).trim();
+    const contentText = assistantMessageText(m.content).trim();
+    const t = contentText || providerReplayText(m.providerReplay).trim();
     if (t && t !== '.') parts.push(t);
   }
   return parts.length ? parts.join('\n\n') : '';
@@ -80,9 +100,14 @@ function collectSessionAssistantHandoffText(session, messageStartIndex = 0) {
 
 export function watchdogPartialHandoffFromError(error, session, messageStartIndex = 0) {
   if (!isAgentProgressWatchdogAbortError(error)) return null;
-  // A canonical agent turn cannot expose its live transcript here, so it
-  // reads the text itself and carries it on the stall error.
-  if (typeof error.partialHandoff === 'string') return error.partialHandoff;
+  return turnPartialHandoff(error, session, messageStartIndex);
+}
+
+/** Assistant text a stopped turn produced before it stopped. A canonical agent
+ *  turn cannot expose its live transcript here, so it reads the text itself
+ *  and carries it on the turn's error. */
+export function turnPartialHandoff(error, session, messageStartIndex = 0) {
+  if (typeof error?.partialHandoff === 'string') return error.partialHandoff;
   return partialHandoffTextFromSession(session, messageStartIndex);
 }
 

package/src/runtime/agent/orchestrator/agent-runtime/cache-strategy.mjs

@@ -210,6 +210,16 @@ export function resolveProviderCacheKey(opts, provider) {
   );
 }
 
+/** A cache identity as a prompt_cache_key: kept as is up to the 64-char wire
+ *  limit, a bounded hash identity beyond it, '' when empty. */
+export function normalizeCacheKey(value) {
+  const text = String(value ?? '').trim();
+  if (!text) return '';
+  const wellFormed = text.toWellFormed();
+  if (wellFormed.length <= 64) return wellFormed;
+  return `pc_${createHash('sha256').update(wellFormed).digest('hex').slice(0, 32)}`;
+}
+
 function shortHash(value, chars = 18) {
   return createHash('sha256').update(stableHashStringify(value)).digest('hex').slice(0, chars);
 }