rewritable 0.6.0 → 0.7.0

package/README.md

@@ -125,10 +125,10 @@ The agent loop retries up to 3 times when the model emits plain text instead of
 
 | Flag | Effect |
 |---|---|
-| `--backend <name>` | `openrouter` (default), `ollama`, `lmstudio`. Falls back to `$RWA_BACKEND`. `bridge` is browser-only by design. |
+| `--backend <name>` | `openrouter` (default), `ollama`, `lmstudio`, `atomic`. Falls back to `$RWA_BACKEND`. `bridge` is browser-only by design. |
 | `--model <id>` | model id passed to the backend. Falls back to `$RWA_MODEL`, then `google/gemini-3.5-flash`. |
-| `--base-url <url>` | OpenAI-compatible base URL override. Defaults: `https://openrouter.ai/api/v1`, `http://localhost:11434/v1` (or `$RWA_OLLAMA_URL`), `http://localhost:1234/v1` (or `$RWA_LMSTUDIO_URL`). |
-| `--api-key <key>` | openrouter only; falls back to `$RWA_OPENROUTER_KEY`. ollama / lmstudio run locally without auth. |
+| `--base-url <url>` | OpenAI-compatible base URL override. Defaults: `https://openrouter.ai/api/v1`, `http://localhost:11434/v1` (or `$RWA_OLLAMA_URL`), `http://localhost:1234/v1` (or `$RWA_LMSTUDIO_URL`), `http://127.0.0.1:1337/v1` (or `$RWA_ATOMIC_URL`). |
+| `--api-key <key>` | openrouter only; falls back to `$RWA_OPENROUTER_KEY`. ollama / lmstudio / atomic run locally without auth. |
 
 #### Other edit flags
 

package/bin/rwa.mjs

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 import { newCmd, importCmd, version, KNOWN_KINDS, openWithPrefill, SEED_CANDIDATES } from '../src/commands.mjs';
-import { resolveApiKey } from '../src/backend.mjs';
+import { resolveApiKey, backendMaxTokens } from '../src/backend.mjs';
 import { parseCreateArgs, createCmd } from '../src/create.mjs';
 import { relative } from 'node:path';
 
@@ -101,7 +101,7 @@ Flags:
   --open, -o     open the resulting file in the default app. First-paint
                  sessionStorage is pre-populated from env / ./.env:
                    OPENROUTER_API_KEY → ?key=…    (lifted into rwa_apikey)
-                   RWA_BACKEND        → ?backend= (openrouter|ollama|lmstudio|bridge)
+                   RWA_BACKEND        → ?backend= (openrouter|ollama|lmstudio|atomic|bridge)
                    RWA_MODEL          → ?model=…  (model name string)
                  The bootstrap lifts each into sessionStorage and scrubs the
                  URL bar on first paint, so the values don't sit in history.
@@ -137,8 +137,8 @@ Flags:
                  the raw body; on failure, the \`{code, subcode, details}\`
                  object goes to stderr.
   --backend <n>  (edit instruction path / skin --l1) backend name. One of:
-                 openrouter (default), ollama, lmstudio. Falls back to
-                 \$RWA_BACKEND if unset.
+                 openrouter (default), ollama, lmstudio, atomic. Falls back
+                 to \$RWA_BACKEND if unset.
   --model <id>   (edit instruction path / skin --l1) model id passed to the
                  backend. Falls back to \$RWA_MODEL, then a
                  sensible default for the backend.
@@ -248,6 +248,7 @@ function envBaseUrl(name) {
     case 'openrouter': return 'https://openrouter.ai/api/v1';
     case 'ollama':     return process.env.RWA_OLLAMA_URL || 'http://localhost:11434/v1';
     case 'lmstudio':   return process.env.RWA_LMSTUDIO_URL || 'http://localhost:1234/v1';
+    case 'atomic':     return process.env.RWA_ATOMIC_URL || 'http://127.0.0.1:1337/v1';
     default:           return undefined;
   }
 }
@@ -377,7 +378,7 @@ function detectProductKind(fileText) {
 
         // Reject unknown backends fast. `bridge` is browser-only by design
         // (single-shot via web_cli_bridge); the CLI has no equivalent.
-        if (!['openrouter', 'ollama', 'lmstudio'].includes(backendName)) {
+        if (!['openrouter', 'ollama', 'lmstudio', 'atomic'].includes(backendName)) {
           emitEdit({ code: 'usage_error', subcode: 'unknown_backend', details: { backend: backendName } }, jsonMode);
           process.exitCode = 1;
           return;
@@ -457,7 +458,7 @@ function detectProductKind(fileText) {
             currentDoc: promptDoc,
             instruction,
             frozenZoneNames,
-            backend: { baseUrl, model: modelId, apiKey },
+            backend: { baseUrl, model: modelId, apiKey, maxTokens: backendMaxTokens(backendName) },
             onRetry: r => {
               if (jsonMode) {
                 process.stderr.write(JSON.stringify({ phase: 'retry', attempt: r.attempt, reason: r.reason }) + '\n');
@@ -845,7 +846,7 @@ function detectProductKind(fileText) {
         const baseUrl     = baseUrlFlag.value || envBaseUrl(backendName);
         const apiKey      = resolveApiKey(backendName, apiKeyFlag.value);
 
-        if (!['openrouter', 'ollama', 'lmstudio'].includes(backendName)) {
+        if (!['openrouter', 'ollama', 'lmstudio', 'atomic'].includes(backendName)) {
           emitSkin({ code: 'usage_error', subcode: 'unknown_backend', details: { backend: backendName } });
           process.exitCode = 1; return;
         }
@@ -882,7 +883,7 @@ function detectProductKind(fileText) {
           result = await skinCmdL1(filePath, action, {
             systemPrompt,
             toolSchemas: TOOL_SCHEMAS,
-            backend: { baseUrl, model: modelId, apiKey },
+            backend: { baseUrl, model: modelId, apiKey, maxTokens: backendMaxTokens(backendName) },
             onRetry: r => {
               if (jsonMode) process.stderr.write(JSON.stringify({ phase: 'retry', attempt: r.attempt, reason: r.reason }) + '\n');
               else process.stderr.write(`rwa skin: attempt ${r.attempt}/3 retrying — ${r.reason}\n`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "rewritable",
-  "version": "0.6.0",
+  "version": "0.7.0",
   "description": "CLI for re-writeable: emit and import single-file rwa documents.",
   "type": "module",
   "bin": {

package/seeds/rewritable.html

@@ -380,6 +380,13 @@ const RWA = {
   // not block these requests even when the container is served over HTTPS.
   DEFAULT_OLLAMA_URL:'http://localhost:11434/v1',
   DEFAULT_LMSTUDIO_URL:'http://localhost:1234/v1',
+  // atomic.chat: a local OpenAI-compatible inference server (MLX-backed on
+  // Apple Silicon) with the same /v1/chat/completions + /v1/models shape and
+  // real multi-turn tool_calls. CORS note: it allows http(s) page origins out
+  // of the box but NOT the file:// null origin — serve the container from a
+  // local origin (or a hosted projection) to use it.
+  DEFAULT_ATOMIC_URL:'http://127.0.0.1:1337/v1',
+  K_BASE_URL_ATOMIC:'rwa_base_url_atomic',
   // Optional alternative agent backend: a localhost CLI bridge
   // (https://github.com/martintreiber/web_cli_bridge style) that runs
   // arbitrary shell commands. When the user picks "bridge" in settings, ⌘K
@@ -411,7 +418,7 @@ try {
     qs.delete('key');
   }
   const b = qs.get('backend');
-  if (b && ['openrouter','ollama','lmstudio','bridge','bridge-session'].includes(b)) {
+  if (b && ['openrouter','ollama','lmstudio','atomic','bridge','bridge-session'].includes(b)) {
     sessionStorage.setItem(RWA.K_BACKEND, b);
     qs.delete('backend');
   }
@@ -1117,7 +1124,7 @@ function buildUI() {
   <button class="rwa-st-btn pri" id="rwa-st-commit">⌘S</button>
 </div>
 <div id="rwa-set-panel">
-  <div class="rwa-set-row"><label>Backend</label><select id="rwa-backend"><option value="openrouter">OpenRouter (API key)</option><option value="ollama">Ollama (localhost)</option><option value="lmstudio">LM Studio (localhost)</option><option value="bridge">Bridge (claude -p, localhost)</option><option value="bridge-session">Bridge session (claude, persistent)</option></select></div>
+  <div class="rwa-set-row"><label>Backend</label><select id="rwa-backend"><option value="openrouter">OpenRouter (API key)</option><option value="ollama">Ollama (localhost)</option><option value="lmstudio">LM Studio (localhost)</option><option value="atomic">atomic.chat (localhost)</option><option value="bridge">Bridge (claude -p, localhost)</option><option value="bridge-session">Bridge session (claude, persistent)</option></select></div>
   <div class="rwa-set-row" id="rwa-set-row-key"><label>OpenRouter Key</label><input type="password" id="rwa-key" placeholder="sk-or-..." autocomplete="off"></div>
   <div class="rwa-set-row" id="rwa-set-row-bridge-token" style="display:none;"><label>Bridge Token</label><input type="password" id="rwa-bridge-token" placeholder="WEB_CLI_BRIDGE_TOKEN" autocomplete="off"></div>
   <div class="rwa-set-row" id="rwa-set-row-bridge-cwd" style="display:none;"><label>Session Dir</label><input type="text" id="rwa-bridge-cwd" placeholder="/path/on/bridge/host" autocomplete="off" spellcheck="false"></div>
@@ -1198,6 +1205,12 @@ function buildUI() {
       defaultUrl: RWA.DEFAULT_LMSTUDIO_URL,
       storageKey: RWA.K_BASE_URL_LMSTUDIO,
     },
+    atomic: {
+      showKey: false, showBaseUrl: true, showModel: true, showHint: true,
+      hintHTML: 'atomic.chat serves an OpenAI-compatible API on <code>127.0.0.1:1337</code>, no key needed — use <code>Test</code> to list its models. CORS allows http(s) page origins out of the box but <strong>not <code>file://</code> pages</strong> (the null origin): open this container from a local web server or a hosted projection to use it.',
+      defaultUrl: RWA.DEFAULT_ATOMIC_URL,
+      storageKey: RWA.K_BASE_URL_ATOMIC,
+    },
     bridge: {
       showKey: false, showBaseUrl: false, showModel: false, showHint: false,
       hintHTML: '',
@@ -4321,6 +4334,18 @@ function resolveBackendConfig() {
       apiKey: null, extraHeaders: {}, requiresKey: false,
     };
   }
+  if (backend === 'atomic') {
+    return {
+      backend, kind:'openai_compat',
+      baseUrl: (sessionStorage.getItem(RWA.K_BASE_URL_ATOMIC) || '').trim() || RWA.DEFAULT_ATOMIC_URL,
+      apiKey: null, extraHeaders: {}, requiresKey: false,
+      // atomic.chat REJECTS (400) requests whose prompt + max generation
+      // exceed its MAX_KV_SIZE (16384 default) rather than clamping like
+      // ollama/lmstudio — half the window for generation, half for the
+      // prompt + document. Callers read cfg.maxTokens || 32000.
+      maxTokens: 8192,
+    };
+  }
   if (backend === 'bridge') {
     return { backend, kind:'bridge' };
   }
@@ -4379,7 +4404,7 @@ async function callAgentSingleShot(prompt) {
   const model = sessionStorage.getItem(RWA.K_MODEL) || RWA.MODEL;
   const data = await openAiCompatChat(cfg, {
     model,
-    max_tokens: 32000,
+    max_tokens: cfg.maxTokens || 32000,
     messages: [{ role:'user', content: prompt }],
   });
   const msg = data.choices?.[0]?.message;
@@ -6578,7 +6603,7 @@ async function modify(instr, lensMeta = null, opts = null) {
       try {
         data = await openAiCompatChat(cfg, {
           model,
-          max_tokens: 32000,
+          max_tokens: cfg.maxTokens || 32000,
           messages,
           tools: TOOL_SCHEMAS,
           tool_choice: 'auto',

package/src/agent-loop.mjs

@@ -128,12 +128,14 @@ export async function runAgentLoop({
   throw new AgentError('no_envelope_after_retries', { retries: RETRY_BUDGET });
 }
 
-async function callBackend({ baseUrl, model, apiKey }, body) {
+async function callBackend({ baseUrl, model, apiKey, maxTokens }, body) {
   const headers = { 'Content-Type': 'application/json' };
   if (apiKey) headers['Authorization'] = `Bearer ${apiKey}`;
   const url = baseUrl.replace(/\/+$/, '') + '/chat/completions';
   // Seed parity (seeds/rewritable.html openAiCompatChat caller in modify()):
-  // every request carries max_tokens: 32000 and tool_choice: 'auto'. The
+  // every request carries the backend's max_tokens (32000 historically; 8192
+  // for atomic, whose server REJECTS prompt+generation past MAX_KV_SIZE rather
+  // than clamping — see backendMaxTokens) and tool_choice: 'auto'. The
   // tool_choice default forces the model to call one of the provided tools
   // rather than emitting plain text (which would trip our no_tool_call retry).
   const res = await fetch(url, {
@@ -141,7 +143,7 @@ async function callBackend({ baseUrl, model, apiKey }, body) {
     headers,
     body: JSON.stringify({
       model,
-      max_tokens: 32000,
+      max_tokens: maxTokens || 32000,
       tool_choice: 'auto',
       ...body,
     }),

package/src/backend.mjs

@@ -2,15 +2,15 @@
 // precedence is unit-testable (the bin entrypoint runs on import and can't be
 // imported cleanly).
 //
-// Only the openrouter backend needs a key — ollama and lmstudio run locally
-// without auth. The key resolves in order: an explicit --api-key flag, then the
+// Only the openrouter backend needs a key — ollama, lmstudio, and atomic run
+// locally without auth. The key resolves in order: an explicit --api-key flag, then the
 // project-specific RWA_OPENROUTER_KEY (env conventions match the docker-compose
 // deploy in service/), then the CONVENTIONAL OPENROUTER_API_KEY that agents and
 // users normally have exported. Empty strings count as absent.
 
 /**
  * Resolve the API key for a backend.
- * @param {string} backendName — 'openrouter' | 'ollama' | 'lmstudio'
+ * @param {string} backendName — 'openrouter' | 'ollama' | 'lmstudio' | 'atomic'
  * @param {string|undefined} flagValue — the --api-key flag value, if any
  * @param {Record<string,string|undefined>} [env] — environment (injectable for tests)
  * @returns {string|undefined} the key, or undefined when none applies
@@ -26,10 +26,10 @@ export function resolveApiKey(backendName, flagValue, env = process.env) {
 /**
  * Default OpenAI-compatible base URL for a backend — mirrors the inline
  * `envBaseUrl` in bin/rwa.mjs (and seeds/rewritable.html resolveBackendConfig).
- * ollama and lmstudio honor RWA_*_URL overrides (remote host / non-standard port);
- * openrouter is fixed (the URL has never drifted in the seed). Shared by `rwa edit`
+ * ollama, lmstudio, and atomic honor RWA_*_URL overrides (remote host / non-standard
+ * port); openrouter is fixed (the URL has never drifted in the seed). Shared by `rwa edit`
  * and `rwa create` so the default never diverges between the two.
- * @param {string} name — 'openrouter' | 'ollama' | 'lmstudio'
+ * @param {string} name — 'openrouter' | 'ollama' | 'lmstudio' | 'atomic'
  * @param {Record<string,string|undefined>} [env] — environment (injectable for tests)
  * @returns {string|undefined}
  */
@@ -38,6 +38,24 @@ export function envBaseUrl(name, env = process.env) {
     case 'openrouter': return 'https://openrouter.ai/api/v1';
     case 'ollama':     return env.RWA_OLLAMA_URL || 'http://localhost:11434/v1';
     case 'lmstudio':   return env.RWA_LMSTUDIO_URL || 'http://localhost:1234/v1';
+    case 'atomic':     return env.RWA_ATOMIC_URL || 'http://127.0.0.1:1337/v1';
     default:           return undefined;
   }
 }
+
+/**
+ * Per-backend max_tokens for the agent loop. The historical 32000 stands for
+ * hosted/clamping backends, but atomic.chat REJECTS (400) any request whose
+ * prompt + max generation exceeds its MAX_KV_SIZE (16384 by default) rather
+ * than clamping — so it gets 8192, leaving the other half of the window for
+ * the system prompt + document. RWA_MAX_TOKENS overrides for unusual servers.
+ * Mirrors the seed's resolveBackendConfig() maxTokens.
+ * @param {string} name — backend name
+ * @param {Record<string,string|undefined>} [env]
+ * @returns {number}
+ */
+export function backendMaxTokens(name, env = process.env) {
+  const override = Number(env.RWA_MAX_TOKENS);
+  if (Number.isFinite(override) && override > 0) return override;
+  return name === 'atomic' ? 8192 : 32000;
+}

package/src/commands.mjs

@@ -85,7 +85,7 @@ async function readEnvKey(name) {
 // than throwing — pre-fill is best-effort; an unknown value just means the
 // user sees the default backend (openrouter) on first paint.
 function validBackend(v) {
-  return ['openrouter', 'ollama', 'lmstudio', 'bridge'].includes(v) ? v : null;
+  return ['openrouter', 'ollama', 'lmstudio', 'atomic', 'bridge'].includes(v) ? v : null;
 }
 
 // Collect URL-param pre-fills from env / ./.env. Returns an object whose keys

package/src/create.mjs

@@ -20,7 +20,7 @@ import { runAgentLoop } from './agent-loop.mjs';
 import { applyPlan, CliError } from './edit.mjs';
 import { assertSelfContained } from './self-contained.mjs';
 import { findFrozenZones } from './apply-edits.mjs';
-import { resolveApiKey, envBaseUrl } from './backend.mjs';
+import { resolveApiKey, envBaseUrl, backendMaxTokens } from './backend.mjs';
 import { atomicWrite } from './atomic-write.mjs';
 
 // Hard cap on --data baked into the snapshot. The dataset lands inside INLINE_DOC
@@ -219,6 +219,7 @@ export async function createCmd(parsed, { seedCandidates, cwd = process.cwd(), s
     baseUrl: parsed.backend.baseUrl || envBaseUrl(backendName),
     model:   parsed.backend.model   || process.env.RWA_MODEL || 'google/gemini-3.5-flash',
     apiKey:  resolveApiKey(backendName, parsed.backend.apiKey),
+    maxTokens: backendMaxTokens(backendName),
   };
 
   // Per-kind system prompt + the create-only self-containment directive; the brief