severally-mcp 2.4.2 → 2.5.0

package/CONFIG.md

@@ -78,6 +78,7 @@ A name is lower-case letters, digits, `.` and `-`, at most 64 characters; `all`
 | `effort` | string | `high` | Reasoning effort, translated per CLI: Codex `-c model_reasoning_effort=…`, Claude Code `--effort …`, OpenCode the model variant `provider/model#…`. A value the CLI or model does not support fails the consultation as `model_unavailable` (e.g. glm-5.2 has `high`/`max` only). For OpenCode, a model written with its own `#variant` keeps it, and one written with a bare trailing `#` (`opencode/nemotron-3-ultra-free#`) runs with no variant, for models that have none. Not allowed on an Antigravity consultant: agy takes the effort from the model name (`gemini-3.8-flash-low` / `-medium` / `-high`). |
 | `timeout_ms` | number, 1000–7200000 | `server.timeout_ms` | This consultant's own wall-clock budget, for the one that reliably needs longer or shorter than the rest. |
 | `max_budget_usd` | number, 0.05–20 | `10` | Claude Code consultants only: the spending cap for one consultation. |
+| `fallback_model` | string | – | A model on the same CLI to run instead when `model`'s account quota is used up (`usage_limit`). The consultation is run again once, inside the same time budget, and the answer says that the fallback gave it; `model` on the result is the fallback, and the switch is listed under `retries`. If the fallback fails too, that failure is reported. Must differ from `model`. |
 | `vendor` | string | from the model name | The lineage, for the same-lineage note on an answer. Read from the model name (`gpt` → openai, `claude` → anthropic, `gemini` → google, `glm` → zai, `deepseek` → deepseek, `nemotron` → nvidia); set it for a model the name does not place. |
 | `enabled` | boolean | `true` | Turns this one consultant off, leaving the others on its CLI. |
 | `note` | string | – | Why it is off, as for `clis`. |

package/README.ja.md

@@ -90,7 +90,7 @@ flowchart TB
 macOS、Linux、Windows（PowerShell またはコマンドプロンプト）のどれでも、次の 1 行で入れられます。
 
 ```bash
-npx -y severally-mcp@2.4.2 install
+npx -y severally-mcp@2.5.0 install
 ```
 
 インストーラは、`PATH` で見つかった CLI のそれぞれに severally を登録します。末尾に `--dry-run` を付けると、何も

package/README.md

@@ -92,7 +92,7 @@ routes and the real CLIs have been tried on Linux only.
 One line, on macOS, Linux and Windows (PowerShell or Command Prompt):
 
 ```bash
-npx -y severally-mcp@2.4.2 install
+npx -y severally-mcp@2.5.0 install
 ```
 
 The installer registers severally with each CLI it finds on `PATH`. Add `--dry-run` at the end to print what it

package/config.example.json

@@ -17,7 +17,7 @@
     "opus": { "cli": "claude-code", "model": "claude-opus-5-5", "effort": "max", "max_budget_usd": 5, "timeout_ms": 900000 },
     "fable": { "cli": "claude-code", "model": "claude-fable-5-1" },
     "sol": { "cli": "codex", "model": "gpt-6.1-sol" },
-    "gemini": { "cli": "antigravity", "model": "gemini-4-argon-high" },
+    "gemini": { "cli": "antigravity", "model": "gemini-4-argon-high", "fallback_model": "gemini-3.8-flash-high" },
     "glm": { "cli": "opencode", "model": "zai-coding-plan/glm-5.3" },
     "deepseek": { "cli": "opencode", "model": "opencode/deepseek-v4.1-flash-free", "enabled": false, "note": "trial ended" },
     "nemotron": { "cli": "opencode", "model": "opencode/nemotron-3-ultra-free#", "vendor": "nvidia" }

package/docs/how-it-works.ja.md

@@ -37,6 +37,8 @@ OpenCode    ──(skill)──> severally server ──> codex exec | claude -p
 - 相談を始めても、主担当は相談している計画そのものは進めません。
 - 一時的な失敗（`rate_limited`、`cli_error`）は、少し待ってから、同じ時間の枠の中で、同じ相談相手とモデルで最大 2 回
   やり直します。それまでの失敗は `retries` に残ります。
+- 相談相手のモデルがアカウントの使用枠を使い切ったとき（`usage_limit`）は、設定に `fallback_model` があれば、その
+  モデルで 1 回だけやり直します。答えには代わりのモデルが答えたことが書かれ、`model` には実際に答えたモデルが入ります。
 - 時間切れになったときは、相談相手がそこまでに書いていた内容の末尾を、途中までのものという印を付けて失敗の隣に
   残します。これは読む材料として残すもので、答えとしては扱いません。
 - 相談の結果は、サーバを再起動したあとも、同じマシンで動くほかの severally のプロセスからも読めます。ほかのプロセスで

package/docs/how-it-works.md

@@ -38,6 +38,8 @@ OpenCode    ──(skill)──> severally server ──> codex exec | claude -p
 - While a consultation runs, the lead does not go ahead with the plan in question.
 - A transient failure (`rate_limited`, `cli_error`) is run again up to twice, after a short wait and inside the
   same time budget, with the same consultant and model; the earlier failures are listed under `retries`.
+- A consultant whose model has used up its account quota (`usage_limit`) is run again once on its `fallback_model`,
+  if the config names one. The answer says that the fallback gave it, and `model` shows the one that answered.
 - A consultant that runs out of time leaves the end of what it had written beside the failure, marked partial:
   something to read, not an answer.
 - A consultation stays readable after the server restarts, and from any other severally process on the machine.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "severally-mcp",
-  "version": "2.4.2",
+  "version": "2.5.0",
   "description": "A plugin that lets your AI agent ask other coding agents for independent opinions, returned separately.",
   "type": "module",
   "license": "Apache-2.0",

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

@@ -1,7 +1,7 @@
 {
   "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
   "name": "severally",
-  "version": "2.4.2",
+  "version": "2.5.0",
   "description": "Ask Codex, Claude Code, Antigravity or OpenCode for an independent opinion, review or structured debate, in a sandboxed child session that can read files and search the web but cannot edit files, change the system, or start other agents.",
   "author": {
     "name": "coz.a",

package/plugins/severally/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "severally",
-  "version": "2.4.2",
+  "version": "2.5.0",
   "description": "Ask Codex, Claude Code, Antigravity or OpenCode for an independent opinion, review or structured debate, in a sandboxed child session that can read files and search the web but cannot edit files, change the system, or start other agents.",
   "author": {
     "name": "coz.a",

package/plugins/severally/dist/severally-mcp.mjs

@@ -863,7 +863,7 @@ function buildConsultants(config2) {
         if (!CONSULTANT_KEYS.includes(k)) problems.push(`${at}.${k}: unknown key (allowed: ${CONSULTANT_KEYS.join(", ")})`);
       if (!cliIds.includes(def.cli)) problems.push(`${at}.cli: must be one of ${CLI_IDS.join(", ")}`);
       if (typeof def.model !== "string" || def.model.trim() === "") problems.push(`${at}.model: required`);
-      for (const k of ["effort", "vendor", "note"]) {
+      for (const k of ["effort", "fallback_model", "vendor", "note"]) {
         if (def[k] !== void 0 && typeof def[k] !== "string") problems.push(`${at}.${k}: must be a string`);
       }
       if (def.enabled !== void 0 && typeof def.enabled !== "boolean") problems.push(`${at}.enabled: must be true or false`);
@@ -871,6 +871,10 @@ function buildConsultants(config2) {
         if (def[k] !== void 0 && !(typeof def[k] === "number" && Number.isFinite(def[k])))
           problems.push(`${at}.${k}: must be a number`);
       }
+      if (typeof def.fallback_model === "string" && def.fallback_model.trim() === "")
+        problems.push(`${at}.fallback_model: must not be empty`);
+      else if (typeof def.fallback_model === "string" && def.fallback_model.trim() === String(def.model).trim())
+        problems.push(`${at}.fallback_model: must differ from model; the fallback is for when the model's quota is used up`);
       if (def.max_budget_usd !== void 0 && def.cli !== "claude-code")
         problems.push(`${at}.max_budget_usd: only a claude-code consultant has a budget`);
       if (def.effort !== void 0 && def.cli === "antigravity")
@@ -905,6 +909,7 @@ function make(name, def) {
     effort: typeof def.effort === "string" && def.effort.trim() !== "" ? def.effort.trim() : null,
     timeoutMs: typeof def.timeout_ms === "number" ? clamp(def.timeout_ms, 1e3, 72e5) : null,
     maxBudgetUsd: typeof def.max_budget_usd === "number" ? clamp(def.max_budget_usd, 0.05, 20) : null,
+    fallbackModel: typeof def.fallback_model === "string" && def.fallback_model.trim() !== "" ? def.fallback_model.trim() : null,
     vendor: vendorOf(model, def.vendor),
     enabled: def.enabled !== false,
     note: typeof def.note === "string" && def.note.trim() !== "" ? def.note.trim() : null
@@ -958,7 +963,7 @@ var init_consultants = __esm({
     };
     TOP_KEYS = ["server", "clis", "consultants", "ranking"];
     CLI_KEYS = ["enabled", "note", "bin", "args", "env", "provider"];
-    CONSULTANT_KEYS = ["cli", "model", "effort", "timeout_ms", "max_budget_usd", "vendor", "enabled", "note"];
+    CONSULTANT_KEYS = ["cli", "model", "effort", "timeout_ms", "max_budget_usd", "fallback_model", "vendor", "enabled", "note"];
     isObject = (v) => v !== null && typeof v === "object" && !Array.isArray(v);
     clamp = (v, min, max) => Math.min(max, Math.max(min, v));
   }
@@ -39946,9 +39951,9 @@ function unreadCaveat(calls, readTools) {
 }
 function refusedToolCaveat(calls, allowed) {
   if (!calls || !allowed) return null;
-  const names = Object.keys(calls).filter((k) => k.endsWith(":error")).map((k) => k.slice(0, -":error".length)).filter((t) => !allowed.includes(t));
+  const names = Object.keys(calls).filter((k) => k.endsWith(":error")).map((k) => k.slice(0, -":error".length)).filter((t) => !allowed.includes(t) && /read|view|open/i.test(t));
   if (!names.length) return null;
-  return `the consultant's CLI called ${names.join(", ")}, which severally does not allow, and was refused; a new CLI release may have renamed a tool`;
+  return `the consultant's CLI called ${names.join(", ")}, which severally does not allow, and was refused; a new CLI release may have renamed its file-reading tool`;
 }
 function sameVendorCaveat(caller, callerModel, consultant) {
   const callerCli = CLI_IDS.find((c) => c === caller);
@@ -40951,20 +40956,26 @@ var init_jobs = __esm({
           this.#mark(job);
           const t0 = Date.now();
           if (adapter.prepare) await adapter.prepare();
-          const invocation = adapter.buildInvocation({
-            workdir,
-            schemaPath,
-            guardrails: renderGuardrails(Boolean(req._snapshot), req.network),
-            sandbox,
-            consultant,
-            hasExposedPaths: Boolean(req._snapshot),
-            network: req.network
-          });
-          const argsProblem = extraArgsProblem(cli, invocation.args, t.extraArgs);
-          if (argsProblem) return this.#fail(job, "spawn_error", `config clis.${cli}.args: ${argsProblem}`);
-          invocation.args = withExtraArgs(invocation.args, t.extraArgs);
-          assertNoForbiddenFlags(cli, invocation.args);
-          job.boundary = adapter.boundary?.({ args: invocation.args, sandbox, network: req.network, hasExposedPaths: Boolean(req._snapshot) }) ?? null;
+          const launch = (c) => {
+            const invocation2 = adapter.buildInvocation({
+              workdir,
+              schemaPath,
+              guardrails: renderGuardrails(Boolean(req._snapshot), req.network),
+              sandbox,
+              consultant: c,
+              hasExposedPaths: Boolean(req._snapshot),
+              network: req.network
+            });
+            const problem = extraArgsProblem(cli, invocation2.args, t.extraArgs);
+            if (problem) return problem;
+            invocation2.args = withExtraArgs(invocation2.args, t.extraArgs);
+            assertNoForbiddenFlags(cli, invocation2.args);
+            job.boundary = adapter.boundary?.({ args: invocation2.args, sandbox, network: req.network, hasExposedPaths: Boolean(req._snapshot) }) ?? null;
+            return invocation2;
+          };
+          const first = launch(consultant);
+          if (typeof first === "string") return this.#fail(job, "spawn_error", `config clis.${cli}.args: ${first}`);
+          let invocation = first;
           const attempt = (timeout) => {
             const { handle, done } = runChild({
               command: invocation.command,
@@ -40988,6 +40999,7 @@ var init_jobs = __esm({
           let interpreted;
           let authTried = false;
           let autoRetries = 0;
+          let fallenBack = false;
           for (; ; ) {
             job.tool_calls = adapter.toolCalls?.(run) ?? null;
             job.duration_ms = Date.now() - t0;
@@ -41044,6 +41056,22 @@ var init_jobs = __esm({
                 continue;
               }
             }
+            if (interpreted.failureKind === "usage_limit" && consultant.fallbackModel && !fallenBack && remaining >= MIN_RETRY_MS) {
+              fallenBack = true;
+              const fallback = consultant.fallbackModel;
+              const next = launch({ ...consultant, model: fallback });
+              if (typeof next === "string") return this.#fail(job, "spawn_error", `config clis.${cli}.args: ${next}`);
+              invocation = next;
+              job.retries.push({
+                failure_kind: interpreted.failureKind,
+                reason: `${job.model} quota used up; re-run on ${fallback}`,
+                at: (/* @__PURE__ */ new Date()).toISOString(),
+                usage: adapter.usageRecord(interpreted.usageRaw)
+              });
+              job.model = fallback;
+              run = await attempt(remaining);
+              continue;
+            }
             const wait = POLICY.retryBackoffMs * 3 ** autoRetries;
             if (AUTO_RETRY_KINDS.includes(interpreted.failureKind) && autoRetries < POLICY.maxAutoRetries && remaining - wait >= MIN_RETRY_MS) {
               autoRetries += 1;
@@ -41101,6 +41129,7 @@ var init_jobs = __esm({
             sameVendorCaveat(job.caller, job.caller_model, consultant),
             interpreted.recoveredError && `the consultant's CLI reported an error mid-run and then finished its answer: ${redact(interpreted.recoveredError.message).slice(0, 300)}`,
             leftover && "the answer parsed, but a process the consultant started kept its output open after it exited and was stopped (one that left its process group may still be running)",
+            fallenBack && `answered by ${job.model}: ${consultant.model}'s quota was used up`,
             req._snapshot && unreadCaveat(job.tool_calls, adapter.FILE_READ_TOOLS),
             refusedToolCaveat(job.tool_calls, adapter.ALLOWED_TOOLS)
           ].filter(Boolean);
@@ -42112,7 +42141,7 @@ var init_server = __esm({
     init_schema();
     init_jobs();
     init_export();
-    SEVERALLY_VERSION = "2.4.2";
+    SEVERALLY_VERSION = "2.5.0";
     DEFAULT_WAIT_MS = process.env.SEVERALLY_MAX_WAIT_MS ? POLICY.maxWaitMs : Math.min(POLICY.maxWaitMs, 45e3);
     PROGRESS_EVERY_MS = 15e3;
     semver = (v) => String(v).split(".").map(Number);

package/plugins/severally/skills/_template/SKILL.md.tmpl

@@ -50,7 +50,8 @@ actually reach. When the user names one, pass **the word the user used** as `tar
 
 **Do not pick a model yourself.** When the user named no one, ask everyone; overriding the operator's
 consultants on your own judgement — including after a `usage_limit` failure — substitutes a different mind
-for the one the user asked for.
+for the one the user asked for. The operator may have set a fallback model for that case; a caveat reading
+"answered by X: Y's quota was used up" means it answered, and your report names X as the one who answered.
 
 Consult your own CLI (`target: "{{SELF_TARGET}}"`) freely, alone or alongside the peers, and the same goes
 for a **different model of your own lineage** when the user wants one (an Opus lead asking Fable). Either is a

package/plugins/severally/skills/antigravity/severally/SKILL.md

@@ -61,7 +61,8 @@ actually reach. When the user names one, pass **the word the user used** as `tar
 
 **Do not pick a model yourself.** When the user named no one, ask everyone; overriding the operator's
 consultants on your own judgement — including after a `usage_limit` failure — substitutes a different mind
-for the one the user asked for.
+for the one the user asked for. The operator may have set a fallback model for that case; a caveat reading
+"answered by X: Y's quota was used up" means it answered, and your report names X as the one who answered.
 
 Consult your own CLI (`target: "antigravity"`) freely, alone or alongside the peers, and the same goes
 for a **different model of your own lineage** when the user wants one (an Opus lead asking Fable). Either is a

package/plugins/severally/skills/claude/severally/SKILL.md

@@ -62,7 +62,8 @@ actually reach. When the user names one, pass **the word the user used** as `tar
 
 **Do not pick a model yourself.** When the user named no one, ask everyone; overriding the operator's
 consultants on your own judgement — including after a `usage_limit` failure — substitutes a different mind
-for the one the user asked for.
+for the one the user asked for. The operator may have set a fallback model for that case; a caveat reading
+"answered by X: Y's quota was used up" means it answered, and your report names X as the one who answered.
 
 Consult your own CLI (`target: "claude-code"`) freely, alone or alongside the peers, and the same goes
 for a **different model of your own lineage** when the user wants one (an Opus lead asking Fable). Either is a

package/plugins/severally/skills/codex/severally/SKILL.md

@@ -61,7 +61,8 @@ actually reach. When the user names one, pass **the word the user used** as `tar
 
 **Do not pick a model yourself.** When the user named no one, ask everyone; overriding the operator's
 consultants on your own judgement — including after a `usage_limit` failure — substitutes a different mind
-for the one the user asked for.
+for the one the user asked for. The operator may have set a fallback model for that case; a caveat reading
+"answered by X: Y's quota was used up" means it answered, and your report names X as the one who answered.
 
 Consult your own CLI (`target: "codex"`) freely, alone or alongside the peers, and the same goes
 for a **different model of your own lineage** when the user wants one (an Opus lead asking Fable). Either is a

package/plugins/severally/skills/opencode/severally/SKILL.md

@@ -61,7 +61,8 @@ actually reach. When the user names one, pass **the word the user used** as `tar
 
 **Do not pick a model yourself.** When the user named no one, ask everyone; overriding the operator's
 consultants on your own judgement — including after a `usage_limit` failure — substitutes a different mind
-for the one the user asked for.
+for the one the user asked for. The operator may have set a fallback model for that case; a caveat reading
+"answered by X: Y's quota was used up" means it answered, and your report names X as the one who answered.
 
 Consult your own CLI (`target: "opencode"`) freely, alone or alongside the peers, and the same goes
 for a **different model of your own lineage** when the user wants one (an Opus lead asking Fable). Either is a

package/scripts/init-config.mjs

@@ -47,6 +47,7 @@ export function renderConfig() {
     '  //   effort           reasoning effort (default high); antigravity takes it from the model name instead',
     "  //   timeout_ms       this consultant's own budget in ms (default: server timeout_ms, max 7200000)",
     '  //   max_budget_usd   claude-code only: spending cap per consultation (default 10, 0.05-20)',
+    "  //   fallback_model   a model to run instead, once, when model's quota is used up (usage_limit)",
     '  //   vendor           the lineage, when the model name does not say (e.g. "nvidia")',
     '  //   enabled, note    as for clis, for this consultant alone',
     '  //',
@@ -117,7 +118,7 @@ function main() {
   }
   console.log('');
   console.log('clis.<cli>: enabled  note  bin  args  env');
-  console.log('consultants.<name>: cli  model  effort  timeout_ms  max_budget_usd  vendor  enabled  note');
+  console.log('consultants.<name>: cli  model  effort  timeout_ms  max_budget_usd  fallback_model  vendor  enabled  note');
   console.log('   more: CONFIG.md (what each key does), config.example.json (a working example).');
   console.log('   Restart the client after editing.');
   console.log('   comments and trailing commas are fine; name it config.jsonc if your editor prefers that');

package/src/consultants.mjs

@@ -15,6 +15,7 @@
  * @property {string | null} effort
  * @property {number | null} timeoutMs
  * @property {number | null} maxBudgetUsd
+ * @property {string | null} fallbackModel
  * @property {string | null} vendor
  * @property {boolean} enabled
  * @property {string | null} note
@@ -333,7 +334,7 @@ export function sameModel(callerModel, consultantModel) {
 
 const TOP_KEYS = ['server', 'clis', 'consultants', 'ranking'];
 const CLI_KEYS = ['enabled', 'note', 'bin', 'args', 'env', 'provider'];
-const CONSULTANT_KEYS = ['cli', 'model', 'effort', 'timeout_ms', 'max_budget_usd', 'vendor', 'enabled', 'note'];
+const CONSULTANT_KEYS = ['cli', 'model', 'effort', 'timeout_ms', 'max_budget_usd', 'fallback_model', 'vendor', 'enabled', 'note'];
 /**
  * @param {unknown} v
  * @returns {v is Record<string, unknown>}
@@ -460,7 +461,7 @@ export function buildConsultants(config) {
         if (!CONSULTANT_KEYS.includes(k)) problems.push(`${at}.${k}: unknown key (allowed: ${CONSULTANT_KEYS.join(', ')})`);
       if (!cliIds.includes(def.cli)) problems.push(`${at}.cli: must be one of ${CLI_IDS.join(', ')}`);
       if (typeof def.model !== 'string' || def.model.trim() === '') problems.push(`${at}.model: required`);
-      for (const k of ['effort', 'vendor', 'note']) {
+      for (const k of ['effort', 'fallback_model', 'vendor', 'note']) {
         if (def[k] !== undefined && typeof def[k] !== 'string') problems.push(`${at}.${k}: must be a string`);
       }
       if (def.enabled !== undefined && typeof def.enabled !== 'boolean') problems.push(`${at}.enabled: must be true or false`);
@@ -468,6 +469,10 @@ export function buildConsultants(config) {
         if (def[k] !== undefined && !(typeof def[k] === 'number' && Number.isFinite(def[k])))
           problems.push(`${at}.${k}: must be a number`);
       }
+      if (typeof def.fallback_model === 'string' && def.fallback_model.trim() === '')
+        problems.push(`${at}.fallback_model: must not be empty`);
+      else if (typeof def.fallback_model === 'string' && def.fallback_model.trim() === String(def.model).trim())
+        problems.push(`${at}.fallback_model: must differ from model; the fallback is for when the model's quota is used up`);
       if (def.max_budget_usd !== undefined && def.cli !== 'claude-code')
         problems.push(`${at}.max_budget_usd: only a claude-code consultant has a budget`);
       // agy takes the effort from the model name and rejects --effort.
@@ -500,6 +505,7 @@ export function buildConsultants(config) {
  * @property {unknown} [effort]
  * @property {unknown} [timeout_ms]
  * @property {unknown} [max_budget_usd]
+ * @property {unknown} [fallback_model]
  * @property {unknown} [vendor]
  * @property {unknown} [enabled]
  * @property {unknown} [note]
@@ -519,6 +525,7 @@ function make(name, def) {
     effort: typeof def.effort === 'string' && def.effort.trim() !== '' ? def.effort.trim() : null,
     timeoutMs: typeof def.timeout_ms === 'number' ? clamp(def.timeout_ms, 1_000, 7_200_000) : null,
     maxBudgetUsd: typeof def.max_budget_usd === 'number' ? clamp(def.max_budget_usd, 0.05, 20) : null,
+    fallbackModel: typeof def.fallback_model === 'string' && def.fallback_model.trim() !== '' ? def.fallback_model.trim() : null,
     vendor: vendorOf(model, def.vendor),
     enabled: def.enabled !== false,
     note: typeof def.note === 'string' && def.note.trim() !== '' ? def.note.trim() : null,

package/src/jobs.mjs

@@ -1123,26 +1123,35 @@ export class JobManager {
       // its latency is part of this job's duration and budget, not free time
       // before the meter starts.
       if (adapter.prepare) await adapter.prepare();
-      const invocation = adapter.buildInvocation({
-        workdir,
-        schemaPath,
-        guardrails: renderGuardrails(Boolean(req._snapshot), req.network),
-        sandbox,
-        consultant,
-        hasExposedPaths: Boolean(req._snapshot),
-        network: req.network,
-      });
-
-      const argsProblem = extraArgsProblem(cli, invocation.args, t.extraArgs);
-      if (argsProblem) return this.#fail(job, 'spawn_error', `config clis.${cli}.args: ${argsProblem}`);
-      invocation.args = withExtraArgs(invocation.args, t.extraArgs);
-      assertNoForbiddenFlags(cli, invocation.args);
-      // What this consultant was actually launched with, read off the
-      // invocation and the sandbox rather than restated from intent, so the
-      // record says what held for this run.
-      job.boundary =
-        adapter.boundary?.({ args: invocation.args, sandbox, network: req.network, hasExposedPaths: Boolean(req._snapshot) }) ??
-        null;
+      // Built again for the fallback model, so the operator's extra args and
+      // the forbidden-flag check apply to that run exactly as to the first.
+      // Returns the invocation, or the problem with the operator's args.
+      /** @param {import('./consultants.mjs').Consultant} c @returns {import('./run.mjs').Invocation | string} */
+      const launch = (c) => {
+        const invocation = adapter.buildInvocation({
+          workdir,
+          schemaPath,
+          guardrails: renderGuardrails(Boolean(req._snapshot), req.network),
+          sandbox,
+          consultant: c,
+          hasExposedPaths: Boolean(req._snapshot),
+          network: req.network,
+        });
+        const problem = extraArgsProblem(cli, invocation.args, t.extraArgs);
+        if (problem) return problem;
+        invocation.args = withExtraArgs(invocation.args, t.extraArgs);
+        assertNoForbiddenFlags(cli, invocation.args);
+        // What this consultant was actually launched with, read off the
+        // invocation and the sandbox rather than restated from intent, so the
+        // record says what held for this run.
+        job.boundary =
+          adapter.boundary?.({ args: invocation.args, sandbox, network: req.network, hasExposedPaths: Boolean(req._snapshot) }) ??
+          null;
+        return invocation;
+      };
+      const first = launch(consultant);
+      if (typeof first === 'string') return this.#fail(job, 'spawn_error', `config clis.${cli}.args: ${first}`);
+      let invocation = first;
 
       const attempt = (/** @type {number} */ timeout) => {
         const { handle, done } = runChild({
@@ -1176,6 +1185,7 @@ export class JobManager {
       let interpreted;
       let authTried = false;
       let autoRetries = 0;
+      let fallenBack = false;
       for (;;) {
         // Read off the consultant's own event stream, last attempt only; null
         // where the CLI does not stream its tool calls (Claude Code).
@@ -1247,6 +1257,25 @@ export class JobManager {
             continue;
           }
         }
+        // The operator's fallback for a model whose account quota is used up
+        // (a re-run on the same model would fail the same way). It is a
+        // different mind from the one asked, so the answer says so below.
+        if (interpreted.failureKind === 'usage_limit' && consultant.fallbackModel && !fallenBack && remaining >= MIN_RETRY_MS) {
+          fallenBack = true;
+          const fallback = consultant.fallbackModel;
+          const next = launch({ ...consultant, model: fallback });
+          if (typeof next === 'string') return this.#fail(job, 'spawn_error', `config clis.${cli}.args: ${next}`);
+          invocation = next;
+          job.retries.push({
+            failure_kind: interpreted.failureKind,
+            reason: `${job.model} quota used up; re-run on ${fallback}`,
+            at: new Date().toISOString(),
+            usage: adapter.usageRecord(interpreted.usageRaw),
+          });
+          job.model = fallback;
+          run = await attempt(remaining);
+          continue;
+        }
         const wait = POLICY.retryBackoffMs * 3 ** autoRetries;
         if (
           AUTO_RETRY_KINDS.includes(interpreted.failureKind) &&
@@ -1319,6 +1348,7 @@ export class JobManager {
           `the consultant's CLI reported an error mid-run and then finished its answer: ${redact(interpreted.recoveredError.message).slice(0, 300)}`,
         leftover &&
           'the answer parsed, but a process the consultant started kept its output open after it exited and was stopped (one that left its process group may still be running)',
+        fallenBack && `answered by ${job.model}: ${consultant.model}'s quota was used up`,
         req._snapshot && unreadCaveat(job.tool_calls, adapter.FILE_READ_TOOLS),
         refusedToolCaveat(job.tool_calls, adapter.ALLOWED_TOOLS),
       ].filter(Boolean);
@@ -2106,8 +2136,12 @@ function unreadCaveat(calls, readTools) {
   return `files were exposed, but no call to read one succeeded (${how})`;
 }
 
-// A refused call to a tool outside the allowlist is most often a tool the CLI
-// renamed in a new release; naming it lets one consultation show the cause.
+// A refused call to a tool outside the allowlist that looks like a file
+// reader is most often the CLI's read tool renamed in a new release; naming
+// it lets one consultation show the cause. The tools severally refuses on
+// purpose (agy tries run_command even for a trivial question) are left out:
+// reporting them would be noise, and they are not renamed readers.
+// ponytail: a name heuristic; a reader named otherwise goes unreported until the unread note shows it.
 /**
  * @param {Record<string, number> | null | undefined} calls
  * @param {readonly string[] | undefined} allowed
@@ -2118,9 +2152,9 @@ function refusedToolCaveat(calls, allowed) {
   const names = Object.keys(calls)
     .filter((k) => k.endsWith(':error'))
     .map((k) => k.slice(0, -':error'.length))
-    .filter((t) => !allowed.includes(t));
+    .filter((t) => !allowed.includes(t) && /read|view|open/i.test(t));
   if (!names.length) return null;
-  return `the consultant's CLI called ${names.join(', ')}, which severally does not allow, and was refused; a new CLI release may have renamed a tool`;
+  return `the consultant's CLI called ${names.join(', ')}, which severally does not allow, and was refused; a new CLI release may have renamed its file-reading tool`;
 }
 
 // A lead may legitimately name its own CLI as the consultant: a fresh child

package/src/server.mjs

@@ -11,7 +11,7 @@ import { exportChain, ExportError } from './export.mjs';
 
 /** @typedef {import('@modelcontextprotocol/server').CallToolResult} CallToolResult */
 
-export const SEVERALLY_VERSION = '2.4.2';
+export const SEVERALLY_VERSION = '2.5.0';
 
 // consult_get waits this long when no wait_ms is given: a call that forgot it
 // would otherwise return at once and be repeated. Up to 45 s, under the 60 s