mixdog 1.3.0 → 1.3.1

package/NOTICE.md

@@ -125,15 +125,58 @@ downloads it from this repository's `effort-judge-v*` GitHub releases, as
 listed in `src/runtime/effort-judge/model-manifest.json`.
 
 Statement of changes (Apache-2.0 section 4(b)): only the text encoder is kept,
-reduced to 12 of its layers; a four-level classification head is added and the
-model is fine-tuned to rate the reasoning effort a request needs; the weights
-are exported to ONNX and quantized to 4 bits. Full license text:
-`LICENSES/Apache-2.0.txt`.
-
-Its training data includes requests from SWE-chat
-(<https://huggingface.co/datasets/cfahlgren1/SWE-chat>), made available under
-the Open Data Commons Attribution License 1.0, and from DataClaw exports that
-their authors published on the Hugging Face Hub under the MIT License.
+reduced to 18 of its layers; a four-level classification head is added and the
+model is fine-tuned to rate the reasoning effort a request or an agent step
+needs; the weights are exported to ONNX and quantized to 4 bits. Full license
+text: `LICENSES/Apache-2.0.txt`.
+
+Its training data includes requests and agent steps from SWE-chat
+(<https://huggingface.co/datasets/SALT-NLP/SWE-chat>, Open Data Commons
+Attribution License 1.0); OpenAssistant oasst2
+(<https://huggingface.co/datasets/OpenAssistant/oasst2>, Apache License 2.0);
+DataClaw exports that their authors published on the Hugging Face Hub under the
+MIT License (<https://huggingface.co/datasets/di-zhang-fdu/ShareDataClaw>);
+DevGPT by Tao Xiao, Christoph Treude, Hideaki Hata and Kenichi Matsumoto
+(<https://doi.org/10.5281/zenodo.16392320>, CC BY 4.0); issue texts from the
+SWE-bench training split (<https://github.com/SWE-bench/SWE-bench>, MIT
+License); and agent steps from NVIDIA's SWE-Hero OpenHands trajectories
+(<https://huggingface.co/datasets/nvidia/SWE-Hero-openhands-trajectories>,
+CC BY 4.0).
+
+Contains information from WildChat-4.8M
+(<https://huggingface.co/datasets/allenai/WildChat-4.8M>) which is made
+available under the ODC Attribution License
+(<https://opendatacommons.org/licenses/by/1-0/>).
+
+Since `effort-judge-v7` it also includes questions from Stack Exchange sites
+(via <https://huggingface.co/datasets/HuggingFaceTB/stackexchange_2025_md>;
+each post is licensed CC BY-SA 2.5, 3.0 or 4.0 by its author, and its URL and
+author are kept with the training item); GitHub issues and pull requests of
+permissively licensed repositories (via
+<https://huggingface.co/datasets/common-pile/github_archive>, with each item's
+repository, URL, license and authors kept); issues and pull requests from
+Nebius SWE-rebench-V2 (<https://huggingface.co/datasets/nebius/SWE-rebench-V2>,
+<https://huggingface.co/datasets/nebius/SWE-rebench-V2-PRs>, CC BY 4.0,
+permissively licensed repositories only); and agent steps from Nebius
+SWE-agent trajectories
+(<https://huggingface.co/datasets/nebius/SWE-agent-trajectories>, CC BY 4.0),
+Nebius SWE-rebench OpenHands trajectories
+(<https://huggingface.co/datasets/nebius/SWE-rebench-openhands-trajectories>,
+CC BY 4.0), Kwai-Klear SWE-smith mini-swe-agent-plus trajectories
+(<https://huggingface.co/datasets/Kwai-Klear/SWE-smith-mini_swe_agent_plus-trajectories-66k>,
+MIT License), SWE-Factory DeepSWE-Agent Kimi-K2 trajectories
+(<https://huggingface.co/datasets/SWE-Factory/DeepSWE-Agent-Kimi-K2-Trajectories-2.8K>,
+MIT License), SWE-Gym OpenHands SFT trajectories
+(<https://huggingface.co/datasets/SWE-Gym/OpenHands-SFT-Trajectories>, MIT
+License), SWE-ZERO-12M trajectories
+(<https://huggingface.co/datasets/AlienKevin/SWE-ZERO-12M-trajectories>, Apache
+License 2.0), SWE-Lego real and synthetic data
+(<https://huggingface.co/datasets/Lego-X/SWE-Lego-Real-Data>,
+<https://huggingface.co/datasets/Lego-X/SWE-Lego-Synthetic-Data>, Apache
+License 2.0) and TIGER-Lab SWE-Next SFT trajectories
+(<https://huggingface.co/datasets/TIGER-Lab/SWE-Next-SFT-Trajectories>, MIT
+License). Each `effort-judge-v*` release lists its files, hashes and
+attributions.
 
 ## CC0-1.0
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "1.3.0",
+  "version": "1.3.1",
   "private": false,
   "type": "module",
   "description": "Open-source coding agent that matches Codex on Terminal-Bench 2.1 with 63% fewer tokens.",

package/src/defaults/shipped-history.json

@@ -116,6 +116,7 @@
       "8d343d0a406fd5817b7fd98a3ac47010",
       "a1b73ce404125d3ce0a5762e4a1e1555",
       "a28fb01bb21ceebbde1bc6c4a02d3db6",
+      "c7fdabb946eed7283db656493d95749b",
       "d40c23517c4803afdac3fc056bb65cda",
       "db1af5417849abc2f6bb1d1d2b63939f"
     ],
@@ -344,6 +345,7 @@
       "0d6926aaca2890da9347f15a826546e3",
       "0f14ba0a3d8b3896b717dc144bbe49a4",
       "15d86146cfbd58d85adf0c25cf1d683f",
+      "18dc665b96ee0ba5a1645180e173b1af",
       "1c317db86e1ff80a5c4102e09e36f5fb",
       "242f11864abdef90fa0cc9cb4599a579",
       "29a347c4d3a2f773fa92c627260301ba",
@@ -374,7 +376,8 @@
       "ff18ac1a9f5961925aec27ffb870bae0"
     ],
     "setup-guide": [
-      "633e15b84dbe22f2b578b67cc35853d4"
+      "633e15b84dbe22f2b578b67cc35853d4",
+      "920d5b29ccfe6c88f623cf0892b29285"
     ],
     "skill-creator": [
       "020817866da73a69544ba35086bdcd5f",

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

@@ -52,7 +52,7 @@ Read this when choosing a `setup` mutation.
 | Remove stored provider authentication (one OAuth account at a time) | `forget_provider_auth` |
 | 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` |
-| Connect a catalog service (Slack, Notion, GitHub, …) for the user: shows a one-click connect card | `connect_connector` |
+| Connect a catalog service (Notion, Linear, Gmail, …; not GitHub, which is the GitHub CLI) for the user: shows a one-click connect card | `connect_connector` |
 | Add an MCP server | `add_mcp_server` |
 | Edit or rename an MCP server | `save_mcp_server` |
 | Remove an MCP server | `remove_mcp_server` |
@@ -112,6 +112,9 @@ workflow` before changing the corresponding domain.
   `originalName` and `name`. Skill deletion is not exposed; disable it instead.
   Built-in and plugin skills accept only `toolDependencies` edits (`null`
   restores their declared dependencies).
+- A created agent is a bot (`definition.kind: "bot"`) or a workflow agent
+  (`"agent"`); omitted, the representative bot's chat creates a bot and any other chat a
+  workflow agent. The kind is fixed after creation.
 - Changing orchestration does not rewrite workflow instructions or agents.
 
 ## Providers and authentication

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

@@ -16,7 +16,7 @@ through file edits. One named setting the user already chose belongs to `setup`.
 
 ## Style
 
-Reply as in agent mode: one step or one concept per message, short plain
+Reply as in bot mode: one step or one concept per message, short plain
 sentences, no tables or headings, in the user's language. In Korean use the
 formal polite style (~합니다 / ~하세요, never ~해요). Say "봇" for Mixy and
 other contacts, "에이전트" only for workflow agents. Never print ids, file
@@ -67,8 +67,8 @@ connector links an outside service so bots and agents can read and act in it,
 and anything that sends or posts asks first through an approval card. Then
 suggest by what the user does, only these: Gmail (with a Google app password;
 it needs 2-Step Verification on the Google account, and the connect window
-opens the app-password page in the browser), GitHub (comes with Git & GitHub),
-Notion, Linear or Jira. Do not suggest Google Calendar or Google
+opens the app-password page in the browser), Notion, Linear or Jira. GitHub is
+not a connector; it is the signed-in GitHub CLI of Git & GitHub. Do not suggest Google Calendar or Google
 Drive; their Google sign-in is still under review. Connect one with
 `connect_connector`; it only leaves a card, so tell the user to press it and
 confirm with `status mcp`.
@@ -142,7 +142,7 @@ step; do not insist.
 
 ## Glossary
 
-- Bot (봇): an agent-mode contact you chat with. Mixy is the representative
+- Bot (봇): a contact you chat with in bot mode (봇 모드). Mixy is the representative
   bot; the user can have it make others, such as a GitHub Watcher.
 - Agent (에이전트): a workflow sub-agent such as Worker or Reviewer that
   does part of a task. Users add their own on the Workflows page with the
@@ -150,12 +150,15 @@ step; do not insist.
 - Default agents (기본 에이전트): built-in helpers listed on Workflows, such as
   Web Search (runs the web search tool) and the Maintainer (memory upkeep and
   session titles). They need a model route to work.
-- Sessions and projects: a coding session runs in a Project, a folder whose
+- Agent mode (에이전트 모드): a coding work session run by Lead and agents.
+  A session a bot starts for a job is an agent task session (에이전트 작업
+  세션); others are just sessions (세션).
+- Sessions and projects: an agent-mode session runs in a Project, a folder whose
   instructions and memories it loads.
 - Workflows and work rules: a workflow sets how a request gets done; work
   rules are the instructions it applies.
-- Connectors and extensions: connectors are MCP services such as GitHub or
-  Notion (`connect_connector`, the user presses the card); extensions are
+- Connectors and extensions: connectors are MCP services such as Notion or
+  Linear (`connect_connector`, the user presses the card); extensions are
   plugins that add tools and skills that teach how to work.
 - Group rooms (그룹 대화방): one chat with several bots. `@name` or "봇 부르기"
   addresses one bot; otherwise the Maintainer decides who answers next.

package/src/output-styles/common.md

@@ -20,15 +20,20 @@ partial: true
   destructive-action confirmations; quote needed error text exactly.
 
 Layout (GitHub-flavored Markdown, `$$…$$` for display math): the content picks
-the format at every depth; depth sets only how much is said.
+the format at every depth; depth sets only how much is said. There is no fixed
+template: a reply's shape follows what it says, so consecutive replies look
+different when their content differs.
 
-- Mixed content gets one short block per kind (result, cause, finding,
-  decision needed, remaining work) under a plain label in the reader's
-  language, never one running paragraph; what the reader must decide or do
-  comes last, on its own.
-- A single answer or cause is short prose paragraphs without labels. Three or
-  more parallel items are a flat one-line list, numbered only for order, never
-  nested. A comparison across items is a small table with short cells.
+- An explanation, a single answer, or a cause and its effect is short prose.
+- Items that share two or more attributes (status of several tasks, options
+  and their costs, before/after, provider × behavior) are a small table with
+  short cells, not a list of labelled bullets.
+- Ordered steps are a numbered list; three or more parallel, independent
+  points are a flat list; never nested.
+- Mixed content opens with one sentence carrying the outcome, then uses
+  whichever of the forms above fits each part. Plain labels only separate
+  parts that are long and genuinely different; never label every short part.
+- What the reader must decide or do comes last, on its own.
 - Commands, paths, identifiers, and literals go in code; multi-line code,
   diffs, and output in fenced blocks.
 - Bold only what the reader scans for: the decision or warning they must not

package/src/rules/agent-mode/AGENT_MODE.md

@@ -1,4 +1,4 @@
-# Agent Mode
+# Bot Mode
 
 - Speak naturally in your persona, like a capable colleague texting: plain
   everyday sentences, technical detail only when asked. No opener, filler,

package/src/rules/agent-mode/MIXY.md

@@ -12,10 +12,17 @@
   requested scope and existing permissions. You alone create, edit or remove
   bots, and you manage every bot's routines and settings; other bots manage
   only their own.
-- Users call agent-mode contacts (you and the ones made for them) "bots"
+- Users call bot-mode contacts (you and the ones made for them) "bots"
   (봇), and the workflow sub-agents (Worker, Reviewer, …) "agents"
-  (에이전트); use those words with the user.
+  (에이전트). Chatting with a bot is "bot mode" (봇 모드); a coding work
+  session run by Lead and agents is "agent mode" (에이전트 모드). A session
+  you start with `session create` is an "agent task session" (에이전트 작업
+  세션); any other is a "session" (세션). Use those words with the user.
 - You coordinate the user's bots and agents: route a request to the one best
-  suited for it (a bot, a sub-agent, or a new coding-mode task session for
+  suited for it (a bot, a sub-agent, or a new agent task session for
   substantial coding) with a self-contained brief, start it when already
-  requested, otherwise suggest it, and follow up with the result.
+  requested, otherwise suggest it, and follow up with the result. Promise a
+  follow-up only for work that reports back: a bot, a sub-agent or an agent
+  task session you created, or a `session send` whose result says
+  reportsBack: true. Otherwise say it was sent and that the user can check
+  that session.

package/src/runtime/agent/orchestrator/context/role-instructions.mjs

@@ -2,8 +2,9 @@
 
 import { createHash } from 'node:crypto';
 import { existsSync, readdirSync } from 'node:fs';
-import { join } from 'node:path';
+import { basename, dirname, join } from 'node:path';
 import { compareCodePoints } from '../../../shared/code-point-order.mjs';
+import { BOTS_DIR, MIXY_BOT_ID, botStoreActive, isStoreBot, mixyIdentityDir } from '../../../shared/agent-store.mjs';
 import { mixdogRoot } from '../../../shared/plugin-paths.mjs';
 import { readMarkdownDocument } from '../../../shared/markdown-frontmatter.mjs';
 import { listHiddenAgentsByKind, getAgentCatalogShareAgents } from '../internal-agents.mjs';
@@ -89,9 +90,35 @@ function agentSectionDirs(pluginRoot) {
   } catch {
     /* unresolvable data dir — built-in roles still load */
   }
+  // Bots of the store (agent-store.mjs), last: a bot's own folder wins over
+  // the agents/ original the split left behind.
+  const root = userDataRoot();
+  if (root && botStoreActive(root)) dirs.push(join(root, BOTS_DIR));
   return dirs;
 }
 
+// The user data dir (parent of agents/), or '' when it cannot be resolved.
+function userDataRoot() {
+  try {
+    const userDir = mixdogGlobalDir('agents');
+    return userDir ? dirname(userDir) : '';
+  } catch {
+    return '';
+  }
+}
+
+// A store bot's own folder, or ''.
+function storeBotDir(agent) {
+  const root = userDataRoot();
+  return root && isStoreBot(root, agent) ? join(root, BOTS_DIR, agent) : '';
+}
+
+// Mixy's identity folder: bots/mixy with the store, agents/lead before.
+function mixyDir() {
+  const root = userDataRoot();
+  return root ? mixyIdentityDir(root) : '';
+}
+
 function loadAgentSections(pluginRoot) {
   // agents/ accepts both the compatibility flat layout and the current
   // nested agents/<agent>/AGENT.md layout.
@@ -117,6 +144,7 @@ function loadAgentSections(pluginRoot) {
       // Mixy's own AGENT.md (agents/lead/) rides only Mixy's agent-mode
       // prompt (loadMixyRoleSection), never another role catalog.
       if (!name || !raw || name === 'lead') continue;
+      if (name === MIXY_BOT_ID && basename(agentsDir) === BOTS_DIR) continue;
       const { body } = readMarkdownDocument(raw);
       if (!body) continue;
       byName.set(name, roleSection(name, body));
@@ -255,6 +283,12 @@ export function truncatePersona(text, max = PERSONA_MAX_CHARS) {
 export function loadAgentPersonaSection(agent) {
   if (typeof agent !== 'string' || !PERSONA_AGENT_ID.test(agent) || agent.includes('..')) return '';
   const isLead = agent === 'lead';
+  // Mixy's persona is in its own folder; a store bot's in its bots/ folder.
+  const own = isLead ? mixyDir() : storeBotDir(agent);
+  if (own) {
+    const persona = truncatePersona(readSafe(join(own, 'PERSONA.md')) || '');
+    return persona ? personaSection(persona) : '';
+  }
   const pluginRoot = mixdogRoot();
   const roots = [];
   try {
@@ -272,17 +306,21 @@ export function loadAgentPersonaSection(agent) {
       if (isLead) continue;
       return '';
     }
-    return [
-      '## Persona',
-      '',
-      'Tone and style guide for how you speak, address the user and carry yourself. It never overrides tool, safety or role rules; those always win.',
-      '',
-      persona,
-    ].join('\n');
+    return personaSection(persona);
   }
   return '';
 }
 
+function personaSection(persona) {
+  return [
+    '## Persona',
+    '',
+    'Tone and style guide for how you speak, address the user and carry yourself. It never overrides tool, safety or role rules; those always win.',
+    '',
+    persona,
+  ].join('\n');
+}
+
 // --- Mixy's working style (agents/lead/AGENT.md, user data dir) ---
 // The user-editable 일하는 방식 of Mixy, framed like any agent's role section.
 // Agent mode only; empty or missing means no section. Re-read per
@@ -290,8 +328,8 @@ export function loadAgentPersonaSection(agent) {
 export function loadMixyRoleSection() {
   let raw = '';
   try {
-    const userDir = mixdogGlobalDir('agents');
-    raw = userDir ? readSafe(join(userDir, 'lead', 'AGENT.md')) || '' : '';
+    const dir = mixyDir();
+    raw = dir ? readSafe(join(dir, 'AGENT.md')) || '' : '';
   } catch {
     return '';
   }
@@ -306,6 +344,15 @@ export function loadMixyRoleSection() {
 // (the Lead configuration, id `lead`) has its own line, under the display name
 // the user gave it (session/mixy-name.mjs), Mixy by default.
 function agentManifest(agent) {
+  const own = storeBotDir(agent);
+  if (own) {
+    try {
+      const value = JSON.parse(readSafe(join(own, 'agent.json')) || '{}');
+      return value && typeof value === 'object' ? value : {};
+    } catch {
+      return {};
+    }
+  }
   const roots = [];
   try {
     const userDir = mixdogGlobalDir('agents');
@@ -344,9 +391,13 @@ export function agentModeIdentityLine(agent) {
 // the shipped tree — as one hash. A session stamps it when its prompt is
 // composed; another value means the profile was edited since
 // (session-runtime/tool-policy-refresh.mjs refreshAgentProfile). '' for an
-// invalid id.
-export function agentProfileRevision(agent) {
+// invalid id. A coding Lead (`lead` outside Mixy's own thread) embeds no
+// profile file at all, so its stamp is a constant: editing Mixy never
+// rebuilds a coding session's prompt.
+export const CODING_LEAD_PROFILE_REVISION = 'coding-lead';
+export function agentProfileRevision(agent, { mixy = true } = {}) {
   if (typeof agent !== 'string' || !PERSONA_AGENT_ID.test(agent) || agent.includes('..')) return '';
+  if (agent === 'lead' && !mixy) return CODING_LEAD_PROFILE_REVISION;
   const roots = [mixdogGlobalDir('agents')];
   const pluginRoot = mixdogRoot();
   if (pluginRoot) roots.push(join(pluginRoot, 'agents'));
@@ -356,6 +407,13 @@ export function agentProfileRevision(agent) {
       hash.update(`${root}\0${file}\0${readSafe(join(root, agent, file)) ?? '\u0001'}\0`);
     }
   }
+  // Mixy's or a store bot's own folder, when it has one.
+  const own = agent === 'lead' ? mixyDir() : storeBotDir(agent);
+  if (own && own !== join(roots[0] || '', agent)) {
+    for (const file of ['agent.json', 'AGENT.md', 'PERSONA.md']) {
+      hash.update(`${own}\0${file}\0${readSafe(join(own, file)) ?? '\u0001'}\0`);
+    }
+  }
   return hash.digest('hex').slice(0, 24);
 }
 
@@ -417,8 +475,8 @@ export function composeSystemPrompt(opts) {
         ? loadMixyRoleSection()
         : loadScopedRoleInstructions(opts.agent || null, opts.provider || null);
   }
-  // Persona follows the role; the Lead has no role section but may own a
-  // persona through agents/lead/PERSONA.md.
+  // Persona follows the role; Mixy's own thread carries Mixy's persona
+  // (bots/mixy/PERSONA.md) — a coding Lead's prompt never does (skipPersona).
   const personaContext = opts.agent && !opts.skipPersona ? loadAgentPersonaSection(opts.agent) : '';
   const coreMemory = skip.memory ? '' : trimmedPromptSlice(opts.coreMemoryContext);
   // User AGENTS.md / CLAUDE.md block (context/instruction-files.mjs), already

package/src/runtime/agent/orchestrator/mcp/agent-connectors.mjs

@@ -14,6 +14,7 @@ import { readFileSync, statSync } from 'node:fs';
 import { join } from 'node:path';
 import { cwdWithinProjects } from '../../../shared/extension-scopes.mjs';
 import { resolvePluginData } from '../../../shared/plugin-paths.mjs';
+import { userDefinitionDir } from '../../../shared/agent-store.mjs';
 import { isSelfMcpToolName, parseMcpToolName } from '../../../shared/tool-primitives.mjs';
 import { connectorToolDefault } from '../../../shared/connector-catalog.mjs';
 import { serverConnector } from './server-connectors.mjs';
@@ -149,7 +150,7 @@ export function readConnectorsFile(path) {
 export function agentConnectors(agentId, dataDir = resolvePluginData()) {
   const id = clean(agentId);
   if (!id || !AGENT_DIR_ID.test(id)) return null;
-  return cachedRead(join(dataDir, 'agents', id, 'agent.json'), readConnectorsFile);
+  return cachedRead(join(userDefinitionDir(dataDir, id), 'agent.json'), readConnectorsFile);
 }
 
 /** Uncached: the project policy file → `[{ root, off }]`, null without one,

package/src/runtime/agent/orchestrator/providers/antigravity-oauth-catalog.mjs

@@ -187,9 +187,14 @@ export function normalizeAntigravityCatalog(rawModels) {
   return models.sort((a, b) => a.display.localeCompare(b.display));
 }
 
-/** Catalog to resolve wire ids against: in-memory mirror, disk cache, then the curated fallback. */
+/**
+ * Catalog to resolve wire ids against: in-memory mirror, disk cache, then the
+ * curated fallback. Sends never refresh the catalog, so the account's last live
+ * list is used even past its TTL: it names the wire ids this account was
+ * offered, which the curated list can only approximate.
+ */
 function antigravityCatalogModels() {
-  if (!_mirror) _mirror = antigravityModelCache.loadSync();
+  if (!_mirror) _mirror = antigravityModelCache.loadSync({ allowStale: true });
   return _mirror || ANTIGRAVITY_MODELS;
 }
 

package/src/runtime/agent/orchestrator/providers/auto-effort.mjs

@@ -31,14 +31,14 @@ export const AUTO_EFFORT_LEVELS = Object.freeze(['easy', 'normal', 'hard', 'very
 const STEP_OF_LEVEL = [-1, 0, 1, 2];
 
 // The chosen level minimises the expected cost under the judge distribution.
-// A miss costs one per level, plus extra for an opposite call and for a
-// needless "very hard" (xhigh). Opposite calls are asymmetric: "easy" for a
-// hard turn degrades the answer, "hard" for an easy turn only costs tokens.
-// On the held-out test splits the heavier under-call cost cut hard-as-easy
-// calls by a third (20 -> 13 of 7,180) for -0.2 points of exact accuracy.
+// A miss costs one per level, plus extra for an opposite call. Opposite calls
+// are asymmetric: "easy" for a hard turn degrades the answer, "hard" for an
+// easy turn only costs tokens. On the held-out test splits the heavier
+// under-call cost cut hard-as-easy calls by a third (20 -> 13 of 7,180) for
+// -0.2 points of exact accuracy. "Very hard" carries no extra cost: an extra
+// penalty there let judge v7 pick xhigh on only 17% of very hard turns.
 export const AUTO_EFFORT_UNDER_OPPOSITE_COST = 3;
 export const AUTO_EFFORT_OVER_OPPOSITE_COST = 1;
-export const AUTO_EFFORT_FALSE_VERY_HARD_COST = 1;
 // Ties keep the default first, then the cheaper side.
 const TIE_ORDER = [1, 0, 2, 3];
 
@@ -49,8 +49,7 @@ function expectedCost(values, level) {
       p *
       (Math.abs(level - actual) +
         (level === 0 && actual >= 2 ? AUTO_EFFORT_UNDER_OPPOSITE_COST : 0) +
-        (actual === 0 && level >= 2 ? AUTO_EFFORT_OVER_OPPOSITE_COST : 0) +
-        (level === 3 && actual < 3 ? AUTO_EFFORT_FALSE_VERY_HARD_COST : 0));
+        (actual === 0 && level >= 2 ? AUTO_EFFORT_OVER_OPPOSITE_COST : 0));
   });
   return cost;
 }

package/src/runtime/agent/orchestrator/providers/lib/anthropic-request-utils.mjs

@@ -10,6 +10,7 @@ import { mergeFlatSchemaProperty } from './flat-schema-merge.mjs';
 import { reportedCostUsd } from './usage-primitives.mjs';
 import { reportInvalidTool, wireInputSchema } from './claude-bridge-schema.mjs';
 import { normalizeServerToolUse } from './anthropic-server-tool-usage.mjs';
+import { ANTHROPIC_TOOL_CALL_IDS, withWireToolCallIds } from './tool-call-wire-ids.mjs';
 export const ANTHROPIC_CACHE_TTL_STABLE = { type: 'ephemeral', ttl: '1h' };
 const ANTHROPIC_CACHE_TTL_VOLATILE = { type: 'ephemeral' };
 
@@ -232,7 +233,8 @@ const ANTHROPIC_NATIVE_PROVIDERS = new Set(['anthropic', 'anthropic-oauth']);
 
 // `provider`: the engine's provider name. A gateway on a native Messages
 // route (wire-route.mjs) also replays the references it tagged itself.
-export function toAnthropicMessages(messages, availableTools, provider = '') {
+export function toAnthropicMessages(transcript, availableTools, provider = '') {
+  const messages = withWireToolCallIds(transcript, ANTHROPIC_TOOL_CALL_IDS);
   const ownProvider = String(provider || '').toLowerCase();
   const availableNames = anthropicToolNameSet(availableTools);
   const result = [];

package/src/runtime/agent/orchestrator/providers/lib/tool-call-wire-ids.mjs

@@ -0,0 +1,65 @@
+// Tool-call ids at the wire. A transcript keeps each tool call's id as the
+// provider that produced it minted it; after a model switch another provider
+// may refuse it (Anthropic accepts only [A-Za-z0-9_-]; OpenAI chat caps the
+// length). Each request builder projects the ids onto its own rule just
+// before lowering: a valid id passes unchanged (so native replay items and
+// the provider's own ids stay byte-exact), an invalid one becomes a
+// deterministic, collision-free stand-in, and a call and its result always
+// map to the same value. Persisted history is never rewritten.
+import { createHash } from 'node:crypto';
+
+export const ANTHROPIC_TOOL_CALL_IDS = Object.freeze({ allowed: /^[A-Za-z0-9_-]+$/, maxLength: 64 });
+export const OPENAI_RESPONSES_TOOL_CALL_IDS = Object.freeze({ allowed: /^[A-Za-z0-9_.:-]+$/, maxLength: 64 });
+export const OPENAI_CHAT_TOOL_CALL_IDS = Object.freeze({ allowed: /^[A-Za-z0-9_.:-]+$/, maxLength: 40 });
+
+const HASH_CHARS = 10;
+
+function standIn(id, rule, salt) {
+  const hash = createHash('sha256').update(`${salt}\u0000${id}`).digest('hex').slice(0, HASH_CHARS);
+  const stem = String(id)
+    .replace(/[^A-Za-z0-9_-]+/g, '_')
+    .replace(/^_+|_+$/g, '')
+    .slice(0, Math.max(0, rule.maxLength - HASH_CHARS - 1));
+  return stem ? `${stem}_${hash}` : `call_${hash}`.slice(0, rule.maxLength);
+}
+
+const isValid = (id, rule) => id.length > 0 && id.length <= rule.maxLength && rule.allowed.test(id);
+
+/** id → wire id for every tool call and result in `messages` under `rule`. */
+export function toolCallWireIdMap(messages, rule) {
+  const ids = new Set();
+  for (const message of Array.isArray(messages) ? messages : []) {
+    if (message?.role === 'assistant')
+      for (const call of message.toolCalls || []) if (call?.id) ids.add(String(call.id));
+    if (message?.role === 'tool' && message.toolCallId) ids.add(String(message.toolCallId));
+  }
+  const map = new Map();
+  const taken = new Set([...ids].filter((id) => isValid(id, rule)));
+  for (const id of ids) {
+    if (isValid(id, rule)) continue;
+    let salt = 0;
+    let wire = standIn(id, rule, salt);
+    while (taken.has(wire)) wire = standIn(id, rule, ++salt);
+    taken.add(wire);
+    map.set(id, wire);
+  }
+  return map;
+}
+
+/** `messages` with tool-call ids projected onto `rule`; the same array when none needs it. */
+export function withWireToolCallIds(messages, rule) {
+  const map = toolCallWireIdMap(messages, rule);
+  if (!map.size) return messages;
+  return messages.map((message) => {
+    if (message?.role === 'tool' && map.has(String(message.toolCallId))) {
+      return { ...message, toolCallId: map.get(String(message.toolCallId)) };
+    }
+    if (message?.role !== 'assistant' || !message.toolCalls?.some((call) => map.has(String(call?.id)))) return message;
+    return {
+      ...message,
+      toolCalls: message.toolCalls.map((call) =>
+        map.has(String(call?.id)) ? { ...call, id: map.get(String(call.id)) } : call
+      ),
+    };
+  });
+}

package/src/runtime/agent/orchestrator/providers/openai-compat-retry-profile.mjs

@@ -104,10 +104,16 @@ export const COMPAT_RETRY_PROFILES = Object.freeze({
   }),
 });
 
-/** withRetry options (attempt budget, schedule, jitter) for a provider; {} for the shared defaults. */
+// A provider without a profile (custom endpoints, gateways, local servers)
+// keeps the shared ladder, but a Retry-After beyond a minute is not waited
+// out: it surfaces with its wait so the account pool can rotate and the
+// notice can name when to try again.
+export const COMPAT_DEFAULT_MAX_RETRY_AFTER_MS = 60_000;
+
+/** withRetry options (attempt budget, schedule, jitter) for a provider; only the advice ceiling for the shared defaults. */
 export function compatRetryOptions(providerName) {
   const profile = COMPAT_RETRY_PROFILES[providerName];
-  if (!profile) return {};
+  if (!profile) return { maxRetryAfterMs: COMPAT_DEFAULT_MAX_RETRY_AFTER_MS };
   return {
     maxAttempts: profile.maxAttempts,
     backoffMs: profile.backoffMs,

package/src/runtime/agent/orchestrator/providers/openai-compat-wire.mjs

@@ -17,6 +17,11 @@ import {
 } from './media-normalization.mjs';
 import { customToolCallFromResponseItem, nativeToolSearchCallFromArguments } from './custom-tool-wire.mjs';
 import { providerReplayItems } from './lib/provider-replay.mjs';
+import {
+  OPENAI_CHAT_TOOL_CALL_IDS,
+  OPENAI_RESPONSES_TOOL_CALL_IDS,
+  withWireToolCallIds,
+} from './lib/tool-call-wire-ids.mjs';
 import { acceptedWireTools } from './lib/claude-bridge-schema.mjs';
 import { ensureResponsesCallOutputs } from './lib/wire-pairing.mjs';
 import { responsesEffortUpdates } from './effort-configuration.mjs';
@@ -34,7 +39,8 @@ export function resolveCompatMaxOutputTokens(opts = {}) {
   );
 }
 
-export function toOpenAIMessages(messages, providerName, options = {}) {
+export function toOpenAIMessages(transcript, providerName, options = {}) {
+  const messages = withWireToolCallIds(transcript, OPENAI_CHAT_TOOL_CALL_IDS);
   // NOTE: chat.completions has no equivalent slot for replaying reasoning
   // encrypted_content the way the Responses API does (no `type:'reasoning'`
   // input item). Whatever reasoningItems may be attached to assistant
@@ -421,7 +427,8 @@ export function xaiSystemInstructions(messages) {
   return instructions || undefined;
 }
 
-export function toXaiResponsesInput(messages, providerState, options = {}) {
+export function toXaiResponsesInput(transcript, providerState, options = {}) {
+  const messages = withWireToolCallIds(transcript, OPENAI_RESPONSES_TOOL_CALL_IDS);
   const includeSystem = options.includeSystem !== false;
   // Generic Responses callers (opencode-go gateway) keep their own state
   // slot and replay tag so a provider switch never replays foreign items.

package/src/runtime/agent/orchestrator/providers/openai-responses-input.mjs

@@ -1,6 +1,7 @@
 // Transcript lowering is independent of request/turn metadata. A retained
 // message must have the same wire representation in every subsequent request.
 import { providerReplayItems } from './lib/provider-replay.mjs';
+import { OPENAI_RESPONSES_TOOL_CALL_IDS, withWireToolCallIds } from './lib/tool-call-wire-ids.mjs';
 import { responsesEffortUpdates } from './effort-configuration.mjs';
 import { ensureResponsesCallOutputs } from './lib/wire-pairing.mjs';
 import { normalizeContentForOpenAIResponses, splitToolContentForOpenAIResponses } from './media-normalization.mjs';
@@ -147,7 +148,8 @@ function pushAssistantToolCalls(toolCalls, { out, customToolCallIds, nativeSearc
  *  - `includeSystem`, `nativePdf`, `splitToolContent`, `startIndex`: the
  *    gateway's system/media/continuation handling.
  */
-export function convertMessagesToResponsesInput(messages, opts = {}) {
+export function convertMessagesToResponsesInput(transcript, opts = {}) {
+  const messages = withWireToolCallIds(transcript, OPENAI_RESPONSES_TOOL_CALL_IDS);
   const out = [];
   const pendingToolMedia = [];
   const pendingToolLoads = [];