@link-assistant/agent 0.26.15 → 0.27.1

package/src/cli/argv.ts

@@ -105,7 +105,7 @@ export function getCompactionSafetyMarginFromProcessArgv(): string | null {
 /**
  * Extract --compaction-models argument directly from process.argv
  * The value is a links notation references sequence, e.g.:
- *   "(kilo/nemotron-3-super-free kilo/nemotron-3-ultra-free kilo/nemotron-3.5-lightning-free same)"
+ *   "(kilo/nemotron-3-ultra-free kilo/nemotron-3.5-lightning-free same)"
  * @returns The compaction models argument from CLI or null if not found
  * @see https://github.com/link-assistant/agent/issues/232
  */

package/src/cli/cmd/auth.ts

@@ -359,6 +359,10 @@ export const AuthLoginCommand = cmd({
       togetherai: 'https://api.together.xyz/settings/api-keys',
       deepseek: 'https://platform.deepseek.com/api_keys',
       xai: 'https://console.x.ai',
+      poe: 'https://poe.com/api_key',
+      digitalocean: 'https://cloud.digitalocean.com/gen-ai/model-access-keys',
+      gitlab: 'https://gitlab.com/-/user_settings/personal_access_tokens',
+      azure: 'https://ai.azure.com/',
     };
 
     if (apiKeyUrls[provider]) {

package/src/cli/cmd/models.ts

@@ -1,10 +1,17 @@
 import type { Argv } from 'yargs';
 import { Instance } from '../../project/instance';
 import { Provider } from '../../provider/provider';
+import { ModelsDev } from '../../provider/models';
+import { ModelDiscovery } from '../../provider/model-discovery';
+import { ExtraProviders } from '../../provider/extra-providers';
 import { cmd } from './cmd';
 import { UI } from '../ui';
 import { EOL } from 'os';
 
+function isFree(model: ModelsDev.Model) {
+  return model.cost?.input === 0 && model.cost?.output === 0;
+}
+
 export const ModelsCommand = cmd({
   command: 'models [provider]',
   describe: 'list all available models',
@@ -19,39 +26,87 @@ export const ModelsCommand = cmd({
         describe:
           'use more verbose model output (includes metadata like costs)',
         type: 'boolean',
+      })
+      .option('discover', {
+        describe:
+          "also list models from each provider's own /models endpoint (cached for an hour)",
+        type: 'boolean',
+      })
+      .option('refresh', {
+        describe:
+          'refetch the models.dev catalogue and provider model listings, ignoring caches (implies --discover)',
+        type: 'boolean',
+      })
+      .option('free', {
+        describe: 'only list models that cost nothing to use',
+        type: 'boolean',
       });
   },
   handler: async (args) => {
+    if (args.refresh) await ModelsDev.refresh();
+    const discover = Boolean(args.discover || args.refresh);
+
     await Instance.provide({
       directory: process.cwd(),
       async fn() {
-        const providers = await Provider.list();
-
-        function printModels(providerID: string, verbose?: boolean) {
-          const provider = providers[providerID];
-          for (const [modelID, model] of Object.entries(provider.info.models)) {
-            process.stdout.write(`${providerID}/${modelID}`);
+        async function printModels(providerID: string) {
+          const provider = await Provider.getProvider(providerID);
+          if (!provider) return false;
+          const models: Record<string, ModelsDev.Model> = {
+            ...provider.info.models,
+          };
+          const unpriced = new Set<string>();
+          const keyless = ExtraProviders.isKeyless(provider.info.id);
+          if (discover) {
+            const result = await Provider.discoverModels(providerID, {
+              force: Boolean(args.refresh),
+            });
+            if (result.status === 'error') {
+              UI.error(
+                `${providerID}: model discovery failed: ${result.error}`
+              );
+            }
+            const s = await Provider.state();
+            for (const found of result.models) {
+              if (models[found.id]) continue;
+              const metadata = ModelDiscovery.catalogueMetadata(
+                s.database,
+                found.model ?? found.id,
+                provider.info.id
+              );
+              // A listing without prices says nothing about the cost, except
+              // on keyless providers (gpt4free, local servers), which are free.
+              if (!keyless && !found.free && !found.cost && !metadata?.cost) {
+                unpriced.add(found.id);
+              }
+              models[found.id] = ModelDiscovery.toModelInfo(found, {
+                metadata,
+              });
+            }
+          }
+          for (const [modelID, model] of Object.entries(models)) {
+            if (args.free && (unpriced.has(modelID) || !isFree(model)))
+              continue;
+            process.stdout.write(`${provider.info.id}/${modelID}`);
             process.stdout.write(EOL);
-            if (verbose) {
+            if (args.verbose) {
               process.stdout.write(JSON.stringify(model, null, 2));
               process.stdout.write(EOL);
             }
           }
+          return true;
         }
 
         if (args.provider) {
-          const provider = providers[args.provider];
-          if (!provider) {
+          if (!(await printModels(args.provider))) {
             UI.error(`Provider not found: ${args.provider}`);
-            return;
           }
-
-          printModels(args.provider, args.verbose);
           return;
         }
 
+        const providers = await Provider.list();
         for (const providerID of Object.keys(providers)) {
-          printModels(providerID, args.verbose);
+          await printModels(providerID);
         }
       },
     });

package/src/cli/cmd/run.ts

@@ -9,6 +9,7 @@ import { EOL } from 'os';
 import { select } from '@clack/prompts';
 import { createOpencodeClient, type OpencodeClient } from '@opencode-ai/sdk';
 import { Provider } from '../../provider/provider';
+import { legacyErrorObject } from '../../util/legacy-error-names';
 
 const TOOL: Record<string, [string, string]> = {
   todowrite: ['Todo', UI.Style.TEXT_WARNING_BOLD],
@@ -267,7 +268,12 @@ export const RunCommand = cmd({
             // Always include the human-readable string alongside the
             // machine-readable object so consumers never print
             // "[object Object]" (#289).
-            if (outputJsonEvent('error', { message: err, error: props.error }))
+            if (
+              outputJsonEvent('error', {
+                message: err,
+                error: legacyErrorObject(props.error),
+              })
+            )
               continue;
             UI.error(err);
           }

package/src/cli/continuous-mode.js

@@ -356,6 +356,10 @@ export async function runContinuousServerMode(
       const parts = [{ type: 'text', text: messageText }];
 
       // Create a promise to wait for this message to complete
+      // `stopIdleWait` ends the wait when the prompt settles without ever
+      // going idle (it can throw before the loop starts); otherwise the
+      // subscription leaked and the input queue hung (issue #335).
+      let stopIdleWait = () => {};
       const messagePromise = new Promise((resolve) => {
         const checkIdle = Bus.subscribeAll((event) => {
           if (
@@ -366,6 +370,10 @@ export async function runContinuousServerMode(
             resolve();
           }
         });
+        stopIdleWait = () => {
+          checkIdle();
+          resolve();
+        };
       });
 
       // Send message
@@ -383,21 +391,28 @@ export async function runContinuousServerMode(
             temperature,
           }),
         }
-      ).catch((error) => {
-        hasError = true;
-        eventHandler.output({
-          type: 'error',
-          timestamp: Date.now(),
-          sessionID,
-          message: errorText(error),
-          error: error instanceof Error ? error.message : String(error),
+      )
+        // Read the reply to the end: it completes only after the prompt has
+        // returned (and released the connection), successful or not.
+        .then((response) => response.text())
+        .catch((error) => {
+          hasError = true;
+          eventHandler.output({
+            type: 'error',
+            timestamp: Date.now(),
+            sessionID,
+            message: errorText(error),
+            error: error instanceof Error ? error.message : String(error),
+          });
         });
-      });
 
       // See the direct-mode comment below: the request rejection must be
-      // observed before the process can exit (#290).
-      await messagePromise;
+      // observed before the process can exit (#290). `session.idle` is
+      // published before the reply is written, so once the reply is complete
+      // the idle wait can stop.
       await requestPromise;
+      stopIdleWait();
+      await messagePromise;
       isProcessing = false;
 
       // Process next pending message if any
@@ -582,6 +597,10 @@ export async function runContinuousDirectMode(
       const parts = [{ type: 'text', text: messageText }];
 
       // Create a promise to wait for this message to complete
+      // `stopIdleWait` ends the wait when the prompt settles without ever
+      // going idle (it can throw before the loop starts); otherwise the
+      // subscription leaked and the input queue hung (issue #335).
+      let stopIdleWait = () => {};
       const messagePromise = new Promise((resolve) => {
         const checkIdle = Bus.subscribeAll((event) => {
           if (
@@ -592,6 +611,10 @@ export async function runContinuousDirectMode(
             resolve();
           }
         });
+        stopIdleWait = () => {
+          checkIdle();
+          resolve();
+        };
       });
 
       // Send message directly
@@ -618,8 +641,11 @@ export async function runContinuousDirectMode(
       // publishes `session.idle` while unwinding, so waiting only on the idle
       // event let the process exit before the rejection handler above ran —
       // no error record, exit code 0 (#290).
-      await messagePromise;
+      // `session.idle` is published synchronously before `prompt` settles,
+      // so after it settles the idle wait can stop.
       await promptPromise;
+      stopIdleWait();
+      await messagePromise;
       isProcessing = false;
 
       // Process next pending message if any

package/src/cli/event-handler.js

@@ -5,6 +5,7 @@
 
 import { Bus } from '../bus/index.ts';
 import { errorText } from '../util/error-text.ts';
+import { legacyErrorObject } from '../util/legacy-error-names.ts';
 
 /**
  * Create a subscription for session bus events that outputs events in the selected JSON format.
@@ -142,7 +143,8 @@ export function outputBusEvent({
       timestamp: Date.now(),
       sessionID,
       message: errorText(props.error),
-      error: props.error,
+      // --legacy-error-names maps new error names to the previous ones.
+      error: legacyErrorObject(props.error),
     });
   }
 }

package/src/cli/model-config.js

@@ -5,6 +5,8 @@ import {
   getCompactionSafetyMarginFromProcessArgv,
 } from './argv.ts';
 import { Log } from '../util/log.ts';
+import { ExtraProviders } from '../provider/extra-providers.ts';
+import { resolveKiloDeprecatedAlias } from '../provider/kilo-deprecation.ts';
 import {
   DEFAULT_PROVIDER_ID,
   DEFAULT_MODEL_ID,
@@ -30,6 +32,17 @@ export class ModelResolutionError extends Error {
   }
 }
 
+/**
+ * Map a withdrawn built-in Kilo alias to its replacement, with a warning (#335).
+ * Other providers and model IDs are returned unchanged.
+ * @param {string} providerID
+ * @param {string} modelID
+ * @returns {string}
+ */
+function replaceDeprecatedModel(providerID, modelID) {
+  return providerID === 'kilo' ? resolveKiloDeprecatedAlias(modelID) : modelID;
+}
+
 /**
  * Parse model config from argv. Supports "provider/model" or short "model" format.
  * @param {object} argv - Parsed command line arguments
@@ -124,8 +137,9 @@ export async function parseModelConfig(
   if (modelArg.includes('/')) {
     // Explicit provider/model format - respect user's choice
     const modelParts = modelArg.split('/');
-    providerID = modelParts[0];
-    modelID = modelParts.slice(1).join('/');
+    // Accept alternative spellings such as `g4f/...` for `gpt4free/...` (#334)
+    providerID = ExtraProviders.canonicalID(modelParts[0]);
+    modelID = replaceDeprecatedModel(providerID, modelParts.slice(1).join('/'));
 
     // Validate that providerID and modelID are not empty
     // Do NOT fall back to defaults - if the user provided an invalid format, fail clearly (#196)
@@ -152,9 +166,10 @@ export async function parseModelConfig(
     const isDefaultModel = !cliModelArg && modelArg === defaultModel;
     try {
       const { Provider } = await import('../provider/provider.ts');
-      const s = await Provider.state();
-      const provider = s.providers[providerID];
-      if (provider && !provider.info.models[modelID]) {
+      // getProvider also enables keyless providers (gpt4free, local servers)
+      // on explicit request; routers among them accept unlisted IDs (#334).
+      const provider = await Provider.getProvider(providerID);
+      if (provider && !(await Provider.hasModel(providerID, modelID))) {
         const liveInfo = await Provider.refreshLiveModelInfo(
           providerID,
           modelID
@@ -344,7 +359,7 @@ async function resolveCompactionModelEntry(
     const parts = modelArg.split('/');
     return {
       providerID: parts[0],
-      modelID: parts.slice(1).join('/'),
+      modelID: replaceDeprecatedModel(parts[0], parts.slice(1).join('/')),
       useSameModel: false,
     };
   }
@@ -396,7 +411,7 @@ async function resolveEntryWithinProvider(
     }
     return {
       providerID: parts[0],
-      modelID: parts.slice(1).join('/'),
+      modelID: replaceDeprecatedModel(parts[0], parts.slice(1).join('/')),
       useSameModel: false,
     };
   }
@@ -405,7 +420,11 @@ async function resolveEntryWithinProvider(
   if (!(await Provider.hasModel(baseProviderID, modelArg))) {
     return null;
   }
-  return { providerID: baseProviderID, modelID: modelArg, useSameModel: false };
+  return {
+    providerID: baseProviderID,
+    modelID: replaceDeprecatedModel(baseProviderID, modelArg),
+    useSameModel: false,
+  };
 }
 
 /**

package/src/cli/run-options.js

@@ -158,82 +158,148 @@ export function buildRunOptions(yargs, defaultOptions = {}) {
         }, true)
       : parser;
 
-  return normalizedParser
-    .option('generate-title', {
-      type: 'boolean',
-      description:
-        'Generate session titles using AI (default: false). Disabling saves tokens and prevents rate limit issues.',
-      default: false,
-    })
-    .option('retry-timeout', {
-      type: 'number',
-      description:
-        'Maximum total retry time in seconds for rate limit errors (default: 604800 = 7 days)',
-    })
-    .option('retry-on-rate-limits', {
-      type: 'boolean',
-      description:
-        'Retry AI completions API requests when rate limited (HTTP 429). Use --no-retry-on-rate-limits in integration tests to fail fast instead of waiting.',
-      default: true,
-    })
-    .option('output-response-model', {
-      type: 'boolean',
-      description: 'Include model info in step_finish output',
-      default: true,
-    })
-    .option('summarize-session', {
-      type: 'boolean',
-      description:
-        'Generate AI session summaries (default: true). Use --no-summarize-session to disable. Automatically off for a model declared as having no context limit (`unlimited: true` or `limit.context: null` in the provider config).',
-      default: true,
-    })
-    .option('compaction-model', {
-      type: 'string',
-      description:
-        'Model to use for context compaction and session summaries, in format providerID/modelID. Use "same" to use the base model. Default: kilo/nemotron-3-super-free (free, Kilo authentication recommended, 262K context), which falls back to the --model model when --model points at another provider. Overridden by --compaction-models if both are specified.',
-      default: defaultCompactionModel,
-    })
-    .option('compaction-models', {
-      type: 'string',
-      description:
-        'Ordered cascade of compaction models in links notation sequence format: "(model1 model2 ... same)". ' +
-        "Models are tried from smallest/cheapest context to largest. If used context exceeds a model's limit or its rate limit is reached, the next model is tried. " +
-        'The special value "same" uses the base model. Overrides --compaction-model when specified. ' +
-        'The default cascade is OpenCode-specific: when --model points at another provider, its entries from other providers are dropped so secondary calls inherit --model. A cascade given explicitly is used as written.',
-      default: defaultCompactionModels,
-    })
-    .option('compaction-safety-margin', {
-      type: 'number',
-      description:
-        'Safety margin (%) of usable context window before triggering compaction. Only applies when the compaction model has equal or smaller context than the base model. Default: 25.',
-      default: defaultCompactionSafetyMarginPercent,
-    })
-    .option('temperature', {
-      type: 'number',
-      description:
-        'Override the temperature for model completions. When not set, the default per-model temperature is used.',
-    })
-    .option('read-only', {
-      type: 'boolean',
-      description:
-        'Enforceable read-only / planning mode. Disables all filesystem-mutating and shell tools (bash, edit, write, multiedit, patch) so the agent can only read, search and plan. Maps to agent-commander --read-only.',
-      default: false,
-    })
-    .option('disable-tools', {
-      type: 'string',
-      description:
-        'Comma-separated list of tool ids to disable (e.g. "bash,write,edit"). Disabled tools are never exposed to the model and are rejected if invoked via batch.',
-    })
-    .option('permission-mode', {
-      type: 'string',
-      choices: ['auto', 'plan', 'readonly', 'ask'],
-      description:
-        'Permission policy for mutating tools. "auto" (default) allows everything; "plan" allows read-only shell and asks before mutations; "readonly" denies all mutations; "ask" asks before every mutating tool. Approvals are driven over JSON (permission_request / permission_response).',
-      default: 'auto',
-    })
-    .option('permission', {
-      type: 'string',
-      description:
-        'OpenCode-compatible permission override as JSON, merged on top of --permission-mode. Example: \'{"edit":"ask","bash":{"git push*":"ask","*":"allow"},"webfetch":"allow"}\'.',
-    });
+  return (
+    normalizedParser
+      .option('generate-title', {
+        type: 'boolean',
+        description:
+          'Generate session titles using AI (default: false). Disabling saves tokens and prevents rate limit issues.',
+        default: false,
+      })
+      .option('retry-timeout', {
+        type: 'number',
+        description:
+          'Maximum total retry time in seconds for rate limit errors (default: 604800 = 7 days)',
+      })
+      .option('retry-on-rate-limits', {
+        type: 'boolean',
+        description:
+          'Retry AI completions API requests when rate limited (HTTP 429). Use --no-retry-on-rate-limits in integration tests to fail fast instead of waiting.',
+        default: true,
+      })
+      .option('output-response-model', {
+        type: 'boolean',
+        description: 'Include model info in step_finish output',
+        default: true,
+      })
+      .option('summarize-session', {
+        type: 'boolean',
+        description:
+          'Generate AI session summaries (default: true). Use --no-summarize-session to disable. Automatically off for a model declared as having no context limit (`unlimited: true` or `limit.context: null` in the provider config).',
+        default: true,
+      })
+      .option('compaction-model', {
+        type: 'string',
+        description:
+          'Model to use for context compaction and session summaries, in format providerID/modelID. Use "same" to use the base model. Default: kilo/nemotron-3-ultra-free (free, Kilo authentication recommended, 1M context), which falls back to the --model model when --model points at another provider. Overridden by --compaction-models if both are specified.',
+        default: defaultCompactionModel,
+      })
+      .option('compaction-models', {
+        type: 'string',
+        description:
+          'Ordered cascade of compaction models in links notation sequence format: "(model1 model2 ... same)". ' +
+          "Models are tried from smallest/cheapest context to largest. If used context exceeds a model's limit or its rate limit is reached, the next model is tried. " +
+          'The special value "same" uses the base model. Overrides --compaction-model when specified. ' +
+          'The default cascade is Kilo-specific: when --model points at another provider, its entries from other providers are dropped so secondary calls inherit --model. A cascade given explicitly is used as written.',
+        default: defaultCompactionModels,
+      })
+      .option('compaction-safety-margin', {
+        type: 'number',
+        description:
+          'Safety margin (%) of usable context window before triggering compaction. Only applies when the compaction model has equal or smaller context than the base model. Default: 25.',
+        default: defaultCompactionSafetyMarginPercent,
+      })
+      .option('temperature', {
+        type: 'number',
+        description:
+          'Override the temperature for model completions. When not set, the default per-model temperature is used.',
+      })
+      .option('read-only', {
+        type: 'boolean',
+        description:
+          'Enforceable read-only / planning mode. Disables all filesystem-mutating and shell tools (bash, edit, write, multiedit, patch) so the agent can only read, search and plan. Maps to agent-commander --read-only.',
+        default: false,
+      })
+      .option('disable-tools', {
+        type: 'string',
+        description:
+          'Comma-separated list of tool ids to disable (e.g. "bash,write,edit"). Disabled tools are never exposed to the model and are rejected if invoked via batch.',
+      })
+      .option('permission-mode', {
+        type: 'string',
+        choices: ['auto', 'plan', 'readonly', 'ask'],
+        description:
+          'Permission policy for mutating tools. "auto" (default) allows everything; "plan" allows read-only shell and asks before mutations; "readonly" denies all mutations; "ask" asks before every mutating tool. Approvals are driven over JSON (permission_request / permission_response).',
+        default: 'auto',
+      })
+      .option('permission', {
+        type: 'string',
+        description:
+          'OpenCode-compatible permission override as JSON, merged on top of --permission-mode. Example: \'{"edit":"ask","bash":{"git push*":"ask","*":"allow"},"webfetch":"allow"}\'.',
+      })
+      // Compatibility switches (issue #335). They are resolved (together with
+      // their LINK_ASSISTANT_AGENT_* env vars) by initConfig() in
+      // src/config/config.ts; they are declared here so they show up in --help.
+      // No defaults here, so an env var is not shadowed by a yargs default.
+      .option('bash-shell', {
+        type: 'string',
+        description:
+          'Shell for the bash tool: "auto" (default, $SHELL if POSIX, else bash, else /bin/sh), "legacy" (/bin/sh as before) or a shell path. Env: LINK_ASSISTANT_AGENT_BASH_SHELL',
+      })
+      .option('legacy-tool-output', {
+        type: 'boolean',
+        description:
+          'Restore previous tool output formats (empty edit/write output, head-truncated bash output, empty read past EOF). Env: LINK_ASSISTANT_AGENT_LEGACY_TOOL_OUTPUT',
+      })
+      .option('legacy-edit-matching', {
+        type: 'boolean',
+        description:
+          'Restore previous edit matching (empty oldString overwrites a file, loose fuzzy thresholds). Env: LINK_ASSISTANT_AGENT_LEGACY_EDIT_MATCHING',
+      })
+      .option('glob-follow-symlinks', {
+        type: 'boolean',
+        description:
+          'Follow symlinked directories in glob and ls (previous behaviour). Env: LINK_ASSISTANT_AGENT_GLOB_FOLLOW_SYMLINKS',
+      })
+      .option('grep-hidden', {
+        type: 'boolean',
+        description:
+          'Search hidden files in grep (default true; --no-grep-hidden restores previous behaviour). Env: LINK_ASSISTANT_AGENT_GREP_HIDDEN',
+      })
+      .option('grep-timeout-ms', {
+        type: 'number',
+        description:
+          'Timeout of one grep search in ms (default 30000, 0 = none). Env: LINK_ASSISTANT_AGENT_GREP_TIMEOUT_MS',
+      })
+      .option('read-max-lines', {
+        type: 'number',
+        description:
+          'Max lines one read call returns (default 2000, 0 = no cap). Env: LINK_ASSISTANT_AGENT_READ_MAX_LINES',
+      })
+      .option('read-max-bytes', {
+        type: 'number',
+        description:
+          'Max bytes one read call returns (default 51200, 0 = no cap). Env: LINK_ASSISTANT_AGENT_READ_MAX_BYTES',
+      })
+      .option('legacy-error-names', {
+        type: 'boolean',
+        description:
+          'Report ContextOverflowError as APIError in JSON output (previous name). Env: LINK_ASSISTANT_AGENT_LEGACY_ERROR_NAMES',
+      })
+      .option('legacy-token-usage', {
+        type: 'boolean',
+        description:
+          'Report tokens.output including reasoning tokens (previous accounting). Env: LINK_ASSISTANT_AGENT_LEGACY_TOKEN_USAGE',
+      })
+      .option('legacy-loop-exit', {
+        type: 'boolean',
+        description:
+          'End the turn on a "stop" finish even if tools ran in that step (previous behaviour). Env: LINK_ASSISTANT_AGENT_LEGACY_LOOP_EXIT',
+      })
+      .option('mcp-connect-timeout', {
+        type: 'number',
+        description:
+          'Default MCP connect/initialize timeout in ms (default 30000, 0 = none). Env: LINK_ASSISTANT_AGENT_MCP_CONNECT_TIMEOUT',
+      })
+  );
 }