pi-cache-optimizer 2.6.7 → 2.6.10

package/index.ts

@@ -1,5 +1,6 @@
 import { createHash } from "node:crypto";
 import { copyFile, mkdir, readFile, rename, unlink, writeFile } from "node:fs/promises";
+import { readFileSync } from "node:fs";
 import { homedir } from "node:os";
 import { dirname, join } from "node:path";
 import type { BuildSystemPromptOptions, ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
@@ -958,12 +959,12 @@ function getNonNegativeNumber(record: UnknownRecord, key: string): number | unde
  */
 function getCompat(model: PiModel | undefined): CacheCompat {
   if (!model) return {} as CacheCompat;
-  
+
   // Pi merges provider.compat with model.compat (model wins on conflicts)
   // We approximate this by reading from ctx.model which should already have merged compat
   // However, for safety, we check both levels if available
   const modelCompat = (model.compat ?? {}) as CacheCompat;
-  
+
   // Note: ctx.model from Pi should already contain merged compat,
   // but we document the two-level structure for clarity
   return modelCompat;
@@ -1941,6 +1942,13 @@ function describeMissingOpenAICompatibleProxyCompat(model: PiModel): string[] {
     missing.push("sendSessionAffinityHeaders");
   }
 
+  // NOTE: supportsLongCacheRetention is intentionally NOT checked here.
+  // Per spec, it is optional/risky advisory text only and must NOT trigger
+  // the ⚠️ compat marker. The before_provider_request hook proactively
+  // strips prompt_cache_retention for models without explicit opt-in,
+  // so 400 errors are prevented regardless of this compat flag.
+  // Doctor/compat may mention it as optional guidance separately.
+
   return missing;
 }
 
@@ -1963,6 +1971,9 @@ function buildSafeOpenAIProxyCompatSuggestion(missing: string[]): Record<string,
   if (missing.includes("sendSessionAffinityHeaders")) {
     suggestion.sendSessionAffinityHeaders = true;
   }
+  // supportsLongCacheRetention is NOT suggested here — per spec it is
+  // optional/risky and must not appear in the copyable safe snippet.
+  // The proactive stripping in before_provider_request handles 400 prevention.
   return suggestion;
 }
 
@@ -1984,6 +1995,10 @@ function hasPromptCacheRetentionUnsupportedSignal(headers: Record<string, string
     "unknown parameter",
     "not supported",
     "unsupported field",
+    "extra inputs",
+    "not permitted",
+    "unrecognized",
+    "bad request",
   ].some((needle) => normalized.includes(needle));
 }
 
@@ -4714,6 +4729,308 @@ function locateModelInJsonc(
   };
 }
 
+/**
+ * Scan produced by `analyzeModelsJsonForMissingEntry` when
+ * `locateModelInJsonc` cannot find the target provider/model.
+ */
+type MissingEntryDiagnosis =
+  | { scenario: "provider_missing"; providersEnd: number }
+  | { scenario: "model_missing"; modelsEnd: number; providerBrace: number; providerEndBrace: number }
+  | { scenario: "provider_without_models"; providerBrace: number; providerEndBrace: number };
+
+/**
+ * Light second-pass scan that determines *why* `locateModelInJsonc` failed.
+ * Returns structured diagnostic so the fix handler can compose targeted
+ * guidance and an optional surgical insertion for API-logged-in models
+ * (e.g. opencode go) that never appear in `models.json`.
+ */
+function analyzeModelsJsonForMissingEntry(
+  text: string,
+  providerLabel: string,
+  modelId: string,
+): MissingEntryDiagnosis | undefined {
+  const clean = stripJsoncComments(text);
+  const rootBrace = skipJsonWhitespace(clean, 0);
+  if (clean[rootBrace] !== "{") return undefined;
+
+  const providersKey = findJsonObjectKey(clean, rootBrace, "providers");
+  if (!providersKey) {
+    // Root has no "providers" key at all — we don't auto-create one.
+    return undefined;
+  }
+  const providersBrace = skipJsonWhitespace(clean, providersKey.valueStart);
+  if (clean[providersBrace] !== "{") return undefined;
+  const providersEnd = findMatchingBracket(clean, providersBrace);
+  if (providersEnd === undefined) return undefined;
+
+  const providerKey = findJsonObjectKey(clean, providersBrace, providerLabel);
+  if (!providerKey || providerKey.keyStart > providersEnd) {
+    return { scenario: "provider_missing", providersEnd };
+  }
+
+  // Provider exists. Check for a models array so we know where to append.
+  const providerBrace = skipJsonWhitespace(clean, providerKey.valueStart);
+  if (clean[providerBrace] !== "{") return undefined;
+  const providerEndBrace = findMatchingBracket(clean, providerBrace);
+  if (providerEndBrace === undefined || providerEndBrace > providersEnd) return undefined;
+
+  const modelsKey = findJsonObjectKey(clean, providerBrace, "models");
+  if (modelsKey && modelsKey.keyStart < providerEndBrace) {
+    let mScan = skipJsonWhitespace(clean, modelsKey.valueStart);
+    if (clean[mScan] === "[") {
+      const modelsEnd = findMatchingBracket(clean, mScan);
+      if (modelsEnd !== undefined && modelsEnd <= providerEndBrace) {
+        return { scenario: "model_missing", modelsEnd, providerBrace, providerEndBrace };
+      }
+    }
+  }
+
+  // Provider exists, but there's no discoverable models array — treat as
+  // a provider that needs one.
+  return { scenario: "provider_without_models", providerBrace, providerEndBrace };
+}
+
+/**
+ * Build a copyable manual-edit snippet for the missing entry. Used when the
+ * terminal is non-interactive or the user chooses to edit by hand.
+ * Returns a complete provider→model→compat JSON block that the user can
+ * paste into `models.json` under `providers`.
+ */
+function formatMissingEntryManualSnippet(
+  providerLabel: string,
+  modelId: string,
+  compatKeys: Record<string, unknown>,
+): string {
+  const lines: string[] = [];
+  const sorted = Object.entries(compatKeys).sort(([a], [b]) => a.localeCompare(b));
+  const compatItems = sorted.map(([k, v]) => `      ${JSON.stringify(k)}: ${JSON.stringify(v)}`);
+  lines.push(`"${providerLabel}": {`);
+  lines.push(`    "models": [`);
+  lines.push(`      {`);
+  lines.push(`        "id": ${JSON.stringify(modelId)},`);
+  lines.push(`        "compat": {`);
+  lines.push(compatItems.join(",\n"));
+  lines.push(`        }`);
+  lines.push(`      }`);
+  lines.push(`    ]`);
+  lines.push(`  }`);
+  return lines.join("\n");
+}
+
+/**
+ * Surgically insert the missing provider/model entry into the original
+ * JSONC text. Returns the modified text and placement descriptor.
+ *
+ * Handles three scenarios:
+ * - `model_missing`: append a new model object to the provider's `models` array.
+ * - `provider_missing`: append a new provider block to the root `providers` object.
+ * - `provider_without_models`: inject a `"models": [...]` key into the existing provider.
+ */
+function composeMissingEntryInsertion(
+  originalText: string,
+  diagnosis: MissingEntryDiagnosis,
+  providerLabel: string,
+  modelId: string,
+  compatKeys: Record<string, unknown>,
+): { modifiedText: string; placementLabel: string } {
+  // Comments preserve length when stripped (`stripJsoncComments` replaces
+  // comment bytes 1-for-1 with spaces), so offsets derived from the
+  // comment-stripped text map cleanly back to the original. However,
+  // `lastIndexOf("{", pos)` / `lastIndexOf("[", pos)` must NOT be run
+  // against the raw original: a comment like `// add [more] here with a {
+  // brace` would surface `[` / `{` bytes that have no structural meaning,
+  // contaminating the `hasExisting`/`hasExistingElements` decision below
+  // and producing a stray leading comma. Run the structural searches
+  // against the comment-stripped version; keep the indentation lookups
+  // (which only care about newlines + leading whitespace) on the
+  // original since comments never contain forward-scan-relevant bytes.
+  const cleanText = stripJsoncComments(originalText);
+
+  // Resolve a sensible indentation step from an arbitrary byte offset in
+  // the original file.
+  const indentUnitAt = (offset: number): string => {
+    const ls = originalText.lastIndexOf("\n", offset);
+    const line = originalText.slice(ls < 0 ? 0 : ls + 1, offset);
+    const m = line.match(/^(\s+)/);
+    return m ? m[1] : "  ";
+  };
+
+  // Figure out the base indent from the insertion point's own line.
+  // Then derive inner indents (+1 and +2 levels).
+  const sorted = Object.entries(compatKeys).sort(([a], [b]) => a.localeCompare(b));
+  const formatCompactCompat = (indent: string): string => {
+    // Single-line compact when there's only one key, multi-line otherwise.
+    if (sorted.length === 1) {
+      const [k, v] = sorted[0];
+      return `{ ${JSON.stringify(k)}: ${JSON.stringify(v)} }`;
+    }
+    return (
+      "{\n" +
+      sorted.map(([k, v]) => `${indent}${JSON.stringify(k)}: ${JSON.stringify(v)}`).join(",\n") +
+      "\n" +
+      indent.slice(0, -2) +
+      "}"
+    );
+  };
+
+  if (diagnosis.scenario === "model_missing") {
+    // Append to the provider's models array, right before `]`.
+    const unit = indentUnitAt(diagnosis.modelsEnd);
+    const inner0 = unit + unit; // indent of model object's own keys
+    const inner1 = inner0 + unit; // indent of compat keys inside the model
+    const inner2 = inner1 + unit; // indent of compat values
+
+    // Determine whether the array is empty (need to skip the leading comma).
+    // Search for the models `[` on the comment-stripped text so a `[` inside
+    // a comment cannot be mistaken for the array opener.
+    const arrayInterior = cleanText.slice(
+      cleanText.lastIndexOf("[", diagnosis.modelsEnd) + 1,
+      diagnosis.modelsEnd,
+    ).trim();
+    const hasExistingElements = arrayInterior.length > 0;
+
+    const compatBlock = formatCompactCompat(inner2);
+    const modelBlock = [
+      hasExistingElements ? "," : "",
+      inner0 + "{",
+      inner1 + `"id": ${JSON.stringify(modelId)},`,
+      inner1 + `"compat": ` + compatBlock,
+      inner0 + "}",
+      unit,
+    ].filter(Boolean).join("\n");
+
+    const insertionPoint = diagnosis.modelsEnd;
+    const prefix = originalText.slice(0, insertionPoint);
+    const suffix = originalText.slice(insertionPoint); // starts with `]`
+    return {
+      modifiedText: prefix + modelBlock + suffix,
+      placementLabel: `providers["${providerLabel}"] -> models -> (new entry for "${modelId}")`,
+    };
+  }
+
+  if (diagnosis.scenario === "provider_missing") {
+    // Append a new provider entry to the root `providers` object, right
+    // before its closing `}`.
+    const unit = indentUnitAt(diagnosis.providersEnd);
+    const inner0 = unit + unit;
+    const inner1 = inner0 + unit;
+    const inner2 = inner1 + unit;
+    const inner3 = inner2 + unit;
+
+    const compatBlock = formatCompactCompat(inner3);
+    // Search for the providers `{` on the comment-stripped text so a `{`
+    // inside a comment cannot be mistaken for the providers object opener.
+    const providersInterior = cleanText.slice(
+      cleanText.lastIndexOf("{", diagnosis.providersEnd) + 1,
+      diagnosis.providersEnd,
+    ).trim();
+    const hasExisting = providersInterior.length > 0;
+
+    const providerBlock = [
+      hasExisting ? "," : "",
+      inner0 + `"${providerLabel}": {`,
+      inner1 + `"models": [`,
+      inner2 + "{",
+      inner3 + `"id": ${JSON.stringify(modelId)},`,
+      inner3 + `"compat": ` + compatBlock,
+      inner2 + "}",
+      inner1 + "]",
+      inner0 + "}",
+      unit,
+    ].filter(Boolean).join("\n");
+
+    const insertionPoint = diagnosis.providersEnd;
+    const prefix = originalText.slice(0, insertionPoint);
+    const suffix = originalText.slice(insertionPoint);
+    return {
+      modifiedText: prefix + providerBlock + suffix,
+      placementLabel: `providers -> (new entry "${providerLabel}")`,
+    };
+  }
+
+  // `provider_without_models`: inject a models array key into the
+  // existing provider block, right after the provider's opening `{`.
+  const unit = indentUnitAt(diagnosis.providerBrace);
+  const inner0 = unit + unit;
+  const inner1 = inner0 + unit;
+  const inner2 = inner1 + unit;
+
+  const compatBlock = formatCompactCompat(inner2);
+  const afterBrace = diagnosis.providerBrace + 1;
+  const modelsBlock = [
+    "",
+    inner0 + `"models": [`,
+    inner1 + "{",
+    inner2 + `"id": ${JSON.stringify(modelId)},`,
+    inner2 + `"compat": ` + compatBlock,
+    inner1 + "}",
+    inner0 + "],",
+    unit,
+  ].join("\n");
+
+  return {
+    modifiedText: originalText.slice(0, afterBrace) + modelsBlock + originalText.slice(afterBrace),
+    placementLabel: `providers["${providerLabel}"] -> (new "models" array with "${modelId}")`,
+  };
+}
+
+/**
+ * Lightweight self-check for a newly inserted entry.
+ * Parses the modified text as JSONC and confirms:
+ *   1. The target model exists under the provider.
+ *   2. Every compat key has the expected value (merged provider+model).
+ * Returns null on success, an error string on failure.
+ */
+function selfCheckMissingEntryInsertion(
+  originalText: string,
+  modifiedText: string,
+  providerLabel: string,
+  modelId: string,
+  compatKeys: Record<string, unknown>,
+): string | null {
+  try {
+    const modParsed = parseJsonc(modifiedText);
+    const providers = asRecord(asRecord(modParsed)?.providers);
+    if (!providers) return "Modified file: providers object missing or invalid";
+    const provider = asRecord(providers[providerLabel]);
+    if (!provider) return `Modified file: provider "${providerLabel}" not found`;
+    const models = provider.models;
+    if (!Array.isArray(models)) return `Modified file: provider "${providerLabel}".models is not an array`;
+    const targetModel = models.find((m: unknown) => asRecord(m)?.id === modelId);
+    if (!targetModel || typeof targetModel !== "object")
+      return `Modified file: model "${modelId}" not found in provider after insertion`;
+
+    // Validate effective merged compat
+    const provCompatRaw = (provider as Record<string, unknown>).compat;
+    const provCompat = (provCompatRaw && typeof provCompatRaw === "object" && !Array.isArray(provCompatRaw))
+      ? provCompatRaw as Record<string, unknown>
+      : {};
+    const mdlCompatRaw = (targetModel as Record<string, unknown>).compat;
+    const mdlCompat = (mdlCompatRaw && typeof mdlCompatRaw === "object" && !Array.isArray(mdlCompatRaw))
+      ? mdlCompatRaw as Record<string, unknown>
+      : {};
+    const merged = { ...provCompat, ...mdlCompat };
+    for (const [k, v] of Object.entries(compatKeys)) {
+      if (!(k in merged)) return `Modified file: effective compat.${k} not found`;
+      if (merged[k] !== v) return `Modified file: effective compat.${k} wrong value`;
+    }
+
+    if (modifiedText.length < originalText.length)
+      return "Modified file: content is shorter than original (possible truncation)";
+
+    const modClean = stripJsoncComments(modifiedText);
+    const rootStart = skipJsonWhitespace(modClean, 0);
+    const rootEnd = findMatchingBracket(modClean, rootStart);
+    if (rootEnd === undefined) return "Modified file: root bracket mismatch";
+    if (skipJsonWhitespace(modClean, rootEnd + 1) !== modClean.length)
+      return "Modified file: trailing content after root object";
+
+    return null;
+  } catch (e) {
+    return `Self-check error: ${e instanceof Error ? e.message : String(e)}`;
+  }
+}
+
 /**
  * Deep-equal comparison of two values, used for post-write self-check.
  * Compares all keys recursively, allowing `extraKeys` to be present in `a` but not in `b`.
@@ -4979,6 +5296,7 @@ function selfCheckFix(
   providerLabel: string,
   modelId: string,
   compatKeys: Record<string, unknown>,
+  placement: "provider" | "model" = "model",
 ): string | null {
   try {
     // Step 1: Parse both versions as JSONC (comments + trailing commas allowed).
@@ -5003,7 +5321,7 @@ function selfCheckFix(
     if (models.length === 0) {
       return `Modified file: provider "${providerLabel}".models is empty`;
     }
-    
+
     // Step 4: Find and validate target model
     const targetModel = models.find((m: Record<string, unknown>) => m.id === modelId);
     if (!targetModel || typeof targetModel !== 'object') {
@@ -5021,7 +5339,7 @@ function selfCheckFix(
     if (!origProvider || !origTargetModelRecord) {
       return `Original file: provider/model "${providerLabel}/${modelId}" not found`;
     }
-    
+
     // Step 5: Compute the EFFECTIVE merged compat (provider-level + model-level),
     // mirroring Pi's mergeCompat behavior (model wins on conflicts). The fix may
     // have written either level, so validation must check the merged result.
@@ -5045,7 +5363,7 @@ function selfCheckFix(
         return `Modified file: effective compat.${k} has wrong value: expected ${JSON.stringify(v)}, got ${JSON.stringify(mergedCompat[k])}`;
       }
     }
-    
+
     // Step 7: Validate original structure is preserved (no accidental deletions/changes)
 
     function isSubset(origVal: unknown, modVal: unknown, path = ''): boolean {
@@ -5071,7 +5389,18 @@ function selfCheckFix(
           } else {
             const origCompat = origObj[key] as Record<string, unknown>;
             const modCompat = modObj[key] as Record<string, unknown>;
-            const mayRepairThisCompat = origObj === origProvider || origObj === origTargetModelRecord;
+            // Only the compat object at the level ACTUALLY edited may have
+            // its values repaired by this fix. The un-edited level must
+            // remain byte/structure-equivalent, so its same-name keys stay
+            // under full validation. Using a disjunction OR (provider ||
+            // target) here would silently skip validation at the un-edited
+            // level, masking corruption (e.g. a buggy editor accidentally
+            // breaking provider.compat.sendSessionAffinityHeaders while
+            // the fix was a model-level repair). Track placement — only
+            // the placement-resolved object's own compat may be exempt.
+            const mayRepairThisCompat =
+              (placement === "provider" && origObj === origProvider) ||
+              (placement === "model" && origObj === origTargetModelRecord);
             for (const ck of Object.keys(origCompat)) {
               if (!(ck in modCompat)) return false;
               // The fix may repair an existing wrong compat value (for example
@@ -5092,13 +5421,16 @@ function selfCheckFix(
     if (!isSubset(origParsed, modParsed)) {
       return "Modified file: original structure was altered (data loss detected)";
     }
-    
-    // Step 8: Basic format sanity checks
-    if (modified.length < original.length) {
-      return "Modified file: content is shorter than original (possible truncation)";
-    }
-    
-    // Step 9: Validate root bracket integrity with the same string/comment-aware
+
+    // Note: we intentionally do NOT enforce `modified.length >= original.length`.
+    // The surgical editor may replace an existing compat value with a shorter one
+    // (e.g. `false` -> `true`), which legitimately shrinks the file by a byte.
+    // Real data loss / truncation is already caught by Step 7's isSubset
+    // (every original key still present) and Step 8's root-bracket integrity
+    // check below — a surviving length heuristic would false-positive on every
+    // such value repair. (Tracked: the mofas glm-5.2 self-check failure path.)
+
+    // Step 8: Validate root bracket integrity with the same string/comment-aware
     // scanner used for edits. Do not count raw braces: comments or strings may
     // legitimately contain unmatched `{` / `}` bytes.
     const modifiedClean = stripJsoncComments(modified);
@@ -5373,6 +5705,9 @@ export const __internals_for_tests = {
   locateModelInJsonc,
   composeFixInsertion,
   selfCheckFix,
+  analyzeModelsJsonForMissingEntry,
+  composeMissingEntryInsertion,
+  selfCheckMissingEntryInsertion,
   decideFixPlacement,
   chooseFixPlacement,
   findExistingCompatKeysInJsonc,
@@ -5791,9 +6126,56 @@ export default function (pi: ExtensionAPI) {
 
   ensureRoutingRegistry();
 
+  /**
+   * Check whether a model has an EXPLICIT supportsLongCacheRetention: true
+   * opt-in in models.json (either at provider-level or model-level).
+   * Model-level compat takes precedence over provider-level (mirrors Pi's
+   * mergeCompat behaviour: model wins on conflicts).
+   *
+   * Returns true ONLY when the user explicitly opted in. Returns false for:
+   *   - Explicit false (opt-out)
+   *   - In models.json but field absent (Pi defaults to true — unsafe)
+   *   - Not in models.json at all (API-logged-in providers)
+   *   - File missing/unreadable
+   *
+   * The caller strips prompt_cache_retention when this returns false.
+   */
+  function hasExplicitLongRetentionOptIn(model: PiModel): boolean {
+    try {
+      const text = readFileSync(MODELS_JSON_PATH, "utf8");
+      const parsed = parseJsonc(text);
+      const providers = asRecord(asRecord(parsed)?.providers);
+      if (!providers) return false;
+
+      const prov = asRecord(providers[model.provider]);
+      if (!prov) return false;
+
+      // Check model-level first (higher priority in Pi's merge logic)
+      const models = prov.models;
+      if (Array.isArray(models)) {
+        const modelEntry = models.find(m => asRecord(m)?.id === model.id);
+        if (modelEntry) {
+          const modelCompat = asRecord(asRecord(modelEntry)?.compat);
+          if (modelCompat?.supportsLongCacheRetention !== undefined) {
+            return modelCompat.supportsLongCacheRetention === true;
+          }
+        }
+      }
+
+      // Check provider-level
+      const provCompat = asRecord(prov.compat);
+      if (provCompat?.supportsLongCacheRetention !== undefined) {
+        return provCompat.supportsLongCacheRetention === true;
+      }
+
+      return false;
+    } catch {
+      return false;
+    }
+  }
+
   pi.on("session_start", async (event, ctx) => {
     await restoreCacheStats(event.reason, ctx);
-    if (runtimeOptimizerEnabled) notifyCacheCompatIfNeeded(resolveRouteModel(ctx.model, ctx) ?? ctx.model, ctx, warnedModels);
     await publishStatus(ctx);
   });
 
@@ -5915,6 +6297,39 @@ export default function (pi: ExtensionAPI) {
   });
 
   pi.on("before_provider_request", (event, ctx) => {
+    // ── Safety: strip prompt_cache_retention from payload for models that
+    // are not authorised to send it. Pi defaults supportsLongCacheRetention
+    // to true for all openai-completions models, but most third-party APIs
+    // reject the parameter with 400 “Extra inputs are not permitted”.
+    //
+    // Gate order (first match wins):
+    //   1. Official OpenAI          → keep (trusted to support it)
+    //   2. 400 history              → strip (empirical evidence overrides user config)
+    //   3. Explicit opt-in in models.json → keep (user explicitly wants it)
+    //   4. Everything else          → strip (safe default for third-party APIs)
+    //
+    // Gate 2 before Gate 3 is critical: if a user explicitly opted in but
+    // the API returned 400, we must strip — otherwise the 400 repeats forever.
+    if (runtimeOptimizerEnabled) {
+      const payload = event.payload as UnknownRecord;
+      if (payload && typeof payload.prompt_cache_retention === 'string') {
+        const rModel = resolveRouteModel(ctx.model, ctx) ?? ctx.model;
+        if (rModel) {
+          if (isOfficialOpenAIBaseUrl(rModel)) {
+            // Gate 1: Official OpenAI → keep
+          } else if (promptCacheRetention400Models.has(modelKey(rModel))) {
+            // Gate 2: 400 history → strip (overrides user opt-in)
+            delete payload.prompt_cache_retention;
+          } else if (hasExplicitLongRetentionOptIn(rModel)) {
+            // Gate 3: Explicit user opt-in → keep
+          } else {
+            // Gate 4: Safe default → strip
+            delete payload.prompt_cache_retention;
+          }
+        }
+      }
+    }
+
     if (!shouldInjectOpenAIPromptCacheKey()) return undefined;
     const requestModel = resolveRouteModel(ctx.model, ctx) ?? ctx.model;
     if (!isOpenAICompatibleApi(requestModel?.api)) return undefined;
@@ -6110,7 +6525,23 @@ export default function (pi: ExtensionAPI) {
           return;
         }
 
-        const suggestion = buildFixSuggestion(model);
+        let suggestion = buildFixSuggestion(model);
+
+        // If no regular missing compat flags but the model has a recorded
+        // prompt_cache_retention 400 (Pi sent `prompt_cache_retention` and
+        // the provider rejected it), offer to override
+        // `supportsLongCacheRetention` to false in models.json.
+        if (!suggestion && isPromptCacheRetention400Applicable(model) && promptCacheRetention400Models.has(modelKey(model))) {
+          const key = modelKey(model);
+          const slashIdx = key.indexOf("/");
+          const providerLabel = slashIdx > 0 ? key.slice(0, slashIdx) : key;
+          suggestion = {
+            providerLabel,
+            modelId: model.id,
+            compatKeys: { supportsLongCacheRetention: false },
+          };
+        }
+
         if (!suggestion) {
           const key = modelKey(model);
           cmdCtx.ui.notify(`✅ Nothing to fix for "${key}". Compat already configured.`, "info");
@@ -6120,24 +6551,40 @@ export default function (pi: ExtensionAPI) {
         if (!cmdCtx.hasUI) {
           // No UI — refuse to write, show manual guidance instead.
           const compatResult = buildCompatDiagnosis(model);
-          if (compatResult) {
-            cmdCtx.ui.notify(
-              `❌ Non-interactive terminal detected. Auto-fix requires UI confirmation.\n\n` +
-              `Manual steps:\n` +
-              `1. Open ${getModelsJsonDisplayPath()} in your editor.\n` +
-              `2. Go to providers["${suggestion.providerLabel}"] -> models -> entry with id "${suggestion.modelId}" -> compat.\n` +
-              `3. Add the missing keys:\n${formatCompatKeysForInsertion(suggestion.compatKeys)}\n` +
-              `4. Save and run /reload.\n\n` +
-              compatResult,
-              "warning",
+          const snippet = formatMissingEntryManualSnippet(
+            suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys,
+          );
+          const manualLines = [
+            `❌ Non-interactive terminal detected. Auto-fix requires UI confirmation.`,
+            "",
+            `Edit ${getModelsJsonDisplayPath()} and run /reload.`,
+          ];
+          if (promptCacheRetention400Models.has(modelKey(model))) {
+            manualLines.push(
+              "",
+              "💡 This model returned HTTP 400 for prompt_cache_retention.",
+              "Create or edit the entry below to override supportsLongCacheRetention to false.",
             );
-          } else {
-            cmdCtx.ui.notify(
-              `❌ Non-interactive terminal detected. Auto-fix requires UI confirmation.\n` +
-              `Edit ${getModelsJsonDisplayPath()} manually and run /reload.`,
-              "warning",
+          }
+          manualLines.push(
+            "",
+            "If the provider/model already exists in models.json, add these compat keys under",
+            `providers["${suggestion.providerLabel}"] -> models -> entry with id "${suggestion.modelId}" -> compat:`,
+            formatCompatKeysForInsertion(suggestion.compatKeys),
+          );
+          if (snippet.length > 0) {
+            manualLines.push(
+              "",
+              "If the provider/model is missing (common for API-logged-in channels such as",
+              `opencode go), add a minimal entry under "providers" (keep existing auth as-is):`,
+              "",
+              snippet,
             );
           }
+          if (compatResult) {
+            manualLines.push("", compatResult);
+          }
+          cmdCtx.ui.notify(manualLines.join("\n"), "warning");
           return;
         }
 
@@ -6150,17 +6597,127 @@ export default function (pi: ExtensionAPI) {
           return;
         }
 
-        // Locate the model entry
+        // Locate the model entry. API-logged-in providers (e.g. opencode go)
+        // may not appear in models.json at all.
         const location = locateModelInJsonc(originalText, suggestion.providerLabel, suggestion.modelId);
         if (!location) {
-          cmdCtx.ui.notify(
-            `❌ Could not locate model "${suggestion.modelId}" in ${getModelsJsonDisplayPath()}.\n` +
-            `The JSONC scanner could not confidently find the target entry.\n` +
-            `Manual edit required: open the file, find providers["${suggestion.providerLabel}"] -> models, and add:\n` +
-            `${formatCompatKeysForInsertion(suggestion.compatKeys)}\n` +
-            `Then run /reload.`,
-            "warning",
-          );
+          const diagnosis = analyzeModelsJsonForMissingEntry(originalText, suggestion.providerLabel, suggestion.modelId);
+          if (diagnosis && cmdCtx.hasUI) {
+            // Offer to create the missing entry.
+            const plan = composeMissingEntryInsertion(
+              originalText, diagnosis,
+              suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys,
+            );
+            const checkError = selfCheckMissingEntryInsertion(
+              originalText, plan.modifiedText,
+              suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys,
+            );
+            if (checkError !== null) {
+              // Fall through to manual guidance.
+              cmdCtx.ui.notify(
+                `❌ Self-check would fail for auto-created entry: ${checkError}\n` +
+                `Falling back to manual guidance. No changes were made.`,
+                "error",
+              );
+              // Continue to manual guidance below.
+            } else {
+              const keysPreview = JSON.stringify(suggestion.compatKeys, null, 2);
+              const ts = backupTimestamp();
+              const backupPath = `${MODELS_JSON_PATH}.backup-cache-optimizer-${ts}`;
+              const previewLines = [
+                `📝 Preview of changes to ${getModelsJsonDisplayPath()}:`,
+                ``,
+                `Location: ${plan.placementLabel}`,
+                `Compat JSON to write:`,
+                keysPreview,
+                ``,
+                `⚠️  Risk notice:`,
+                `  1. This creates a new entry in models.json. Existing auth (e.g. login API tokens) is not affected.`,
+                `  2. A timestamped backup will be written to: ${backupPath}`,
+                `  3. You must run /reload or restart Pi for the change to take effect.`,
+                `  4. If the file contains comments or unusual formatting, please verify the result after write.`,
+              ];
+              if (promptCacheRetention400Models.has(modelKey(model))) {
+                previewLines.push(
+                  "",
+                  "💡  This fix overrides supportsLongCacheRetention to false because",
+                  "a 400 prompt_cache_retention error was observed for this model.",
+                  "After applying and reloading, Pi will no longer send the",
+                  "prompt_cache_retention parameter to this provider.",
+                );
+              }
+              previewLines.push("", `Apply these changes?`);
+              const confirmed = await cmdCtx.ui.confirm("Cache Optimizer — Fix (new entry)", previewLines.join("\n"));
+              if (confirmed) {
+                try {
+                  await copyFile(MODELS_JSON_PATH, backupPath);
+                  const tempPath = `${MODELS_JSON_PATH}.${process.pid}.${Date.now()}.fix.tmp`;
+                  await writeFile(tempPath, plan.modifiedText, "utf8");
+                  await rename(tempPath, MODELS_JSON_PATH);
+
+                  const writtenText = await readFile(MODELS_JSON_PATH, "utf8");
+                  const postErr = selfCheckMissingEntryInsertion(
+                    originalText, writtenText,
+                    suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys,
+                  );
+                  if (postErr !== null) {
+                    await copyFile(backupPath, MODELS_JSON_PATH);
+                    cmdCtx.ui.notify(
+                      `❌ Post-write self-check failed: ${postErr}\n` +
+                      `The backup at ${backupPath} has been restored. No changes applied.`,
+                      "error",
+                    );
+                    return;
+                  }
+                  cmdCtx.ui.notify(
+                    `✅ Fix applied to ${getModelsJsonDisplayPath()}.\n` +
+                    `Backup saved to: ${backupPath}\n` +
+                    `Run /reload or restart Pi for the change to take effect.`,
+                    "info",
+                  );
+                } catch (e) {
+                  cmdCtx.ui.notify(
+                    `❌ Write failed: ${e instanceof Error ? e.message : String(e)}`,
+                    "error",
+                  );
+                }
+                return;
+              }
+              cmdCtx.ui.notify("No changes were made. Canceled by user.", "info");
+              return;
+            }
+          }
+
+          // Non-interactive or no diagnosis: show manual guidance.
+          const snippet = diagnosis
+            ? formatMissingEntryManualSnippet(suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys)
+            : formatCompatKeysForInsertion(suggestion.compatKeys);
+          const adviceLines: string[] = [];
+          if (!diagnosis) {
+            adviceLines.push(
+              `❌ Could not locate model "${suggestion.modelId}" or provider "${suggestion.providerLabel}" in ${getModelsJsonDisplayPath()}.`,
+              "",
+              "Providers that were added via Pi /login API (e.g. opencode go) do not have",
+              "entries in models.json. You can create a minimal compat-only entry by hand:",
+            );
+          } else if (diagnosis.scenario === "provider_missing") {
+            adviceLines.push(
+              `ℹ️ Provider "${suggestion.providerLabel}" does not exist in ${getModelsJsonDisplayPath()}.`,
+              `This is common for API-logged-in providers (e.g. /login ...).`,
+              "",
+              "Add the following minimal block under the \"providers\" key (keep your",
+              "existing authentication as-is):",
+            );
+          } else {
+            adviceLines.push(
+              `ℹ️ Model "${suggestion.modelId}" was not found in ${getModelsJsonDisplayPath()}`,
+              `under providers["${suggestion.providerLabel}"].`,
+              "",
+              "Add the following entry to the models array (keep existing auth):",
+            );
+          }
+          adviceLines.push("", snippet, "", "Then save and run /reload.");
+          cmdCtx.ui.notify(adviceLines.join("\n"), "warning");
           return;
         }
 
@@ -6170,7 +6727,7 @@ export default function (pi: ExtensionAPI) {
         const modifiedText = composeFixInsertion(originalText, location, suggestion.compatKeys, decision.placement);
 
         // Self-check
-        const checkError = selfCheckFix(originalText, modifiedText, suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys);
+        const checkError = selfCheckFix(originalText, modifiedText, suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys, decision.placement);
         if (checkError !== null) {
           cmdCtx.ui.notify(
             `❌ Self-check failed before write: ${checkError}\n` +
@@ -6203,15 +6760,23 @@ export default function (pi: ExtensionAPI) {
           `Placement: ${decision.placement} level — ${decision.reason}`,
           `Compat JSON to write:`,
           keysPreview,
-          ``, 
+          ``,
           `⚠️  Risk notice:`,
           scopeRiskLine,
           `  2. A timestamped backup will be written to: ${backupPath}`,
           `  3. You must restart Pi / run /reload for the change to take effect.`,
           `  4. If the file contains comments or unusual formatting, please verify the result after write.`,
-          ``, 
-          `Apply these changes?`,
         ];
+        if (promptCacheRetention400Models.has(modelKey(model))) {
+          previewLines.push(
+            "",
+            "💡  This fix overrides supportsLongCacheRetention to false because",
+            "a 400 prompt_cache_retention error was observed for this model.",
+            "After applying and reloading, Pi will no longer send the",
+            "prompt_cache_retention parameter to this provider.",
+          );
+        }
+        previewLines.push("", `Apply these changes?`);
 
         const confirmed = await cmdCtx.ui.confirm("Cache Optimizer — Fix", previewLines.join("\n"));
         if (!confirmed) {
@@ -6231,7 +6796,7 @@ export default function (pi: ExtensionAPI) {
 
           // Post-write self-check (read back)
           const writtenText = await readFile(MODELS_JSON_PATH, "utf8");
-          const postCheckError = selfCheckFix(originalText, writtenText, suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys);
+          const postCheckError = selfCheckFix(originalText, writtenText, suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys, decision.placement);
           if (postCheckError !== null) {
             // Restore from backup
             await copyFile(backupPath, MODELS_JSON_PATH);
@@ -6360,7 +6925,7 @@ export default function (pi: ExtensionAPI) {
 
             const menuDecision = chooseFixPlacement(originalText, location, suggestion.compatKeys, suggestion.providerLabel);
             const modifiedText = composeFixInsertion(originalText, location, suggestion.compatKeys, menuDecision.placement);
-            const checkError = selfCheckFix(originalText, modifiedText, suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys);
+            const checkError = selfCheckFix(originalText, modifiedText, suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys, menuDecision.placement);
             if (checkError !== null) {
               cmdCtx.ui.notify(`❌ Self-check failed: ${checkError}\nNo changes made.`, "error");
               return;
@@ -6383,13 +6948,13 @@ export default function (pi: ExtensionAPI) {
               `Placement: ${menuDecision.placement} level — ${menuDecision.reason}`,
               `Compat JSON to write:`,
               keysPreview,
-              ``, 
+              ``,
               `⚠️  Risk notice:`,
               menuScopeRiskLine,
               `  2. A timestamped backup will be written to: ${backupPath}`,
               `  3. You must restart Pi / run /reload for the change to take effect.`,
               `  4. If the file contains comments, verify the result after write.`,
-              ``, 
+              ``,
               `Apply these changes?`,
             ];
 
@@ -6406,7 +6971,7 @@ export default function (pi: ExtensionAPI) {
               await rename(tempPath, MODELS_JSON_PATH);
 
               const writtenText = await readFile(MODELS_JSON_PATH, "utf8");
-              const postCheck = selfCheckFix(originalText, writtenText, suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys);
+              const postCheck = selfCheckFix(originalText, writtenText, suggestion.providerLabel, suggestion.modelId, suggestion.compatKeys, menuDecision.placement);
               if (postCheck !== null) {
                 await copyFile(backupPath, MODELS_JSON_PATH);
                 cmdCtx.ui.notify(`❌ Post-write check failed: ${postCheck}\nBackup restored.`, "error");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-cache-optimizer",
-  "version": "2.6.7",
+  "version": "2.6.10",
   "description": "Improve Pi prompt/KV cache hit rates with stable prompts, OpenAI-compatible cache keys, proxy compat warnings, and footer cache stats.",
   "keywords": [
     "pi-package",