@modelstatus/cli 0.1.87 → 0.1.89

package/README.md

@@ -75,6 +75,20 @@ place. Boundary-safe (`gpt-4` never rewrites inside `gpt-4o`), style-preserving
 it follows the chain to the first live model. In the TUI, press `f` for a
 red/green diff preview; nothing is written until you confirm.
 
+### Free: or hand the fixes to your AI coding agent
+
+```bash
+mm prompt [dir]              # print a fix-it prompt for Claude Code, Cursor, …
+mm prompt [dir] | pbcopy     # straight to your clipboard
+```
+
+Prints one self-contained prompt: every retired/retiring reference (file:line +
+the exact string in code), the chain-resolved replacement, and the rewrite
+rules. Paste it into your agent and let it do the work — unlike `mm fix`, an
+agent can also re-pin dated ids (`-20250514`), follow config indirection, and
+adjust parameters the new model needs. In the TUI, press `p` on the Here or
+Inventory tab to copy the same prompt to your clipboard.
+
 ### Sign in for cloud features
 
 ```bash
@@ -92,6 +106,7 @@ You get two binaries — `mm` (short) and `llmstatus` (descriptive). Same binary
 |---|---|
 | `mm status [dir]` | Free offline model-health check — no account |
 | `mm fix [dir]` | Rewrite dying model ids to their replacement (`--dry-run` to preview) |
+| `mm prompt [dir]` | Print a fix-it prompt for your AI coding agent (`mm prompt \| pbcopy`) |
 | `mm [dir]` | Launch the TUI on a folder (defaults to the current one) — runs locally |
 | `mm update` | Update the binary in place (Homebrew installs: `brew upgrade`) |
 | `mm login [api_key]` | Browser sign-in with polling (or paste a key) |
@@ -125,6 +140,7 @@ Secret sources shell out to your already-authenticated CLIs, run **read-only**,
 | Resolve + health locally, on-device | ✓ | ✓ |
 | Secret-source aware (`env`, `aws-secrets`, `k8s`, `helm`, `sql`) | ✓ | ✓ |
 | `mm fix` — rewrite dying ids to replacements | ✓ | ✓ |
+| `mm prompt` — fix-it prompt for your AI agent | ✓ | ✓ |
 | Cloud inventory across projects/teams | — | ✓ |
 | GitHub App: PR checks + one-click fix PRs | — | ✓ |
 | Alerts on deprecations/retirements (email/Slack/SMS) | — | ✓ |

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@modelstatus/cli",
-  "version": "0.1.87",
+  "version": "0.1.89",
   "description": "Track which AI models you use, where, and never get surprised by a retirement. Free offline model-health for any repo (mm status), browser sign-in for cloud inventory + alerts.",
   "keywords": [
     "llm",
@@ -38,7 +38,8 @@
     "LICENSE"
   ],
   "publishConfig": {
-    "access": "public"
+    "access": "public",
+    "provenance": false
   },
   "engines": {
     "node": ">=18"

package/src/changelog-data.js

@@ -1,6 +1,17 @@
 /* GENERATED by scripts/gen-changelog.mjs from apps/web/lib/changelog.json — do not edit.
  * Release notes baked into the binary (in-TUI + the on-load what's-new card). */
 export const CHANGELOG = [
+  {
+    "version": "0.1.88",
+    "date": "2026-08-22",
+    "title": "Hand the fixes to your AI coding agent",
+    "items": [
+      "New `mm prompt [dir]`: prints one self-contained prompt describing every retired/retiring model reference in your repo (file:line, the exact string in code, the chain-resolved replacement) plus rewrite rules. Paste it into Claude Code, Cursor, or any coding agent — or pipe it straight to your clipboard: `mm prompt | pbcopy`.",
+      "Unlike `mm fix` (a mechanical in-place rewrite), an agent can also re-pin dated ids like `-20250514`, follow config indirection, and adjust parameters the new model needs — the prompt spells out those rules for it.",
+      "TUI: press `p` on the Here tab or the Inventory tab to copy the same prompt to your clipboard (with a file fallback when no clipboard tool exists). On the Inventory tab the `/` filter scopes it — search `prod`, press `p`, get a prod-only prompt.",
+      "`mm status` and `mm fix` now point at `mm prompt` when there's something an auto-rewrite can't safely touch (like version-pinned ids)."
+    ]
+  },
   {
     "version": "0.1.87",
     "date": "2026-07-07",

package/src/clipboard.js

@@ -0,0 +1,42 @@
+/* Deliver generated text from inside the TUI, where stdout belongs to ink:
+ * copy to the system clipboard via the platform's native tool, or — when no
+ * clipboard tool exists (headless, bare containers) — write a file and hand
+ * back its path. Best-effort and synchronous (the payloads are a few KB).
+ * MM_NO_CLIPBOARD=1 skips the copy (tests + users who hate surprise copies);
+ * MM_PROMPT_DIR redirects the fallback file (tests). */
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+import { spawnSync } from "node:child_process";
+
+const CANDIDATES = process.platform === "darwin"
+  ? [["pbcopy", []]]
+  : process.platform === "win32"
+    ? [["clip", []]]
+    : [["wl-copy", []], ["xclip", ["-selection", "clipboard"]], ["xsel", ["--clipboard", "--input"]]];
+
+/** True when the text landed on the system clipboard. */
+export function copyToClipboard(text) {
+  if (process.env.MM_NO_CLIPBOARD === "1") return false;
+  for (const [cmd, args] of CANDIDATES) {
+    try {
+      const r = spawnSync(cmd, args, { input: text, stdio: ["pipe", "ignore", "ignore"], timeout: 3000 });
+      if (r.status === 0) return true;
+    } catch { /* tool missing — try the next */ }
+  }
+  return false;
+}
+
+/** Clipboard first, file fallback. Returns { method: "clipboard" } or
+ * { method: "file", path } — never throws (a failed file write returns
+ * { method: "none" } so the caller can say so instead of crashing the TUI). */
+export function deliverText(text, { filename = "mm-fix-prompt.md" } = {}) {
+  if (copyToClipboard(text)) return { method: "clipboard" };
+  try {
+    const p = path.join(process.env.MM_PROMPT_DIR || os.tmpdir(), filename);
+    fs.writeFileSync(p, text);
+    return { method: "file", path: p };
+  } catch {
+    return { method: "none" };
+  }
+}

package/src/fix-prompt.js

@@ -0,0 +1,121 @@
+/* Build a copy-paste prompt for an AI coding agent (Claude Code, Cursor, …) to
+ * fix dying model references. `mm fix` is the mechanical path — boundary-safe
+ * string swaps at known file:lines — but it deliberately skips what a dumb
+ * rewrite can't do safely: version-pinned ids, config indirection, parameter
+ * changes the new model needs. An agent CAN do those, so the prompt hands it
+ * the full picture: every dying reference (file:line + the exact string), the
+ * chain-resolved replacement, and the rewrite rules fix.js enforces.
+ *
+ * Used by `mm prompt` (stdout), the Here tab (`p`) and the Inv tab (`p`).
+ * Pure text-building — no I/O, no registry access — so every surface renders
+ * the identical prompt and tests stay hermetic.
+ */
+
+// Same severity order the TUI/status sort by: already-dead first, then aging.
+const TIER = { retired: 0, withdrawn: 0, retiring: 1, deprecating: 2 };
+
+const REFS_CAP = 100; // per model — an agent doesn't need ref #101 to get it
+
+/** One reference line: "- path:line — currently `gpt-4o`" (repo-prefixed when
+ * the ref came from another checkout, so a cross-repo inventory stays honest). */
+function refLine(r) {
+  const loc = `${r.repo ? r.repo + " · " : ""}${r.file}${r.line ? ":" + r.line : ""}`;
+  return `- ${loc}${r.matched ? ` — currently \`${r.matched}\`` : ""}`;
+}
+
+/** Dedupe a model's refs (same file:line:string appears once; at one file:line
+ * a matched string that's a substring of a sibling's is dropped — the detector
+ * can emit both "gpt-4o" and "openai/gpt-4o" for the same spot). */
+function dedupeRefs(refs) {
+  const seen = new Map();
+  for (const r of refs || []) {
+    if (!r?.file) continue;
+    const key = `${r.repo || ""}|${r.file}:${r.line ?? ""}:${(r.matched || "").toLowerCase()}`;
+    if (!seen.has(key)) seen.set(key, r);
+  }
+  const list = [...seen.values()];
+  return list.filter(
+    (r) => !list.some(
+      (q) => q !== r && q.file === r.file && q.line === r.line && (q.repo || "") === (r.repo || "")
+        && q.matched && r.matched && q.matched.toLowerCase() !== r.matched.toLowerCase()
+        && q.matched.toLowerCase().includes(r.matched.toLowerCase()),
+    ),
+  );
+}
+
+/** Human health phrase for a section header. */
+function healthPhrase(f) {
+  if (f.health === "retired") return f.retires_date ? `retired ${f.retires_date}` : "retired";
+  if (f.health === "withdrawn") return "withdrawn by the provider";
+  return `${f.health}${f.retires_date ? `, retires ${f.retires_date}` : ""}`;
+}
+
+/** Dedupe + drop empty findings + sort — the rows the prompt actually renders. */
+function buildRows(findings) {
+  return (findings || [])
+    .map((f) => ({ ...f, refs: dedupeRefs(f.refs) }))
+    .filter((f) => f.refs.length)
+    .sort((a, b) => (TIER[a.health] ?? 3) - (TIER[b.health] ?? 3)
+      || String(a.retires_date || "9999").localeCompare(String(b.retires_date || "9999"))
+      || String(a.slug || a.display).localeCompare(String(b.slug || b.display)));
+}
+
+/** The model/reference counts the rendered prompt will actually contain (after
+ * dedupe and empty-finding drops) — use these for any summary line or toast so
+ * it never disagrees with the prompt body. */
+export function promptCounts(findings) {
+  const rows = buildRows(findings);
+  return { models: rows.length, refs: rows.reduce((n, f) => n + f.refs.length, 0) };
+}
+
+/**
+ * Render the prompt. `findings` = [{ slug, display, health, retires_date,
+ * replacement, refs: [{ file, line, matched, repo? }] }] where `replacement`
+ * is the write-form target ("provider/canonical_id", terminalReplacement's
+ * output) or null when the registry has no successor yet. Findings with zero
+ * usable refs are dropped; returns null when nothing remains.
+ */
+export function buildFixPrompt(findings, { date = new Date() } = {}) {
+  const rows = buildRows(findings);
+  if (!rows.length) return null;
+
+  // Local calendar date — toISOString() is UTC and stamps "tomorrow" for
+  // anyone west of Greenwich in the evening.
+  const day = `${date.getFullYear()}-${String(date.getMonth() + 1).padStart(2, "0")}-${String(date.getDate()).padStart(2, "0")}`;
+  const nRefs = rows.reduce((n, f) => n + f.refs.length, 0);
+  const out = [];
+  out.push("# Task: update retired / retiring AI model references");
+  out.push("");
+  out.push(`This codebase references ${rows.length === 1 ? "an AI model" : `${rows.length} AI models`} that ${rows.length === 1 ? "is" : "are"} retired or scheduled for retirement (checked against the LLM Status registry on ${day}). Update every reference listed below — ${nRefs} in total.`);
+  out.push("");
+  out.push("## Models to replace");
+
+  for (const f of rows) {
+    const name = f.display && f.display !== f.slug ? `${f.display} (\`${f.slug}\`)` : `\`${f.slug || f.display}\``;
+    out.push("");
+    out.push(`### ${name} — ${healthPhrase(f)}`);
+    if (f.replacement) {
+      const i = f.replacement.indexOf("/");
+      const bare = i >= 0 ? f.replacement.slice(i + 1) : f.replacement;
+      out.push(bare !== f.replacement
+        ? `Replace with: \`${bare}\` (provider-prefixed form: \`${f.replacement}\`).`
+        : `Replace with: \`${f.replacement}\`.`);
+    } else {
+      out.push("No registry replacement is listed yet — pick the provider's closest current successor, apply it consistently, and note the choice in your summary.");
+    }
+    out.push("References:");
+    for (const r of f.refs.slice(0, REFS_CAP)) out.push(refLine(r));
+    if (f.refs.length > REFS_CAP) out.push(`- …and ${f.refs.length - REFS_CAP} more (rewrite every occurrence of the old id, not just the lines listed)`);
+  }
+
+  out.push("");
+  out.push("## Rules");
+  out.push("");
+  out.push("1. Preserve each call site's existing style: keep a bare id bare (`claude-sonnet-4-5`) and a provider-prefixed id prefixed (`anthropic/claude-sonnet-4-5`); don't change quoting or surrounding formatting.");
+  out.push("2. Version-pinned ids (dated like `-20250514`, `@`-pins, Bedrock `-v1:0`, fine-tune `:suffix`): switch to the replacement model's own current id — pinned only if the platform requires a pin. Never graft the old model's version pin onto the new id.");
+  out.push("3. If a model id comes from a config file, env var, constant, or database row, update it at the source — not just the call sites listed above.");
+  out.push("4. Review adjacent parameters that may need adjusting for the new model (max output tokens, temperature constraints, flags the old model needed).");
+  out.push("5. When you're done, run the project's tests, then run `mm status` (the LLM Status CLI, https://llmstatus.ai) to confirm no retired or retiring references remain.");
+  out.push("");
+  return out.join("\n");
+}

package/src/index.js

@@ -774,9 +774,79 @@ async function cmdFix(positional, flags) {
   console.log(`\n✓ rewrote ${res.applied.length} reference(s) in ${new Set(res.applied.map((p) => p.file)).size} file(s).`);
   for (const s of res.stale) console.log(`  ! skipped ${s.file}:${s.line} — ${s.error}`);
   for (const f of res.failed) console.log(`  × failed ${f.file}:${f.line} — ${f.error}`);
+  if (res.stale.some((s) => /pinned variant/.test(s.error || ""))) {
+    console.log("\n→ pinned ids need a human (or an agent): `mm prompt` prints a fix-it prompt for your AI coding agent.");
+  }
   if (res.applied.length) console.log("\nRe-run your tests, then `mm status` to confirm everything reads current.");
 }
 
+/** `mm prompt [dir]` — print a copy-paste prompt for an AI coding agent to fix
+ * dying model refs. Where `mm fix` mechanically rewrites what it can prove safe,
+ * the prompt hands EVERYTHING to an agent — including version-pinned ids, models
+ * with no registry replacement yet, and the config/parameter follow-ups a string
+ * swap can't do. stdout is the prompt and nothing else (pipe it: `mm prompt |
+ * pbcopy`); progress + the summary line go to stderr. */
+async function cmdPrompt(positional, flags) {
+  const dir = path.resolve(positional[1] || flags.dir || ".");
+  requireDirectory(dir);
+  const { getRegistry } = await import("./registry/fetch.js");
+  const { resolveLocal, computeHealth, dropResolvedFragments } = await import("./registry/local.js");
+  const { terminalReplacement } = await import("./fix.js");
+  const { buildFixPrompt, promptCounts } = await import("./fix-prompt.js");
+  const prog = startProgress(true, "fetching the model registry…");
+  const snapshot = await getRegistry({
+    offline: !!flags.offline || process.env.LLMSTATUS_REGISTRY_OFFLINE === "1",
+    cacheFile: process.env.LLMSTATUS_REGISTRY_CACHE || undefined,
+    log: (m) => prog.log(m),
+  });
+
+  prog.update("scanning for model references…");
+  const onProgress = ({ filesScanned, candidates: c }) => prog.update(`scanning… ${filesScanned} files, ${c} reference(s)`);
+  let candidates = await collectFrom(["filesystem"], { root: dir }, snapshot.detection, new Set(), onProgress);
+  prog.stop();
+  const resolved = resolveLocal(snapshot, [...new Set(candidates.map((c) => c.model_string))]);
+  const byStr = new Map(resolved.map((r) => [r.input.toLowerCase(), r]));
+  candidates = dropResolvedFragments(candidates, (c) => !!byStr.get(c.model_string.toLowerCase())?.model_slug);
+  const today = new Date();
+
+  // Group EVERY dying model's refs — unlike cmdFix, keep models with no
+  // replacement (the agent picks a successor) and pinned refs (it re-pins).
+  const byModel = new Map(); // slug -> { model, health, refs }
+  for (const c of candidates) {
+    const r = byStr.get(c.model_string.toLowerCase());
+    if (!r?.model_slug || !r.model) continue;
+    const health = computeHealth(r.model, 90, today);
+    if (health === "ok") continue;
+    if (flags.model && r.model_slug !== flags.model) continue;
+    const e = byModel.get(r.model_slug) || { model: r.model, health, refs: [] };
+    e.refs.push(c);
+    byModel.set(r.model_slug, e);
+  }
+
+  const bySlug = new Map(snapshot.models.map((m) => [m.slug, m]));
+  const isCurrent = (m) => computeHealth(m, 90, today) === "ok";
+  const findings = [...byModel.values()].map(({ model, health, refs }) => ({
+    slug: model.slug,
+    display: model.display,
+    health,
+    retires_date: model.retires_date,
+    replacement: model.replacement_slug
+      ? terminalReplacement(model.replacement_slug, (slug) => bySlug.get(slug) ?? null, isCurrent)
+      : null,
+    refs: refs.map((c) => ({ file: c.source_path || c.location_label, line: c.source_line, matched: c.model_string })),
+  }));
+
+  const text = buildFixPrompt(findings, { date: today });
+  if (!text) {
+    process.stderr.write("Nothing to fix — every recognized model here is current. No prompt to generate.\n");
+    return;
+  }
+  console.log(text);
+  const counts = promptCounts(findings);
+  const copyTool = process.platform === "darwin" ? "pbcopy" : process.platform === "win32" ? "clip" : "xclip -selection clipboard";
+  process.stderr.write(`\n(${counts.models} model(s), ${counts.refs} reference(s). Copy it straight to your clipboard: mm prompt | ${copyTool})\n`);
+}
+
 /** List detection sources and whether each can run right now. Live integrations
  * also show their on/off toggle (the `int` column) so toggled state is visible
  * here too. */
@@ -791,7 +861,7 @@ async function cmdSources(_positional, flags) {
   }
   console.log("\nScan with: mm scan --sources env,aws-secrets,k8s,helm,sql   (or --sources all)");
   console.log("Integrations (toggle with `mm integrations enable <id>`): " + INTEGRATION_IDS.join(", "));
-  console.log("Options:   --region <r> · --namespace <ns> · --kube-context <c> · --db <pg-dsn> --sql-table <t>");
+  console.log("Options:   --region <r> · --namespace <ns> · --kube-context <c> · --db <pg-dsn> · --sql-table <t>");
   console.log("           --vercel-project <p> · --vercel-team <t> · --gh-repo <owner/name> · --supabase-ref <ref>");
   console.log("Safety:    secret VALUES never leave your machine — only model ids upload. Use --dry-run to preview.");
 }
@@ -973,6 +1043,7 @@ async function cmdStatus(positional, flags) {
     const fixable = attention.filter((r) => r.model.replacement_slug).length;
     const tips = [];
     if (fixable) tips.push(`\`mm fix\` rewrites the ${fixable} with a known replacement`);
+    tips.push("`mm prompt` prints a fix-it prompt for your AI coding agent (Claude Code, Cursor, …)");
     if (!loadConfig().apiKey) tips.push("`mm login` to get alerted before these dates (Pro)");
     if (tips.length) console.log("\n" + tips.map((t) => "→ " + t).join("\n"));
   }
@@ -989,6 +1060,7 @@ Usage:
   mm config                View or change settings (analytics, …)
   mm scan [dir]            Scan for model usage; interactive TUI, or --ci/--json for pipelines
   mm fix [dir]             Rewrite dying model ids to their replacement, in place (--dry-run previews; --model <slug> limits; --yes skips the confirm)
+  mm prompt [dir]          Print a fix-it prompt for your AI coding agent — paste it into Claude Code, Cursor, … and let it do the rewrites (mm prompt | pbcopy)
   mm ci [dir]              CI gate: fail the build on deprecated/retiring models (GitHub annotations)
                           (--fail-on <none|deprecating|retiring|retired> sets the threshold, default retired;
                            --json-out <file> writes findings JSON; --diff <base> limits to changed files, auto on PRs)
@@ -1011,7 +1083,7 @@ Scan sources (--sources; default = filesystem + enabled integrations; "all" for
   and only ever upload model ids — secret VALUES never leave your machine.
 
 Flags: --update · --api <url> · --key <key> · --project <id|name> · --dir <path> · --yes · --json · --ci · --dry-run
-       --offline (use the cached registry) · --model <slug> (fix) · --fail-on <t> · --json-out <file> (ci) · --all · --force (clear)
+       --offline (use the cached registry) · --model <slug> (fix, prompt) · --fail-on <t> · --json-out <file> (ci) · --all · --force (clear)
        --sources <list> · --region <r> · --namespace <ns> · --kube-context <c> · --db <dsn> · --sql-table <t>
        --vercel-project <p> · --vercel-team <t> · --gh-repo <owner/name> · --supabase-ref <ref>
 
@@ -1042,6 +1114,19 @@ const COMMAND_HELP = {
   --yes             skip the confirmation prompt
   --model <slug>    only fix this one model (full provider/slug)
   --offline         use the cached registry only`,
+  prompt: `mm prompt [dir]   Print a copy-paste prompt for an AI coding agent to fix dying model refs.
+
+  Scans like \`mm status\`, then prints one self-contained prompt to stdout: every
+  retired/retiring model reference (file:line + the exact string in code), the
+  chain-resolved replacement, and the rewrite rules. Paste it into Claude Code,
+  Cursor, or any coding agent. Unlike \`mm fix\` (a mechanical in-place rewrite),
+  an agent can also re-pin dated ids, follow config indirection, and adjust
+  parameters the new model needs. stdout is ONLY the prompt — pipe it:
+  mm prompt | pbcopy
+
+  --model <slug>    only include this one model (full provider/slug)
+  --offline         use the cached registry only
+  --dir <path>      directory to scan (alternative to the positional arg)`,
   ci: `mm ci [dir]   CI gate: fail the build on deprecated/retiring models.
 
   Exits non-zero when a finding is at/above --fail-on. Emits GitHub annotations +
@@ -1217,6 +1302,7 @@ async function main() {
     }
     else if (cmd === "scan") await cmdScan(positional, flags);
     else if (cmd === "fix") await cmdFix(positional, flags);
+    else if (cmd === "prompt") await cmdPrompt(positional, flags);
     else if (cmd === "ci") await cmdCi(positional, flags);
     else if (cmd === "status") await cmdStatus(positional, flags);
     else if (cmd === "sources") await cmdSources(positional, flags);

package/src/telemetry.js

@@ -75,8 +75,8 @@ export function analyticsState() {
  * paths") includes the path the user points mm at. Keep in sync with the
  * dispatch in index.js main(). */
 const KNOWN_COMMANDS = new Set([
-  "login", "signup", "logout", "config", "analytics", "scan", "fix", "ci",
-  "status", "sources", "integrations", "clear", "play", "upgrade", "tui",
+  "login", "signup", "logout", "config", "analytics", "scan", "fix", "prompt",
+  "ci", "status", "sources", "integrations", "clear", "play", "upgrade", "tui",
   "update", "version", "help",
 ]);
 

package/src/tui/views/inventory.js

@@ -24,6 +24,7 @@ export const meta = {
     { k: "g", label: "refresh" },
     { k: "r", label: "rescan" },
     { k: "n", label: "new" },
+    { k: "p", label: "llm prompt" },
     { k: "e", label: "env" },
     { k: "t", label: "tag" },
     { k: "c", label: "critical" },
@@ -100,6 +101,61 @@ export function InventoryView({ client, ui, dir = ".", active, width = 78, heigh
     ui?.reportStatus?.({ counts, context: query ? `${filtered.length} of ${usages.length}` : `${usages.length} tracked` });
   }, [usages, ui, query, filtered.length, q.error, q.loading]);
 
+  /** p: build the LLM fix prompt for every dying usage in the CURRENT view
+   * (the / filter scopes it — search "prod", press p, get a prod-only prompt)
+   * and put it on the clipboard (file fallback). Inventory rows can point at
+   * code on other machines, which is exactly why the prompt exists: paste it
+   * into the agent that lives WITH that code. Registry lookup is best-effort
+   * (cached snapshot) to turn replacement slugs into real API ids; without a
+   * cache the slug is still an unambiguous instruction for an agent. */
+  async function makeLlmPrompt() {
+    const dying = filtered.filter((u) => u.health && u.health !== "ok" && u.health !== "custom");
+    if (!dying.length) return ui.showToast(query ? "no matching usages need fixing" : "all current — nothing to fix", "#16a34a");
+    try {
+      const [{ terminalReplacement }, { buildFixPrompt, promptCounts }, { deliverText }, { computeHealth }] = await Promise.all([
+        import("../../fix.js"), import("../../fix-prompt.js"), import("../../clipboard.js"), import("../../registry/local.js"),
+      ]);
+      const today = new Date();
+      let bySlug = new Map();
+      try {
+        const { getRegistry } = await import("../../registry/fetch.js");
+        const snap = await getRegistry({ offline: true, cacheFile: process.env.LLMSTATUS_REGISTRY_CACHE || undefined });
+        bySlug = new Map((snap.models || []).map((m) => [m.slug, m]));
+      } catch { /* no cached registry — replacement slugs still name the target */ }
+      const isCurrent = (m) => computeHealth(m, 90, today) === "ok";
+      const byModel = new Map(); // one section per model, refs merged across projects
+      for (const u of dying) {
+        const key = u.model_display || u.custom_model_name || u.canonical_id || "?";
+        const e = byModel.get(key) || {
+          slug: u.canonical_id || key,
+          display: u.model_display || key,
+          health: u.health,
+          retires_date: u.retires_date ? String(u.retires_date).slice(0, 10) : null,
+          replacement: u.replacement_slug
+            ? terminalReplacement(u.replacement_slug, (slug) => bySlug.get(slug) ?? null, isCurrent)
+            : null,
+          refs: [],
+        };
+        e.refs.push({
+          file: u.source_path || u.location_label || "(added manually — find its usages)",
+          line: u.source_line,
+          matched: u.custom_model_name || u.canonical_id,
+          repo: u.source_repo || undefined,
+        });
+        byModel.set(key, e);
+      }
+      const text = buildFixPrompt([...byModel.values()], { date: today });
+      if (!text) return ui.showToast("no locatable references to hand off", "#d97706");
+      const counts = promptCounts([...byModel.values()]);
+      const res = deliverText(text);
+      if (res.method === "clipboard") ui.showToast(`${GLYPH.check} LLM fix prompt copied (${counts.models} model${counts.models === 1 ? "" : "s"} · ${counts.refs} refs) — paste into your AI agent`);
+      else if (res.method === "file") ui.showToast(`${GLYPH.check} couldn't copy — prompt saved to ${res.path}`);
+      else ui.showToast("couldn't copy or save the prompt", "#dc2626");
+    } catch (e) {
+      ui.showToast(e.message, "red");
+    }
+  }
+
   async function patch(u, body, label) {
     try {
       await client.patchUsage(u.id, body);
@@ -128,6 +184,7 @@ export function InventoryView({ client, ui, dir = ".", active, width = 78, heigh
       if (input === "g") return q.reload();
       if (input === "r") return ui.switchTo("scan");
       if (input === "n") return ui.switchTo("add");
+      if (input === "p") return makeLlmPrompt();
       if (input === "t") {
         const untagged = usages.filter(isUntagged);
         if (!untagged.length) return ui.showToast("nothing untagged");

package/src/tui/views/local.js

@@ -29,6 +29,7 @@ export const meta = {
     { k: "↑↓", label: "nav" },
     { k: "↵", label: "refs" },
     { k: "f", label: "fix all" },
+    { k: "p", label: "llm prompt" },
     { k: "u", label: "push → Inv" },
     { k: "g", label: "rescan" },
     { k: "/", label: "search" },
@@ -134,6 +135,39 @@ export function LocalView({ client, me, dir, ui, width = 78, height = 14, active
     });
   }
 
+  /** p (while idle): build the LLM fix prompt for EVERY dying model found here
+   * and put it on the clipboard (file fallback) — for the refs `mm fix` can't
+   * safely touch (pinned ids, config indirection) hand the whole job to the
+   * user's AI coding agent instead. Same builder as `mm prompt`. */
+  function makeLlmPrompt() {
+    const dying = items.filter((it) => it.model && it.health !== "ok" && it.health !== "custom");
+    if (!dying.length) return ui?.showToast?.("all current — nothing to fix", "#16a34a");
+    Promise.all([import("../../fix.js"), import("../../fix-prompt.js"), import("../../clipboard.js")]).then(
+      ([{ terminalReplacement }, { buildFixPrompt, promptCounts }, { deliverText }]) => {
+        const bySlug = new Map((scan.snapshot?.models || []).map((m) => [m.slug, m]));
+        const today = new Date();
+        const isCurrent = (m) => computeHealth(m, 90, today) === "ok";
+        const findings = dying.map((it) => ({
+          slug: it.model.slug,
+          display: it.model.display,
+          health: it.health,
+          retires_date: it.model.retires_date,
+          replacement: it.model.replacement_slug
+            ? terminalReplacement(it.model.replacement_slug, (slug) => bySlug.get(slug) ?? null, isCurrent)
+            : null,
+          refs: distinctRefs(it.refs).map((r) => ({ file: r.source_path || r.location_label, line: r.source_line, matched: r.model_string })),
+        }));
+        const text = buildFixPrompt(findings, { date: today });
+        if (!text) return ui?.showToast?.("no file references to hand off", "#d97706");
+        const counts = promptCounts(findings);
+        const res = deliverText(text);
+        if (res.method === "clipboard") ui?.showToast?.(`${GLYPH.check} LLM fix prompt copied (${counts.models} model${counts.models === 1 ? "" : "s"} · ${counts.refs} refs) — paste into your AI agent`);
+        else if (res.method === "file") ui?.showToast?.(`${GLYPH.check} couldn't copy — prompt saved to ${res.path}`);
+        else ui?.showToast?.("couldn't copy or save the prompt", "#dc2626");
+      },
+    );
+  }
+
   const tick = useTick(80, running || pushing);
   const spin = SPINNER[tick % SPINNER.length];
   const search = useSearch();
@@ -281,7 +315,9 @@ export function LocalView({ client, me, dir, ui, width = 78, height = 14, active
       if (key.upArrow || input === "k") return nav.up();
       if (input === "e") return excludeRef(drefs[0]); // exclude the highlighted model's location (editable)
       if (input === "f") return fixRefs(cur?.refs || [], `all ${cur?.count ?? 0} references`);
-      if (input === "p" && running) return scan.togglePause();
+      // p is overloaded by state: pausing only means anything mid-scan, and the
+      // prompt is only trustworthy once the scan is done — no key collision.
+      if (input === "p") return running ? scan.togglePause() : makeLlmPrompt();
       if (input === "g") { justReloadedRef.current = true; return scan.reload(); }
       if (input === "u" && !pushing) return pushToInventory();
     },