@blockrun/franklin 3.15.56 → 3.15.58

package/dist/agent/error-classifier.d.ts

@@ -15,5 +15,14 @@ export interface AgentErrorInfo {
     maxRetries?: number;
     /** User-facing suggestion for how to recover. Appended to error message in UI. */
     suggestion?: string;
+    /**
+     * Upstream-recommended wait time before retrying. Parsed from a
+     * `[retry-after-ms=...]` tag the streaming client appends to the error
+     * message when the response carries a `Retry-After` header (typically
+     * 429 / 503). The agent loop should honor this in place of its
+     * default exponential backoff. Capped at 10 minutes upstream so a
+     * malicious or buggy server can't pin the agent indefinitely.
+     */
+    retryAfterMs?: number;
 }
 export declare function classifyAgentError(message: string): AgentErrorInfo;

package/dist/agent/error-classifier.js

@@ -11,6 +11,17 @@ function includesAny(text, patterns) {
 }
 export function classifyAgentError(message) {
     const err = message.toLowerCase();
+    // Extract Retry-After hint that streaming-client appended (see llm.ts
+    // 429 path). Surfaces on the AgentErrorInfo so the loop can honor the
+    // upstream's recommended wait instead of guessing with exponential
+    // backoff.
+    let retryAfterMs;
+    const retryAfterTag = /\[retry-after-ms=(\d+)\]/i.exec(message);
+    if (retryAfterTag) {
+        const n = parseInt(retryAfterTag[1], 10);
+        if (Number.isFinite(n) && n > 0 && n <= 600_000)
+            retryAfterMs = n;
+    }
     // payment_rejected — the gateway received a SIGNED payment header and
     // rejected it during verification (signature mismatch, replay-nonce
     // reuse, clock skew, wrong-chain wallet). Different remedy from
@@ -76,6 +87,7 @@ export function classifyAgentError(message) {
         return {
             category: 'rate_limit', label: 'RateLimit', isTransient: true, maxRetries: 1,
             suggestion: 'Try /model to switch to a different model, or wait a moment and /retry.',
+            retryAfterMs,
         };
     }
     if (includesAny(err, [

package/dist/agent/llm.js

@@ -430,7 +430,25 @@ export class ModelClient {
             }
             if (!response.ok) {
                 const errorBody = await response.text().catch(() => 'unknown error');
-                const message = extractApiErrorMessage(errorBody);
+                let message = extractApiErrorMessage(errorBody);
+                // 429 with Retry-After header: tag the error message so the
+                // classifier can extract and the loop can honor it. Verified
+                // 2026-05-04 in a live session: a 429 fired with the loop's
+                // exponential backoff (~1-2s) but the upstream's actual
+                // Retry-After window was ~30s — the agent retried prematurely
+                // and burned its rate_limit retry budget. Anthropic + most
+                // gateways send Retry-After as either seconds (integer) or an
+                // HTTP-date; we only honor the seconds form (the date form is
+                // rare in practice and harder to validate against clock skew).
+                if (response.status === 429) {
+                    const retryAfter = response.headers.get('retry-after');
+                    if (retryAfter) {
+                        const seconds = parseInt(retryAfter, 10);
+                        if (Number.isFinite(seconds) && seconds > 0 && seconds <= 600) {
+                            message = `${message} [retry-after-ms=${seconds * 1000}]`;
+                        }
+                    }
+                }
                 // Runtime tool_choice retry. The static allowlist at line ~35
                 // catches the case where the request goes directly to a model
                 // whose name contains `deepseek-reasoner` / `openai/o1` /

package/dist/agent/loop.js

@@ -1156,8 +1156,17 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                         }
                     }
                     recoveryAttempts++;
-                    const backoffMs = getBackoffDelay(recoveryAttempts);
-                    logger.warn(`[franklin] ${classified.label} error — retrying in ${(backoffMs / 1000).toFixed(1)}s (attempt ${recoveryAttempts}/${effectiveMaxRetries}): ${errMsg.slice(0, 100)}`);
+                    // Honor an upstream Retry-After (parsed from the response by
+                    // llm.ts when 429+ Retry-After is present) over our own
+                    // exponential backoff. Verified 2026-05-04: a 429 with
+                    // Retry-After=30s was retried after ~1.5s exponential backoff
+                    // → got 429 again → burned the rate_limit retry budget. Cap at
+                    // 30s so the agent never feels "frozen" — anything longer
+                    // falls back to a different model instead.
+                    const upstreamWaitMs = classified.retryAfterMs;
+                    const honorUpstream = typeof upstreamWaitMs === 'number' && upstreamWaitMs <= 30_000;
+                    const backoffMs = honorUpstream ? upstreamWaitMs : getBackoffDelay(recoveryAttempts);
+                    logger.warn(`[franklin] ${classified.label} error — retrying in ${(backoffMs / 1000).toFixed(1)}s (attempt ${recoveryAttempts}/${effectiveMaxRetries})${honorUpstream ? ' (upstream Retry-After)' : ''}: ${errMsg.slice(0, 100)}`);
                     // Surface the actual error + model so the user can see which model
                     // is failing and what the upstream said. Old "Retrying after Server
                     // error" was uninformative — users couldn't tell whether to wait,
@@ -1230,7 +1239,18 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                     }
                 }
                 // ── Unrecoverable: show error with suggestion from classifier ──
-                const suggestion = classified.suggestion ? `\nTip: ${classified.suggestion}` : '';
+                // For rate_limit specifically, augment the classifier's generic
+                // suggestion with an explicit "all free models exhausted — switch
+                // to a paid model" hint when we got here because pickFreeFallback
+                // returned null. Verified 2026-05-04: the screenshot's session
+                // ended with a bare "[RateLimit] API error: 429" because every
+                // free model had already been ruled out earlier in the turn —
+                // the user had a funded wallet but no signal that paid models
+                // were the way out.
+                let suggestion = classified.suggestion ? `\nTip: ${classified.suggestion}` : '';
+                if (classified.category === 'rate_limit' && turnFailedModels.size > 0) {
+                    suggestion = `\nTip: All free models tried this turn are rate-limited. Switch to a paid model with /model anthropic/claude-sonnet-4.6 (or any other paid model) and retry — your wallet handles it. Or wait ~60s and /retry the same turn.`;
+                }
                 onEvent({
                     kind: 'turn_done',
                     reason: 'error',

package/dist/commands/content.d.ts

@@ -0,0 +1,17 @@
+/**
+ * `franklin content` CLI surface — human-facing read access to the
+ * Content library that lives at ~/.blockrun/content.json.
+ *
+ * Tools (ContentCreate / ContentAddAsset) write the library during agent
+ * sessions; before this command, there was no way to see the resulting
+ * spend without scripting against the JSON file. Verified 2026-05-04 in
+ * a live session: user asked "我花了多少钱做这个", agent ran
+ * `franklin content list` and got "no content subcommand", fell back to
+ * estimating from memory.
+ *
+ * Subcommands:
+ *   - list             : table of id, type, title, status, spent/budget, assets
+ *   - show <idOrPrefix>: full detail of one Content, including each asset
+ */
+import { Command } from 'commander';
+export declare function buildContentCommand(): Command;

package/dist/commands/content.js

@@ -0,0 +1,141 @@
+/**
+ * `franklin content` CLI surface — human-facing read access to the
+ * Content library that lives at ~/.blockrun/content.json.
+ *
+ * Tools (ContentCreate / ContentAddAsset) write the library during agent
+ * sessions; before this command, there was no way to see the resulting
+ * spend without scripting against the JSON file. Verified 2026-05-04 in
+ * a live session: user asked "我花了多少钱做这个", agent ran
+ * `franklin content list` and got "no content subcommand", fell back to
+ * estimating from memory.
+ *
+ * Subcommands:
+ *   - list             : table of id, type, title, status, spent/budget, assets
+ *   - show <idOrPrefix>: full detail of one Content, including each asset
+ */
+import os from 'node:os';
+import path from 'node:path';
+import { Command } from 'commander';
+import { loadLibrary } from '../content/store.js';
+const DEFAULT_PATH = path.join(os.homedir(), '.blockrun', 'content.json');
+function fmtUsd(n) {
+    return `$${n.toFixed(2)}`;
+}
+/**
+ * Resolve a user-typed id-or-prefix to a single Content record. Returns
+ * either the matching record or an error message; the caller prints.
+ *
+ * Accepts the full UUID, a prefix (≥4 chars), or — for convenience — a
+ * substring match against the title (case-insensitive). Ambiguity returns
+ * the candidates so the user can disambiguate without rerunning blind.
+ */
+function resolveContent(all, input) {
+    const q = input.trim();
+    if (!q)
+        return { error: 'Provide an id, id-prefix, or title substring.' };
+    const exact = all.find(c => c.id === q);
+    if (exact)
+        return { found: exact };
+    const prefix = q.length >= 4 ? all.filter(c => c.id.startsWith(q)) : [];
+    if (prefix.length === 1)
+        return { found: prefix[0] };
+    if (prefix.length > 1) {
+        return { error: `Ambiguous prefix "${q}" — matches:\n${prefix.map(c => `  ${c.id}  ${c.title}`).join('\n')}` };
+    }
+    const lower = q.toLowerCase();
+    const titled = all.filter(c => c.title.toLowerCase().includes(lower));
+    if (titled.length === 1)
+        return { found: titled[0] };
+    if (titled.length > 1) {
+        return { error: `Ambiguous title "${q}" — matches:\n${titled.map(c => `  ${c.id}  ${c.title}`).join('\n')}` };
+    }
+    return { error: `No Content matches "${q}".` };
+}
+export function buildContentCommand() {
+    const cmd = new Command('content').description('Inspect Content library (assets, spend, budget)');
+    cmd
+        .command('list')
+        .description('List all Content records, newest first')
+        .action(() => {
+        const lib = loadLibrary(DEFAULT_PATH);
+        if (!lib) {
+            console.log('No Content library yet. Tools like ContentCreate populate it during agent sessions.');
+            return;
+        }
+        const all = lib.list();
+        if (all.length === 0) {
+            console.log('No Content records.');
+            return;
+        }
+        // Header + rows. Truncate id to 8-char prefix and title to 40 chars
+        // so common terminal widths (80-100) fit a row on one line.
+        console.log(['id'.padEnd(8), 'type'.padEnd(8), 'status'.padEnd(10), 'spent/cap'.padEnd(13), 'assets', 'title'].join('  '));
+        for (const c of all) {
+            const id8 = c.id.slice(0, 8);
+            const spend = `${fmtUsd(c.spentUsd)}/${fmtUsd(c.budgetUsd)}`;
+            const title = c.title.length > 40 ? c.title.slice(0, 39) + '…' : c.title;
+            console.log([
+                id8.padEnd(8),
+                c.type.padEnd(8),
+                c.status.padEnd(10),
+                spend.padEnd(13),
+                String(c.assets.length).padEnd(6),
+                title,
+            ].join('  '));
+        }
+        // Footer with rolled-up spend.
+        const totalSpent = all.reduce((s, c) => s + c.spentUsd, 0);
+        const totalBudget = all.reduce((s, c) => s + c.budgetUsd, 0);
+        console.log();
+        console.log(`Total: ${fmtUsd(totalSpent)} spent across ${all.length} content${all.length === 1 ? '' : 's'} (cap ${fmtUsd(totalBudget)}).`);
+    });
+    cmd
+        .command('show <idOrPrefix>')
+        .description('Show full detail for one Content record (id, prefix, or title substring)')
+        .action((input) => {
+        const lib = loadLibrary(DEFAULT_PATH);
+        if (!lib) {
+            console.log('No Content library yet.');
+            process.exit(1);
+        }
+        const result = resolveContent(lib.list(), input);
+        if ('error' in result) {
+            console.error(result.error);
+            process.exit(1);
+        }
+        const c = result.found;
+        console.log(`# ${c.title}`);
+        console.log();
+        console.log(`id:        ${c.id}`);
+        console.log(`type:      ${c.type}`);
+        console.log(`status:    ${c.status}`);
+        console.log(`spent:     ${fmtUsd(c.spentUsd)} / ${fmtUsd(c.budgetUsd)} cap`);
+        console.log(`created:   ${new Date(c.createdAt).toISOString()}`);
+        if (c.publishedAt)
+            console.log(`published: ${new Date(c.publishedAt).toISOString()}`);
+        console.log();
+        if (c.assets.length > 0) {
+            console.log(`## Assets (${c.assets.length})`);
+            for (const a of c.assets) {
+                console.log(`- ${a.kind.padEnd(6)} ${fmtUsd(a.costUsd).padStart(7)}  ${a.source}`);
+                console.log(`    ${a.data}`);
+            }
+            console.log();
+        }
+        if (c.drafts.length > 0) {
+            console.log(`## Drafts (${c.drafts.length})`);
+            c.drafts.forEach((d, i) => {
+                const preview = d.text.length > 80 ? d.text.slice(0, 79) + '…' : d.text;
+                console.log(`- #${i + 1}  ${preview}`);
+            });
+            console.log();
+        }
+        if (c.distribution.length > 0) {
+            console.log(`## Distribution (${c.distribution.length})`);
+            for (const dist of c.distribution) {
+                console.log(`- ${dist.channel}${dist.url ? `  ${dist.url}` : ''}  (${new Date(dist.at).toISOString()})`);
+            }
+        }
+    });
+    return cmd;
+}

package/dist/index.js

@@ -24,6 +24,7 @@ import { initCommand } from './commands/init.js';
 import { uninitCommand } from './commands/uninit.js';
 import { proxyCommand } from './commands/proxy.js';
 import { buildTaskCommand } from './commands/task.js';
+import { buildContentCommand } from './commands/content.js';
 import { VERSION as version } from './config.js';
 const program = new Command();
 program
@@ -221,6 +222,10 @@ program
 // `franklin task <subcmd>` — human-facing CLI for detached background tasks.
 // Defined in src/commands/task.ts; subcommands: list, tail, cancel, wait.
 program.addCommand(buildTaskCommand());
+// `franklin content <subcmd>` — read access to the Content library
+// (~/.blockrun/content.json) so users + agent shell-outs can inspect
+// spend without scripting against the JSON file. Subcommands: list, show.
+program.addCommand(buildContentCommand());
 // Hidden internal subcommand — invoked by startDetachedTask via spawn(detached).
 // The underscore prefix signals "not for humans"; we still register it via
 // commander so exit codes and arg parsing stay consistent with the rest of

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.15.56",
+  "version": "3.15.58",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {