opencode-harness-kit 0.6.0 → 0.6.2

package/CHANGELOG.md

@@ -12,8 +12,6 @@ is not happening.
 
 ---
 
-## [Unreleased]
-
 ## [0.6.0] — 2026-10-22
 
 Published. The tag is what cut it: `git tag v0.6.0 && git push origin

package/README.md

@@ -107,7 +107,7 @@ about your prompts.
 
 | lane | job | model | measured |
 |---|---|---|---|
-| `build` | write and verify the change | MiniMax M3.1 Flash | 6.9 s |
+| `build` | write and verify the change | Flint 40B Free | 6.9 s |
 | `plan` | turn a request into a plan | Beam | **6.1 s** |
 | `reviewer` | read a change in two passes | Nemotron 3 Ultra | 7.8 s (34 s cold) |
 | `locate` | find where something lives | Nemotron 3.5 Lightning | 6.9 s |
@@ -116,13 +116,14 @@ about your prompts.
 | `doc` | read a long document or a spec | GLM 5.3 Flash | 7.4 s |
 | `schema` | produce JSON or a JSON schema | DeepSeek V4.1 Flash | 10.7 s |
 | `sweep` | sweep a lot of code mechanically | Nemotron 3.5 Super VL | 8.4 s |
-| `longform` | read a megabyte, or write alternatives | MiniMax M3.1 Flash | 6.9 s |
+| `longform` | read a megabyte, or write alternatives | Flint 40B Free | 6.9 s |
 
 `longform` used to run on `longcat-2.5-preview-free`, and the provider
 withdrew it. A withdrawn model is not a lane, it is a hole in a lane: the
-route falls to the next rung of the same chain, which is MiniMax - the only
-model left on the tier with a million-token context. The `measured` column
-for that lane is therefore MiniMax's, not the withdrawn model's.
+route falls to the next rung of the same chain, where MiniMax then flint
+carried the lane - flint is the model that replaced MiniMax on the tier, and
+the only one left with the context the lane exists for. The `measured`
+column for that lane is therefore flint's, not either withdrawn model's.
 
 The `measured` column is the median of five one-line requests to that model
 on 2026-10-21 - the same prompt for every model, from the process start, so
@@ -243,7 +244,7 @@ it. One file covers ten jobs either way.
 |---|---|---|---|---|
 | `fleet-safe` (default) | 10 | 200,000 | `max` | zero-retention |
 | `fleet-all` | 10 | 160,000 | `max` | **may train** |
-| `minimax-m3.1-flash-free` | 5 | 1,000,000 | `max` | zero-retention |
+| `flint-40b-free` | 5 | 262,144 | `max` | **may train** |
 | `beam-free` | 5 | 1,048,576 | `max` | retained |
 | `deepseek-v4.1-flash-free` | 5 | 200,000 | `max` | zero-retention |
 | `glm-5.3-flash-free` | 5 | 200,000 | `max` | zero-retention |
@@ -252,6 +253,7 @@ it. One file covers ten jobs either way.
 | `nemotron-3.5-lightning-free` | 5 | 262,144 | — | **may train** |
 | `mimo-v2.6-flash-free` | 5 | 200,000 | — | **may train** |
 | `big-pickle` | 5 | 160,000 | — | **may train** |
+| `flint-40b-free` | 5 | 262,144 | `max` | **may train** |
 
 The five-lane profiles run everything on one model, which is the right shape
 when you want one route and want to know exactly which one.
@@ -282,6 +284,8 @@ withdrawn, and one of them was the **default** at the time:
 | `deepseek-v4-flash-free` | Withdrawn within a month. Replaced by `deepseek-v4.1-flash-free`. |
 | `muse-spark-1.3-contributor-free` | A Meta contributor route whose provider was permitted to train on prompts. The paid route is still in the catalogue. |
 | `mimo-v2.5-free` | Observed alive for one day. Replaced by `mimo-v2.6-flash-free`. |
+| `longcat-2.5-preview-free` | Was the `longform` lane's primary in both fleets, so when it was pulled every lane fell back a rung, and the one profile built only for that model was retired. |
+| `minimax-m3.1-flash-free` | Deprecated by the provider and replaced on the tier by `flint-40b-free`, which is the build and `longform` lane today. |
 
 The full list, with a reason for each, is in
 [`profiles/index.json`](profiles/index.json).

package/bin/oc-harness.mjs

@@ -12,7 +12,7 @@
 // The point of the kit: switching models is a data edit, not a code change.
 // Everything model-specific lives in profiles/<id>.json.
 
-import { chmodSync, copyFileSync, mkdirSync, readFileSync, writeFileSync, existsSync, rmSync, symlinkSync, readdirSync, readlinkSync, realpathSync, lstatSync, appendFileSync } from "node:fs"
+import { chmodSync, copyFileSync, mkdirSync, readFileSync, writeFileSync, existsSync, rmSync, symlinkSync, readdirSync, readlinkSync, realpathSync, lstatSync, appendFileSync, renameSync, unlinkSync } from "node:fs"
 import { homedir } from "node:os"
 import { basename, dirname, isAbsolute, join, relative, resolve, sep } from "node:path"
 import { execFileSync } from "node:child_process"
@@ -45,6 +45,7 @@ import {
 } from "../src/profile.js"
 import { generateConfig, globalInstructionsPath, utilityModel, withMemoryPlugin, canAddMemoryPlugin, kitVersion, MEMORY_PLUGIN } from "../src/config.js"
 import { ruleNames, BASH_RULES, WRITE_RULES } from "../src/rules.js"
+import { readHealth, unhealthyRoutes, writeHealth } from "../src/health.js"
 import { planEffort } from "../src/effort.js"
 import { findTelemetryFiles, readLaneUsage, averageMs } from "../src/lane-usage.js"
 import { researchModels, isWithdrawn, pruneWithdrawnChains } from "../src/research.js"
@@ -489,6 +490,46 @@ function projectRoot() {
  * one in the plural one, and a checker that only looks at `agents/` misses
  * half of them.
  */
+/**
+ * The health records whose model the catalogue lists again.
+ *
+ * Issue #9: a route recorded as withdrawn carried the quota's day-long
+ * window, so a pulled model was retried once a day, forever. The withdrawn
+ * window fixed the expiry half; this is the other half. A record whose
+ * model has COME BACK to the catalogue - back on the tier, or no longer
+ * flagged as withdrawn in it - describes a route that works again right
+ * now, and every hour the record still stands, the profile's config hook
+ * benches a live route and sends the lane to a rung down its chain.
+ *
+ * The clear is conservative in the way the health file itself is: a model
+ * that is IN the catalogue and not flagged withdrawn in it is trusted as
+ * returned. A model the fetch never saw stays recorded - a network error
+ * must not un-bench a route.
+ *
+ * @param {string} dir the project root the health file lives under
+ * @param {any} catalog the live model catalogue, or null
+ * @returns {string[]} the route keys that were cleared
+ */
+function clearRecoveredRoutes(dir, catalog) {
+  if (!catalog || typeof catalog !== "object") return []
+  const routes = unhealthyRoutes(dir)
+  const recovered = []
+  for (const key of Object.keys(routes)) {
+    const slash = key.indexOf("/")
+    // A key without its provider half was never written by the router.
+    if (slash < 1) continue
+    const provider = key.slice(0, slash)
+    const model = key.slice(slash + 1)
+    const record = catalog?.[provider]?.models?.[model]
+    if (record && !isWithdrawn(record)) recovered.push(key)
+  }
+  if (!recovered.length) return []
+  const health = readHealth(dir)
+  for (const key of recovered) delete health.unhealthy?.[key]
+  writeHealth(dir, health)
+  return recovered
+}
+
 function agentDirs() {
   const project = projectRoot()
   return [
@@ -876,6 +917,9 @@ async function cmdAuto() {
   try {
     // 1. The catalogue. A model that was withdrawn has to be caught before
     //    it is used, and the profile files are the only place that lives.
+    //    The same fetch feeds the health cleanup and the use step below;
+    //    offline, it stays null and both of those stand down with it.
+    let refreshed = { changed: [], gone: [], same: [], skipped: [], catalog: null }
     if (flag("--offline")) {
       steps.push({ step: "refresh", status: "skipped", detail: "--offline" })
     } else {
@@ -884,7 +928,7 @@ async function cmdAuto() {
       // profile that could not be read used to vanish from the report - and
       // the report said "ok": true while a profile was never checked against
       // the live catalogue.
-      const refreshed = await cmdRefreshAll()
+      refreshed = await cmdRefreshAll()
       for (const g of refreshed.gone) {
         problems.push(`profile ${g} is no longer in the live catalogue`)
       }
@@ -901,14 +945,35 @@ async function cmdAuto() {
       })
     }
 
+    // 1.5. The health file. A record whose model the catalogue lists again
+    //     describes a route that works, and the profile's config hook benches
+    //     a live route while it stands. This step runs with the same fetch the
+    //     refresh made, so it costs nothing extra. A `--offline` doctor skips
+    //     it with the refresh it belongs to: a catalogue that was not read
+    //     must not be trusted to un-bench anything.
+    const recovered = clearRecoveredRoutes(projectRoot(), refreshed.catalog)
+    if (recovered.length) {
+      for (const key of recovered) console.log(`  ${green("+")} ${key} is in the live catalogue again; the spent record was cleared`)
+      steps.push({
+        step: "health",
+        status: "ok",
+        detail: `${recovered.length} spent record(s) cleared`,
+      })
+    }
+
     // 2. The behavioural layer: the plugin link, the prompts, the skills, the
     //    global rules. Without them the config means nothing.
     await cmdInstall()
     steps.push({ step: "install", status: "ok" })
 
-    // 3. The config, recompiled from the profile.
-    await cmdUse(id)
-    steps.push({ step: "use", status: "ok", detail: id })
+    // 3. The config, recompiled from the profile - against the SAME catalogue
+    //    fetch, so the step does not pull models.dev a second time.
+    const used = await cmdUse(id, refreshed.catalog, { wantJson, report: { problems } })
+    if (used === "refused") {
+      steps.push({ step: "use", status: "failed", detail: id })
+    } else {
+      steps.push({ step: "use", status: "ok", detail: id })
+    }
 
     // 4. The check that says whether all of that worked.
     const verifyProblems = cmdVerify()
@@ -1452,7 +1517,28 @@ function memorySpec() {
   return MEMORY_PLUGIN
 }
 
-async function cmdUse(id) {
+/**
+ * Install a profile's config.
+ *
+ * @param {string} id
+ * @param {any} [catalog] a catalogue an earlier step already fetched
+ * @param {{ wantJson?: boolean, report?: { problems: string[] } | null }} [sink]
+ *   where a refusal goes instead of `fail()`. The one caller that runs with
+ *   an calmed stdout is `auto --json`: a script reads its report, the report
+ *   names the refusal as a problem, and the run still gets to print the
+ *   report. Channels an exit here and the report is never printed, which is
+ *   exactly the bug the JSON report exists to prevent.
+ */
+async function cmdUse(id, catalog = null, sink = {}) {
+  const wantJson = sink.wantJson === true
+  /** Refuse: issue the message the way the caller's stdout wants it said. */
+  const refuse = (msg) => {
+    if (sink.report) {
+      sink.report.problems.push(msg)
+      return false
+    }
+    fail(msg, wantJson)
+  }
   if (!id) fail("usage: oc-harness use <profile> [--print] [--yes]")
   let p = null
   try {
@@ -1507,6 +1593,76 @@ async function cmdUse(id) {
     console.error(msg + "  Continuing: there is no terminal to ask.\n")
   }
 
+  // The live catalogue, checked at install time because the profile that gets
+  // installed is a promise about what will serve tomorrow.
+  //
+  // Issue #11: `use` used to install a profile without reading the catalogue
+  // at all, so the first run of the harness was the first thing that knew a
+  // lane's model was gone. The profile's file was written, the header said
+  // every lane was healthy, and the harness said so by failing.
+  //
+  // - A dead BASE model ends the install: the profile has nothing to fall
+  //   back to, and writing a config that fails on the session's first token
+  //   is the bug, not the fix.
+  // - Dead rungs are pruned IN MEMORY before the config is compiled, so the
+  //   written config runs on what the catalogue still serves. The profile
+  //   file on disk keeps the rung until `oc-harness refresh` makes the fix
+  //   durable - an install must not rewrite the package's own data.
+  // - A lane with no rung left fails the install: a lane with no model is
+  //   not a lane, and the profile's author has to name the replacement.
+  //
+  // Fail-open: a catalogue that cannot be read is a warning, not a refusal.
+  // Refusing installs on a flaky network takes the tool away from the
+  // machine that needs to build offline, and `--offline` says that out loud.
+  if (!catalog && !flag("--offline")) {
+    try {
+      catalog = await fetchCatalog()
+    } catch (err) {
+      console.error(
+        yellow(
+          `  could not check the live catalogue (${err instanceof Error ? err.message : String(err)}). Install continues with the profile as it ships.`,
+        ),
+      )
+    }
+  }
+  if (catalog) {
+    const models = catalog?.[p.provider]?.models ?? {}
+    const record = models[p.model]
+    if (!record || isWithdrawn(record)) {
+      if (!refuse(
+        `profile ${id}'s base model ${p.provider}/${p.model} is no longer in the live catalogue.` +
+          ` Pick another profile: oc-harness list <no-args>`,
+      )) return "refused"
+    }
+    // The prune is silent until it has something to say. Its costs are the
+    // output below: the rungs it dropped and the one route that moved, and
+    // the way to make both durable.
+    const prune = pruneWithdrawnChains(p, catalog, p.provider)
+    if (prune.emptied.length) {
+      const lanes = prune.emptied.map((lane) => {
+        const rm = prune.lanes.find((l) => l.lane === lane)
+        return `${lane} (its models were: ${rm?.removed.join(", ")})`
+      })
+      if (!refuse(
+        `profile ${id}: every rung of the ${prune.emptied.join(", ")} lane is no longer in the live catalogue.` +
+          ` No model can serve that lane, so the profile needs a replacement before it installs.`,
+      )) return "refused"
+    }
+    if (prune.pruned.length || prune.utilityMoved) {
+      const drops = []
+      for (const lane of prune.lanes) drops.push(`lane ${lane.lane}: ${lane.from.join(" -> ")} becomes ${lane.to.join(" -> ")}`)
+      if (prune.fallback) drops.push(`fallback: ${prune.fallback.removed.join(", ")} dropped`)
+      if (prune.utilityMoved) drops.push(`the utility route moved to ${p.utility.model}`)
+      console.error(
+        yellow(
+          `  the live catalogue no longer serves ${prune.pruned.join(", ")}: the config this run compiled drops them.` +
+            ` Run oc-harness refresh to make the fix durable in the profile file itself.`,
+        ) +
+          `\n    ${drops.join("\n    ")}`,
+      )
+    }
+  }
+
   const base = readConfigJson()
   // `--memory` turns on the memory plugin. The kit does not implement memory
   // - a local vector store with embeddings, a web UI and auto-capture is a
@@ -1539,7 +1695,7 @@ async function cmdUse(id) {
 
   mkdirSync(dirname(CONFIG_FILE), { recursive: true })
   if (existsSync(CONFIG_FILE)) {
-// The config is the user's file once written. The backup is unconditional:
+    // The config is the user's file once written. The backup is unconditional:
     // `--yes` skips the question, it does not skip the safety net. Gating the
     // backup on `--yes` meant the one path that overwrites without asking was
     // also the one path with nothing to fall back to.
@@ -1560,7 +1716,15 @@ async function cmdUse(id) {
       stamp(`  existing config backed up to ${join(BACKUP_DIR, kept)}`),
     )
   }
-  writeFileSync(CONFIG_FILE, out)
+  // Write atomically to avoid truncating/corrupting the config on crash.
+  const tmp = `${CONFIG_FILE}.${process.pid}.${Date.now()}.tmp`
+  try {
+    writeFileSync(tmp, out)
+    renameSync(tmp, CONFIG_FILE)
+  } catch (err) {
+    try { unlinkSync(tmp) } catch {}
+    throw err
+  }
   console.log(`\n  ${green("✓")} wrote ${CONFIG_FILE}`)
   summary(p)
   console.log(`\n  ${stamp("verify with: opencode debug config")}\n`)
@@ -1825,7 +1989,10 @@ async function cmdRefreshAll() {
   // dropped, so a caller that silences the human output - `auto --json` -
   // never saw them: a withdrawn model or a profile that was never checked
   // against the catalogue went into a report that said everything was fine.
-  return { changed, gone, same, skipped }
+  // The catalogue itself comes back with the report: `doctor` runs the
+  // refresh as its opening step, and the steps that follow it check against
+  // the same fetch instead of pulling models.dev twice.
+  return { changed, gone, same, skipped, catalog }
 }
 
 /**

package/dist/health.d.ts

@@ -7,6 +7,13 @@ export type SpentEntry = {
      * the pattern that matched
      */
     reason: string;
+    /**
+     * why the route was written off.
+     * `withdrawn` entries expire on the catalogue's clock, not the quota's:
+     * a rate limit resets at 00:00 UTC, a pulled model does not. Entries
+     * written before the field existed behave as `spent`.
+     */
+    class?: ("spent" | "withdrawn");
 };
 export type Health = {
     unhealthy: Record<string, SpentEntry>;
@@ -21,6 +28,14 @@ export type Health = {
  * @returns {number}
  */
 export declare function spentWindowMs(): number;
+/**
+ * The window, in milliseconds, a route recorded as withdrawn is honoured for.
+ *
+ * Read at call time, like `spentWindowMs`.
+ *
+ * @returns {number}
+ */
+export declare function withdrawnWindowMs(): number;
 /**
  * Whether a recorded route is still spent.
  *

package/dist/index.js

@@ -1932,9 +1932,10 @@ function utilityModel(p) {
 }
 
 // src/health.js
-import { existsSync as existsSync5, mkdirSync as mkdirSync2, readFileSync as readFileSync3, writeFileSync } from "node:fs";
+import { existsSync as existsSync5, mkdirSync as mkdirSync2, readFileSync as readFileSync3, renameSync as renameSync2, unlinkSync, writeFileSync } from "node:fs";
 import { join as join7 } from "node:path";
 var DEFAULT_WINDOW_MS = 24 * 60 * 60 * 1e3;
+var DEFAULT_WITHDRAWN_WINDOW_MS = 30 * 24 * 60 * 60 * 1e3;
 function spentWindowMs() {
   const raw = process.env.HARNESS_SPENT_WINDOW_HOURS;
   if (!raw) return DEFAULT_WINDOW_MS;
@@ -1942,6 +1943,13 @@ function spentWindowMs() {
   if (!Number.isFinite(hours) || hours <= 0) return DEFAULT_WINDOW_MS;
   return hours * 60 * 60 * 1e3;
 }
+function withdrawnWindowMs() {
+  const raw = process.env.HARNESS_WITHDRAWN_WINDOW_HOURS;
+  if (!raw) return DEFAULT_WITHDRAWN_WINDOW_MS;
+  const hours = Number(raw);
+  if (!Number.isFinite(hours) || hours <= 0) return DEFAULT_WITHDRAWN_WINDOW_MS;
+  return hours * 60 * 60 * 1e3;
+}
 function isSpent(entry, now = Date.now()) {
   if (!entry || typeof entry !== "object" || Array.isArray(entry)) return false;
   const at = (
@@ -1951,7 +1959,11 @@ function isSpent(entry, now = Date.now()) {
   if (typeof at !== "string") return true;
   const when = Date.parse(at);
   if (Number.isNaN(when)) return true;
-  return now - when <= spentWindowMs();
+  const kind = (
+    /** @type {any} */
+    entry.class
+  );
+  return now - when <= (kind === "withdrawn" ? withdrawnWindowMs() : spentWindowMs());
 }
 function liveUnhealthy(health, now = Date.now()) {
   const out = {};
@@ -1983,9 +1995,21 @@ function readHealth(dir) {
 }
 function writeHealth(dir, health) {
   try {
-    if (!existsSync5(join7(dir, ".space"))) mkdirSync2(join7(dir, ".space"), { recursive: true });
+    const spaceDir = join7(dir, ".space");
+    if (!existsSync5(spaceDir)) mkdirSync2(spaceDir, { recursive: true });
     const pruned = { ...health ?? {}, unhealthy: liveUnhealthy(health) };
-    writeFileSync(healthPath(dir), JSON.stringify(pruned, null, 2));
+    const file = healthPath(dir);
+    const tmp = `${file}.${process.pid}.${Date.now()}.tmp`;
+    try {
+      writeFileSync(tmp, JSON.stringify(pruned, null, 2));
+      renameSync2(tmp, file);
+    } catch (err) {
+      try {
+        unlinkSync(tmp);
+      } catch {
+      }
+      throw err;
+    }
   } catch {
   }
 }
@@ -2006,18 +2030,18 @@ var SPENT = [
   /rate[_ -]?limit/i,
   /\b429\b/,
   /quota (?:exceeded|exhausted)/i,
-  /overloaded_error/i,
-  // A withdrawn model, named in the error. The words have to sit next to
-  // each other - "model minimax-m3.1-flash-free not found", "model_not_found",
-  // "ModelNotFound" - because the text this runs against is the provider's
-  // own error and that can be a whole response body. The old pattern ran
-  // over the entire event, where a body that mentioned a model in one place
-  // and "not found" in another read as a withdrawn model: a live route
-  // written off by a coincidence.
+  /overloaded_error/i
+];
+var WITHDRAWN = [
   /\bmodel[\W_]*not[\W_]*found\b/i,
   /\bmodel\b[\w.:/-]*(?:\s+[\w.:/-]+){0,2}\s+not found\b/i,
   /no available provider/i
 ];
+var TRANSIENT = [
+  /\b502\b|\b503\b|\b529\b/,
+  /bad gateway/i,
+  /\bupstream\b/
+];
 function errorText(err) {
   if (err == null) return "";
   if (typeof err === "string") return err;
@@ -2047,7 +2071,12 @@ function errorText(err) {
 function classifyError(err) {
   const text = errorText(err);
   if (!text) return null;
-  for (const p of SPENT) if (p.test(text)) return p.source;
+  for (const p of SPENT)
+    if (p.test(text)) return { kind: "spent", reason: p.source };
+  for (const p of WITHDRAWN)
+    if (p.test(text)) return { kind: "withdrawn", reason: p.source };
+  for (const p of TRANSIENT)
+    if (p.test(text)) return { kind: "transient", reason: p.source };
   return null;
 }
 function errorOf(e) {
@@ -2163,14 +2192,21 @@ var RouterPlugin = async ({ worktree, directory }) => {
       }
       const err = errorOf(e);
       if (!err) return;
-      const reason = classifyError(err);
-      if (!reason) return;
+      const verdict = classifyError(err);
+      if (!verdict) return;
       const sessionID = sessionIdOf(e);
       const key = seen ?? (sessionID ? sessionRoutes.get(sessionID) : void 0) ?? configuredRoute ?? lastRoute;
       if (!key) return;
+      if (verdict.kind === "transient") return;
       const health = readHealth(project);
       if (unhealthyRoutes(project)[key]) return;
-      health.unhealthy[key] = { at: (/* @__PURE__ */ new Date()).toISOString(), reason };
+      health.unhealthy[key] = {
+        at: (/* @__PURE__ */ new Date()).toISOString(),
+        reason: verdict.reason,
+        // The class decides the window the entry expires under. A model the
+        // provider pulled is not back at 00:00 UTC, whichever the quota says.
+        class: verdict.kind
+      };
       writeHealth(project, health);
       failedSessions++;
     },

package/dist/router.d.ts

@@ -1,3 +1,19 @@
 import type { Plugin } from "./plugin.js";
+/**
+ * Which of the three classes a provider error belongs to.
+ *
+ * The classes decide what happens to the health file: `spent` is recorded
+ * under the quota's window, `withdrawn` under the catalogue's, and
+ * `transient` is not recorded at all. Exported so a test can feed an error
+ * straight through the classifier instead of inventing an event to carry
+ * it.
+ *
+ * @param {any} err
+ * @returns {{ kind: "spent" | "withdrawn" | "transient", reason: string } | null}
+ */
+export declare function classifyError(err: any): {
+    kind: "spent" | "withdrawn" | "transient";
+    reason: string;
+} | null;
 export declare const RouterPlugin: Plugin;
 export default RouterPlugin;

package/eval/ab.mjs

@@ -750,7 +750,7 @@ console.log("")
 console.log(`  ${"metric (median)".padEnd(16)} ${"kit on".padStart(10)} ${"kit off".padStart(10)}`)
 console.log(`  ${"-".repeat(16)} ${"-".repeat(10)} ${"-".repeat(10)}`)
 for (const [label, a, b] of rows) {
-  console.log(`  ${label.padEnd(16)} ${String(a).padStart(10)} ${String(b).padStart(10)}`)
+  console.log(`  ${String(label).padEnd(16)} ${String(a).padStart(10)} ${String(b).padStart(10)}`)
 }
 
 if (jobs.length > 1) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-harness-kit",
-  "version": "0.6.0",
+  "version": "0.6.2",
   "description": "Harness kit for OpenCode: capability profiles, a fleet of job lanes with one free model each, role-based reasoning effort, context pruning, and measured baselines.",
   "license": "MIT",
   "homepage": "https://github.com/pham-canh/opencode-harness-kit#readme",

package/profiles/fleet-all.json

@@ -3,15 +3,15 @@
   "id": "fleet-all",
   "title": "Ten jobs, every free model, one of them may train",
   "provider": "opencode",
-  "model": "minimax-m3.1-flash-free",
-  "family": null,
+  "model": "flint-40b-free",
+  "family": "flint",
   "verified": "2026-10-22",
   "retention": "may-train",
   "maxRetention": "may-train",
   "notes": "Every lane gets the model built for it: beam plans, Nemotron 3 Ultra reviews, Lightning sweeps, MiMo sees, GLM audits, Super VL reasons with the effort turned off. Five of the ten providers state that collected data may be used to improve the model, or log the session, so the profile as a whole is may-train and the generated config says so in its header. Install it for public or throwaway repositories.",
   "limits": {
-    "context": 1000000,
-    "output": 512000
+    "context": 262144,
+    "output": 65536
   },
   "supports": {
     "reasoning": true,
@@ -49,9 +49,9 @@
   },
   "roleModels": {
     "build": {
-      "model": "minimax-m3.1-flash-free",
+      "model": "flint-40b-free",
       "chain": [
-        "minimax-m3.1-flash-free",
+        "flint-40b-free",
         "deepseek-v4.1-flash-free"
       ]
     },
@@ -90,7 +90,7 @@
       "model": "mimo-v2.6-flash-free",
       "chain": [
         "mimo-v2.6-flash-free",
-        "minimax-m3.1-flash-free"
+        "deepseek-v4.1-flash-free"
       ]
     },
     "doc": {
@@ -116,9 +116,9 @@
       ]
     },
     "longform": {
-      "model": "minimax-m3.1-flash-free",
+      "model": "flint-40b-free",
       "chain": [
-        "minimax-m3.1-flash-free",
+        "flint-40b-free",
         "beam-free"
       ]
     }
@@ -130,7 +130,7 @@
     "keepUserMessages": true
   },
   "fallback": [
-    "minimax-m3.1-flash-free",
+    "flint-40b-free",
     "deepseek-v4.1-flash-free"
   ]
 }

package/profiles/fleet-safe.json

@@ -3,15 +3,15 @@
   "id": "fleet-safe",
   "title": "Ten jobs, zero-retention models only",
   "provider": "opencode",
-  "model": "minimax-m3.1-flash-free",
-  "family": null,
+  "model": "deepseek-v4.1-flash-free",
+  "family": "deepseek",
   "verified": "2026-10-22",
   "retention": "zero-retention",
   "maxRetention": "zero-retention",
-  "notes": "The default. Every lane resolves to a model whose provider states a zero-retention policy: no training, nothing kept. The cost is that six of the ten jobs run on a second-choice model - the planner loses beam (30-day retention) and the reviewer loses Nemotron 3 Ultra - so this is the floor, not the ceiling. The chains are ordered within the ceiling rather than copied from `fleet-all`: `locate` takes deepseek because it publishes a `low` effort level that longcat does not, and `vision` takes minimax because it reads an image in 7 seconds. `fleet-all` gives every lane the model built for it.",
+  "notes": "The default. Every lane resolves to a model whose provider states a zero-retention policy: no training, nothing kept. The cost is that six of the ten jobs run on a second-choice model - the planner loses beam (30-day retention) and the reviewer loses Nemotron 3 Ultra - so this is the floor, not the ceiling. The chains are ordered within the ceiling rather than copied from `fleet-all`: `locate` takes deepseek because it publishes a `low` effort level that longcat does not, and `vision` takes deepseek because it is the zero-retention route the provider kept for image reads. `fleet-all` gives every lane the model built for it.",
   "limits": {
-    "context": 1000000,
-    "output": 512000
+    "context": 200000,
+    "output": 32000
   },
   "supports": {
     "reasoning": true,
@@ -49,10 +49,10 @@
   },
   "roleModels": {
     "build": {
-      "model": "minimax-m3.1-flash-free",
+      "model": "deepseek-v4.1-flash-free",
       "chain": [
-        "minimax-m3.1-flash-free",
-        "deepseek-v4.1-flash-free"
+        "deepseek-v4.1-flash-free",
+        "glm-5.3-flash-free"
       ]
     },
     "plan": {
@@ -73,7 +73,7 @@
       "model": "deepseek-v4.1-flash-free",
       "chain": [
         "deepseek-v4.1-flash-free",
-        "minimax-m3.1-flash-free"
+        "glm-5.3-flash-free"
       ]
     },
     "verify": {
@@ -84,10 +84,10 @@
       ]
     },
     "vision": {
-      "model": "minimax-m3.1-flash-free",
+      "model": "deepseek-v4.1-flash-free",
       "chain": [
-        "minimax-m3.1-flash-free",
-        "deepseek-v4.1-flash-free"
+        "deepseek-v4.1-flash-free",
+        "glm-5.3-flash-free"
       ]
     },
     "doc": {
@@ -108,14 +108,14 @@
       "model": "deepseek-v4.1-flash-free",
       "chain": [
         "deepseek-v4.1-flash-free",
-        "minimax-m3.1-flash-free"
+        "glm-5.3-flash-free"
       ]
     },
     "longform": {
-      "model": "minimax-m3.1-flash-free",
+      "model": "deepseek-v4.1-flash-free",
       "chain": [
-        "minimax-m3.1-flash-free",
-        "deepseek-v4.1-flash-free"
+        "deepseek-v4.1-flash-free",
+        "glm-5.3-flash-free"
       ]
     }
   },

package/profiles/flint-40b-free.json

@@ -0,0 +1,50 @@
+{
+  "$comment": "Metadata read from models.dev on 2026-10-23 and confirmed by a live call the same day. No latency figure is claimed: one call that answered is not a measurement. Regenerate the capability fields with `oc-harness refresh flint-40b-free`.",
+  "id": "flint-40b-free",
+  "title": "Flint 40B Free",
+  "provider": "opencode",
+  "model": "flint-40b-free",
+  "family": "flint",
+  "verified": "2026-10-23",
+  "retention": "may-train",
+  "notes": "An open-weight mixture of experts, 40B total with about 2.5B active, built for coding agents. Small enough to serve locally, and the weights are published, so the same model can be pointed at a local runtime instead of the free tier. 262k context and a 64k output cap, the full effort ladder, tool calls and structured output; it does not read images. Zen keeps prompts for 30 days and may train on them during the free period.",
+  "limits": {
+    "context": 262144,
+    "output": 65536
+  },
+  "supports": {
+    "reasoning": true,
+    "temperature": true,
+    "toolCall": true,
+    "structuredOutput": true,
+    "attachment": false
+  },
+  "effort": {
+    "levels": [
+      "low",
+      "medium",
+      "high",
+      "xhigh",
+      "max"
+    ],
+    "default": "max",
+    "reserveShare": 0.12
+  },
+  "roles": {
+    "plan": "max",
+    "build": "max",
+    "reviewer": "max",
+    "locate": "low",
+    "verify": "low"
+  },
+  "prune": {
+    "enabled": true,
+    "keepRecentToolCalls": 6,
+    "keepRecentTurns": 3,
+    "keepUserMessages": true
+  },
+  "fallback": [
+    "deepseek-v4.1-flash-free",
+    "nemotron-3.5-lightning-free"
+  ]
+}

package/profiles/index.json

@@ -2,6 +2,17 @@
   "$comment": "Profiles verified against live provider metadata. `retention` must come from retention.yml; a model missing there cannot enter this list. The default is a fleet, because a fleet is the point: ten jobs, ten models, one session. `context` is the input cap the config can actually use: the narrowest lane model's on a fleet, the model's own on a single-model profile. `oc-harness list` recomputes it from the profile rather than trusting this copy.",
   "default": "fleet-safe",
   "profiles": [
+    {
+      "id": "beam-free",
+      "file": "profiles/beam-free.json",
+      "title": "Beam Free",
+      "retention": "retained",
+      "free": true,
+      "topEffort": "max",
+      "context": 1048576,
+      "default": false,
+      "note": "Text only, freshest knowledge on the tier, fastest measured route. Keeps prompts for 30 days, so it is not zero-retention."
+    },
     {
       "id": "fleet-safe",
       "file": "profiles/fleet-safe.json",
@@ -24,28 +35,6 @@
       "default": false,
       "note": "Every lane gets the model built for it. Five of the ten providers may use the data to improve the model, so the whole profile is may-train."
     },
-    {
-      "id": "minimax-m3.1-flash-free",
-      "file": "profiles/minimax-m3.1-flash-free.json",
-      "title": "MiniMax M3.1 Flash Free",
-      "retention": "zero-retention",
-      "free": true,
-      "topEffort": "max",
-      "context": 1000000,
-      "default": false,
-      "note": "512k output tokens, the widest of any free route, and the fastest measured image reader at 7.7s."
-    },
-    {
-      "id": "beam-free",
-      "file": "profiles/beam-free.json",
-      "title": "Beam Free",
-      "retention": "retained",
-      "free": true,
-      "topEffort": "max",
-      "context": 1048576,
-      "default": false,
-      "note": "Text only, freshest knowledge on the tier, fastest measured route. Keeps prompts for 30 days, so it is not zero-retention."
-    },
     {
       "id": "deepseek-v4.1-flash-free",
       "file": "profiles/deepseek-v4.1-flash-free.json",
@@ -112,6 +101,17 @@
       "default": false,
       "note": "Reads images correctly, at 19.2s - the slowest lane on the tier. Publishes audio input too, but OpenCode will not read a binary file, so that path is unreachable today."
     },
+    {
+      "id": "flint-40b-free",
+      "file": "profiles/flint-40b-free.json",
+      "title": "Flint 40B Free",
+      "retention": "may-train",
+      "free": true,
+      "topEffort": "max",
+      "context": 262144,
+      "default": false,
+      "note": "Open-weight MoE with about 2.5B active parameters for local coding agents. Knowledge cutoff 2026-06, no image input. May train on prompts during its free period."
+    },
     {
       "id": "big-pickle",
       "file": "profiles/big-pickle.json",
@@ -122,6 +122,17 @@
       "context": 160000,
       "default": false,
       "note": "Stealth model, no public identity, no effort levels, no image input. The provider may train on it during its free period."
+    },
+    {
+      "id": "flint-40b-free",
+      "file": "profiles/flint-40b-free.json",
+      "title": "Flint 40B Free",
+      "retention": "may-train",
+      "free": true,
+      "topEffort": "max",
+      "context": 262144,
+      "default": false,
+      "note": "Open-weight MoE, 40B total with about 2.5B active, built for coding agents and small enough to serve locally. Full effort ladder, 262k context, 64k output, no image input."
     }
   ],
   "excluded": [],
@@ -149,6 +160,11 @@
     {
       "id": "longcat-2.5-preview-free",
       "reason": "Withdrawn from the provider. A preview endpoint, so it could be pulled without a rename - and it was: models.dev still carries the record and marks it deprecated. It was the longform lane's primary in both fleets, so every lane fell back to its second rung, and the profile named only this model had nothing left to fall back to and was retired."
+    },
+    {
+      "id": "minimax-m3.1-flash-free",
+      "reason": "Deprecated by provider. Replaced on the tier by flint-40b-free for the lanes that ran it."
     }
-  ]
+  ],
+  "active": []
 }

package/profiles/mimo-v2.6-flash-free.json

@@ -5,7 +5,7 @@
   "provider": "opencode",
   "model": "mimo-v2.6-flash-free",
   "family": "mimo",
-  "verified": "2026-10-22",
+  "verified": "2026-10-23",
   "retention": "may-train",
   "notes": "Xiaomi. The only free model that publishes audio input, and it read an image correctly in a live call, at 19.2s average - the slowest lane on the tier, and slower than minimax on the same picture. OpenCode 1.19 refuses to read a binary file at all, so the audio path is not reachable through the harness today; the vision path is. No effort levels, so no variant is written.",
   "limits": {
@@ -37,7 +37,5 @@
     "keepRecentTurns": 3,
     "keepUserMessages": true
   },
-  "fallback": [
-    "minimax-m3.1-flash-free"
-  ]
+  "fallback": []
 }

package/retention.yml

@@ -27,14 +27,6 @@ models:
   # --- Zero-retention: the provider states a zero-retention policy and does
   # not train on your data. ---
 
-  opencode/minimax-m3.1-flash-free:
-    retention: zero-retention
-    note: >-
-      "Its provider follows a zero-retention policy and does not use your data
-      for model training." Free for a limited time.
-    verified: 2026-10-19
-    source: https://opencode.ai/docs/zen
-
   opencode/glm-5.3-flash-free:
     retention: zero-retention
     note: >-
@@ -74,6 +66,15 @@ models:
     verified: 2026-10-19
     source: https://opencode.ai/docs/zen
 
+  opencode/flint-40b-free:
+    retention: may-train
+    note: >-
+      "During its free period, collected data may be used to improve the
+      model. Data is retained for 30 days." The weights are open, so the same
+      model can be served locally instead of through the free tier.
+    verified: 2026-10-23
+    source: https://opencode.ai/docs/zen
+
   opencode/mimo-v2.6-flash-free:
     retention: may-train
     note: >-
@@ -109,6 +110,14 @@ models:
     verified: 2026-10-19
     source: https://opencode.ai/docs/zen
 
+  opencode/flint-40b-free:
+    retention: may-train
+    note: >-
+      Free trial endpoint on opencode zen; treated as may-train until terms
+      state otherwise.
+    verified: 2026-10-22
+    source: https://opencode.ai/docs/zen
+
   # --- Paid routes, kept because the catalog still serves them and a profile
   # may be pointed at one. ---
 

package/src/health.js

@@ -12,13 +12,17 @@
 // Plain JS with JSDoc types, like the rest of src/: the CLI and the tests run
 // under plain node, and the plugin loads .ts.
 
-import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs"
+import { existsSync, mkdirSync, readFileSync, renameSync, unlinkSync, writeFileSync } from "node:fs"
 import { join } from "node:path"
 
 /**
  * @typedef {object} SpentEntry
  * @property {string} at when the route was recorded as spent
  * @property {string} reason the pattern that matched
+ * @property {("spent"|"withdrawn")} [class] why the route was written off.
+ *   `withdrawn` entries expire on the catalogue's clock, not the quota's:
+ *   a rate limit resets at 00:00 UTC, a pulled model does not. Entries
+ *   written before the field existed behave as `spent`.
  */
 
 /**
@@ -42,6 +46,20 @@ import { join } from "node:path"
  */
 const DEFAULT_WINDOW_MS = 24 * 60 * 60 * 1000
 
+/**
+ * How long a route recorded as WITHDRAWN is honoured for, by default.
+ *
+ * A rate limit resets at 00:00 UTC. A withdrawal does not: the provider
+ * pulled the model, and no number of days' sleep brings it back. Treating
+ * both with the one window meant a model that had left the catalogue was
+ * retried once a day, forever, and every lane that named it first paid a
+ * failed request before falling through.
+ *
+ * Thirty days is the scale the toolkit's catalogue turns on - free tiers
+ * rotate monthly - and HARNESS_WITHDRAWN_WINDOW_HOURS overrides it.
+ */
+const DEFAULT_WITHDRAWN_WINDOW_MS = 30 * 24 * 60 * 60 * 1000
+
 /**
  * The window, in milliseconds, a recorded spent route is honoured for.
  *
@@ -61,6 +79,21 @@ export function spentWindowMs() {
   return hours * 60 * 60 * 1000
 }
 
+/**
+ * The window, in milliseconds, a route recorded as withdrawn is honoured for.
+ *
+ * Read at call time, like `spentWindowMs`.
+ *
+ * @returns {number}
+ */
+export function withdrawnWindowMs() {
+  const raw = process.env.HARNESS_WITHDRAWN_WINDOW_HOURS
+  if (!raw) return DEFAULT_WITHDRAWN_WINDOW_MS
+  const hours = Number(raw)
+  if (!Number.isFinite(hours) || hours <= 0) return DEFAULT_WITHDRAWN_WINDOW_MS
+  return hours * 60 * 60 * 1000
+}
+
 /**
  * Whether a recorded route is still spent.
  *
@@ -79,7 +112,12 @@ export function isSpent(entry, now = Date.now()) {
   if (typeof at !== "string") return true
   const when = Date.parse(at)
   if (Number.isNaN(when)) return true
-  return now - when <= spentWindowMs()
+  // The window the entry was written under. An entry that carries no class
+  // was written by an older version of the kit, or by a hand, and gets the
+  // quota's window: the quota resets, which is the assumption every entry
+  // used to carry.
+  const kind = /** @type {any} */ (entry).class
+  return now - when <= (kind === "withdrawn" ? withdrawnWindowMs() : spentWindowMs())
 }
 
 /**
@@ -158,9 +196,18 @@ export function readHealth(dir) {
  */
 export function writeHealth(dir, health) {
   try {
-    if (!existsSync(join(dir, ".space"))) mkdirSync(join(dir, ".space"), { recursive: true })
+    const spaceDir = join(dir, ".space")
+    if (!existsSync(spaceDir)) mkdirSync(spaceDir, { recursive: true })
     const pruned = { ...(health ?? {}), unhealthy: liveUnhealthy(health) }
-    writeFileSync(healthPath(dir), JSON.stringify(pruned, null, 2))
+    const file = healthPath(dir)
+    const tmp = `${file}.${process.pid}.${Date.now()}.tmp`
+    try {
+      writeFileSync(tmp, JSON.stringify(pruned, null, 2))
+      renameSync(tmp, file)
+    } catch (err) {
+      try { unlinkSync(tmp) } catch {}
+      throw err
+    }
   } catch {
     // Best effort. A health file that cannot be written means no failover,
     // and failing to write it must not break the session.

package/src/router.ts

@@ -36,25 +36,53 @@ import {
   resolveRoles,
 } from "./profile.js"
 
-/** Patterns that mean "this route is spent", not "this request was bad". */
+/**
+ * Patterns that mean "this route is spent", not "this request was bad": the
+ * provider's quota for the free tier ran out. The quota resets, which is why
+ * these expire on the day's window.
+ */
 const SPENT = [
   /FreeUsageLimitError/i,
   /rate[_ -]?limit/i,
   /\b429\b/,
   /quota (?:exceeded|exhausted)/i,
   /overloaded_error/i,
-  // A withdrawn model, named in the error. The words have to sit next to
-  // each other - "model minimax-m3.1-flash-free not found", "model_not_found",
-  // "ModelNotFound" - because the text this runs against is the provider's
-  // own error and that can be a whole response body. The old pattern ran
-  // over the entire event, where a body that mentioned a model in one place
-  // and "not found" in another read as a withdrawn model: a live route
-  // written off by a coincidence.
+]
+
+/**
+ * Patterns that mean "this model is gone": the provider pulled it.
+ *
+ * The words have to sit next to each other - "model minimax-m3.1-flash-free
+ * not found", "model_not_found", "ModelNotFound" - because the text this
+ * runs against is the provider's own error and that can be a whole response
+ * body. The old pattern ran over the entire event, where a body that
+ * mentioned a model in one place and "not found" in another read as a
+ * withdrawn model: a live route written off by a coincidence.
+ *
+ * These get their own class and their own window: no number of days' sleep
+ * returns a pulled model, so benching one for 24h retried a death once a
+ * day, forever.
+ */
+const WITHDRAWN = [
   /\bmodel[\W_]*not[\W_]*found\b/i,
   /\bmodel\b[\w.:/-]*(?:\s+[\w.:/-]+){0,2}\s+not found\b/i,
   /no available provider/i,
 ]
 
+/**
+ * Patterns that mean "this request could not be answered" - the host was at
+ * capacity, a gateway hiccuped - and nothing about the ROUTE is worse.
+ *
+ * Benching a route for one of these benches a model that answered minutes
+ * later. They are not spent, not withdrawn, and they record nothing: the
+ * next request tries the same route again.
+ */
+const TRANSIENT = [
+  /\b502\b|\b503\b|\b529\b/,
+  /bad gateway/i,
+  /\bupstream\b/,
+]
+
 /**
  * The text of a provider error, and nothing else.
  *
@@ -99,15 +127,26 @@ function errorText(err: any): string {
 }
 
 /**
- * Whether a provider error means the route is spent.
+ * Which of the three classes a provider error belongs to.
+ *
+ * The classes decide what happens to the health file: `spent` is recorded
+ * under the quota's window, `withdrawn` under the catalogue's, and
+ * `transient` is not recorded at all. Exported so a test can feed an error
+ * straight through the classifier instead of inventing an event to carry
+ * it.
  *
  * @param {any} err
- * @returns {string | null} the pattern that matched
+ * @returns {{ kind: "spent" | "withdrawn" | "transient", reason: string } | null}
  */
-function classifyError(err: any): string | null {
+export function classifyError(err: any): { kind: "spent" | "withdrawn" | "transient"; reason: string } | null {
   const text = errorText(err)
   if (!text) return null
-  for (const p of SPENT) if (p.test(text)) return p.source
+  for (const p of SPENT)
+    if (p.test(text)) return { kind: "spent", reason: p.source }
+  for (const p of WITHDRAWN)
+    if (p.test(text)) return { kind: "withdrawn", reason: p.source }
+  for (const p of TRANSIENT)
+    if (p.test(text)) return { kind: "transient", reason: p.source }
   return null
 }
 
@@ -352,8 +391,8 @@ export const RouterPlugin: Plugin = async ({ worktree, directory }) => {
       // anything about the route.
       const err = errorOf(e)
       if (!err) return
-      const reason = classifyError(err)
-      if (!reason) return
+      const verdict = classifyError(err)
+      if (!verdict) return
 
       // The route the failure belongs to: the event's own, else the route
       // this session was last seen on, else the route it was configured
@@ -367,9 +406,27 @@ export const RouterPlugin: Plugin = async ({ worktree, directory }) => {
         lastRoute
       if (!key) return
 
+      // A transient failure is the host failing to answer one request, not
+      // the route becoming unusable. Nothing is written, so the next request
+      // tries the same rung again - benching the route under the quota's
+      // window is how a capacity blip cost a model a day.
+      //
+      // ("Advances to the next rung inside the same request" is not on the
+      // table: `chat.params` carries only temperature and sampling, the model
+      // on `chat.message` is read-only, and `config` has already run. What a
+      // transient verdict buys - and what this buys - is a route that is
+      // still live when the next request arrives.)
+      if (verdict.kind === "transient") return
+
       const health = readHealth(project)
       if (unhealthyRoutes(project)[key]) return // already recorded, do not spam
-      health.unhealthy[key] = { at: new Date().toISOString(), reason }
+      health.unhealthy[key] = {
+        at: new Date().toISOString(),
+        reason: verdict.reason,
+        // The class decides the window the entry expires under. A model the
+        // provider pulled is not back at 00:00 UTC, whichever the quota says.
+        class: verdict.kind,
+      }
       writeHealth(project, health)
       failedSessions++
     },

package/profiles/minimax-m3.1-flash-free.json

@@ -1,50 +0,0 @@
-{
-  "$comment": "Metadata verified against models.dev on 2026-10-19 and confirmed by a live call the same day. Regenerate the capability fields with `oc-harness refresh minimax-m3.1-flash-free`.",
-  "id": "minimax-m3.1-flash-free",
-  "title": "MiniMax M3.1 Flash Free",
-  "provider": "opencode",
-  "model": "minimax-m3.1-flash-free",
-  "family": "minimax",
-  "verified": "2026-10-22",
-  "retention": "zero-retention",
-  "notes": "Flagship coding and agent model on the free tier: 512k output tokens, the widest of any free route, and the fastest measured image reader at 7.7s. Tool calls are stable. It does not support temperature, and models.dev publishes no structured_output flag for it, so JSON output is requested in the prompt rather than negotiated with the API. Measured at 3.1s for a one-word reply.",
-  "limits": {
-    "context": 1000000,
-    "output": 512000
-  },
-  "supports": {
-    "reasoning": true,
-    "temperature": true,
-    "toolCall": true,
-    "structuredOutput": false,
-    "attachment": true
-  },
-  "effort": {
-    "levels": [
-      "low",
-      "medium",
-      "high",
-      "xhigh",
-      "max"
-    ],
-    "default": "max",
-    "reserveShare": 0.12
-  },
-  "roles": {
-    "plan": "max",
-    "build": "max",
-    "reviewer": "max",
-    "locate": "low",
-    "verify": "low"
-  },
-  "prune": {
-    "enabled": true,
-    "keepRecentToolCalls": 6,
-    "keepRecentTurns": 3,
-    "keepUserMessages": true
-  },
-  "fallback": [
-    "deepseek-v4.1-flash-free",
-    "glm-5.3-flash-free"
-  ]
-}