clementine-agent 1.18.13 → 1.18.15

package/README.md

@@ -328,8 +328,9 @@ For spend/context tuning, `clementine budgets` gives a safer shortcut:
 
 ```bash
 clementine budgets              # show chat/cron/heartbeat caps and 1M context state
-clementine budgets safe         # lower background budgets and disable Claude 1M context
-clementine budgets 1m on        # enable 1M context for eligible accounts / Extra Usage
+clementine budgets safe         # lower background budgets and force standard 200K context
+clementine budgets 1m auto      # allow included Opus 1M, keep Sonnet on 200K
+clementine budgets 1m on        # force 1M context for Extra Usage/API users
 clementine budgets 1m off       # disable 1M context for maximum compatibility
 clementine budgets set chat 10  # raise one budget cap
 ```
@@ -342,7 +343,8 @@ clementine budgets set chat 10  # raise one budget cap
 | `BUDGET_CRON_T1_USD` | `0.75` | Max spend per tier-1 cron job |
 | `BUDGET_CRON_T2_USD` | `1.50` | Max spend per tier-2 cron job |
 | `BUDGET_HEARTBEAT_USD` | `0.25` | Max spend per heartbeat tick |
-| `CLAUDE_CODE_DISABLE_1M_CONTEXT` | `true` | `true`/`1` keeps Claude Code on 200K context unless the user explicitly enables 1M |
+| `CLEMENTINE_1M_CONTEXT_MODE` | `auto` | `auto` allows included Opus 1M on Max/Team/Enterprise while keeping Sonnet on 200K; `off` forces 200K; `on` forces 1M |
+| `CLAUDE_CODE_DISABLE_1M_CONTEXT` | legacy | Backward-compatible Claude Code switch; `budgets safe` writes `1`, `budgets 1m auto` removes it |
 | `DEFAULT_MODEL_TIER` | `sonnet` | Default model: `haiku` / `sonnet` / `opus` |
 | `HEARTBEAT_INTERVAL_MINUTES` | `30` | How often the agent auto-checks in |
 | `HEARTBEAT_ACTIVE_START` | `8` | First hour of the active window (0–23) |

package/dist/agent/assistant.js

@@ -13,7 +13,7 @@ import fs from 'node:fs';
 import path from 'node:path';
 import { query as rawQuery, listSubagents, getSubagentMessages, SYSTEM_PROMPT_DYNAMIC_BOUNDARY, } from '@anthropic-ai/claude-agent-sdk';
 import pino from 'pino';
-import { BASE_DIR, PKG_DIR, VAULT_DIR, DAILY_NOTES_DIR, SOUL_FILE, AGENTS_FILE, MEMORY_FILE, AGENTS_DIR, ASSISTANT_NAME, OWNER_NAME, MODEL, MODELS, HEARTBEAT_MAX_TURNS, SEARCH_CONTEXT_LIMIT, SEARCH_RECENCY_LIMIT, SYSTEM_PROMPT_MAX_CONTEXT_CHARS, SESSION_EXCHANGE_HISTORY_SIZE, SESSION_EXCHANGE_MAX_CHARS, INJECTED_CONTEXT_MAX_CHARS, UNLEASHED_PHASE_TURNS, UNLEASHED_DEFAULT_MAX_HOURS, UNLEASHED_MAX_PHASES, PROJECTS_META_FILE, CRON_PROGRESS_DIR, CRON_REFLECTIONS_DIR, HANDOFFS_DIR, BUDGET, TASK_BUDGET_TOKENS, IDENTITY_FILE, CLAUDE_CODE_OAUTH_TOKEN, ANTHROPIC_API_KEY as CONFIG_ANTHROPIC_API_KEY, envSnapshot, } from '../config.js';
+import { BASE_DIR, PKG_DIR, VAULT_DIR, DAILY_NOTES_DIR, SOUL_FILE, AGENTS_FILE, MEMORY_FILE, AGENTS_DIR, ASSISTANT_NAME, OWNER_NAME, MODEL, MODELS, HEARTBEAT_MAX_TURNS, SEARCH_CONTEXT_LIMIT, SEARCH_RECENCY_LIMIT, SYSTEM_PROMPT_MAX_CONTEXT_CHARS, SESSION_EXCHANGE_HISTORY_SIZE, SESSION_EXCHANGE_MAX_CHARS, INJECTED_CONTEXT_MAX_CHARS, UNLEASHED_PHASE_TURNS, UNLEASHED_DEFAULT_MAX_HOURS, UNLEASHED_MAX_PHASES, PROJECTS_META_FILE, CRON_PROGRESS_DIR, CRON_REFLECTIONS_DIR, HANDOFFS_DIR, BUDGET, TASK_BUDGET_TOKENS, IDENTITY_FILE, CLAUDE_CODE_OAUTH_TOKEN, ANTHROPIC_API_KEY as CONFIG_ANTHROPIC_API_KEY, claudeCodeDisableOneMillionForModel, currentOneMillionContextMode, normalizeClaudeModelForOneMillionContext, usesOneMillionContext, envSnapshot, } from '../config.js';
 import { summarizeIntegrationStatus } from '../config/integrations-registry.js';
 import { loadToolPreferences, computeAvailability, buildPromptInstruction, buildComposioStatusBlock, } from '../integrations/tool-preferences.js';
 import { loadClaudeIntegrations } from './mcp-bridge.js';
@@ -320,6 +320,8 @@ const MODEL_CONTEXT_WINDOWS = {
     'opus': 200_000,
 };
 function getContextWindow(model) {
+    if (usesOneMillionContext(model))
+        return 1_000_000;
     for (const [family, size] of Object.entries(MODEL_CONTEXT_WINDOWS)) {
         if (model.includes(family))
             return size;
@@ -338,10 +340,6 @@ function resultInputTokens(result) {
     }
     return total;
 }
-function oneMillionContextDisabled() {
-    const value = process.env.CLAUDE_CODE_DISABLE_1M_CONTEXT;
-    return value === undefined || !/^(0|false|no)$/i.test(value);
-}
 export function looksLikeOneMillionContextError(value) {
     const text = String(value ?? '');
     return /extra usage.*1m context|1m context.*extra usage|context-1m/i.test(text);
@@ -537,12 +535,6 @@ function buildSafeEnv() {
         sanitized.ANTHROPIC_API_KEY = apiKeyVal;
     }
     // When all are absent: HOME lets the subprocess find Keychain OAuth automatically.
-    // Preserve trusted Claude Code runtime flags set by config.ts. In
-    // particular, CLAUDE_CODE_DISABLE_1M_CONTEXT defaults on so background
-    // helper queries do not silently re-enable the 1M context beta.
-    if (process.env.CLAUDE_CODE_DISABLE_1M_CONTEXT !== undefined) {
-        sanitized.CLAUDE_CODE_DISABLE_1M_CONTEXT = process.env.CLAUDE_CODE_DISABLE_1M_CONTEXT;
-    }
     // Step 3: Add trusted markers AFTER sanitization
     sanitized.CLEMENTINE_HOME = BASE_DIR;
     return sanitized;
@@ -2148,7 +2140,10 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             && toolsDisabledForCall
             && turnPolicy?.retrievalTier === 'none'
             && turnPolicy.effort === 'low';
-        const resolvedModel = resolveModel(requestedModel) ?? (lightweightModelEligible ? MODELS.haiku : MODEL);
+        const rawResolvedModel = resolveModel(requestedModel) ?? (lightweightModelEligible ? MODELS.haiku : MODEL);
+        const resolvedModel = normalizeClaudeModelForOneMillionContext(rawResolvedModel);
+        const oneMillionModeValue = currentOneMillionContextMode();
+        const oneMillionDisableValue = claudeCodeDisableOneMillionForModel(resolvedModel);
         const modelRouteReason = model
             ? 'explicit'
             : profile?.model
@@ -2421,7 +2416,7 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             systemPrompt: fullSystemPrompt,
             model: resolvedModel,
             ...(fallback ? { fallbackModel: fallback } : {}),
-            ...(oneMillionContextDisabled() ? { betas: [] } : {}),
+            ...(oneMillionDisableValue === '1' ? { betas: [] } : {}),
             permissionMode: effectivePermissionMode,
             allowDangerouslySkipPermissions: true,
             ...(sessionStore ? { sessionStore } : {}),
@@ -2459,6 +2454,10 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                             CLEMENTINE_TEAM_AGENT: profile?.slug ?? 'clementine',
                             CLEMENTINE_INTERACTION_SOURCE: sourceOverride ?? inferInteractionSource(sessionKey),
                             CLEMENTINE_TOOL_ALLOWLIST: clementineToolAllowlist,
+                            CLEMENTINE_1M_CONTEXT_MODE: oneMillionModeValue,
+                            ...(oneMillionDisableValue !== undefined
+                                ? { CLAUDE_CODE_DISABLE_1M_CONTEXT: oneMillionDisableValue }
+                                : {}),
                         },
                     },
                     ...externalMcpServers,
@@ -2472,14 +2471,21 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             // env only when the prompt/job mentions a connector-backed service.
             // Per-MCP-server env isolation still happens inside each mcpServers
             // entry; this only affects the Claude Code subprocess itself.
-            ...(shouldInheritClaudeEnv ? {} : {
-                env: {
+            env: shouldInheritClaudeEnv
+                ? {
+                    ...process.env,
+                    CLEMENTINE_1M_CONTEXT_MODE: oneMillionModeValue,
+                    ...(oneMillionDisableValue !== undefined
+                        ? { CLAUDE_CODE_DISABLE_1M_CONTEXT: oneMillionDisableValue }
+                        : {}),
+                }
+                : {
                     ...SAFE_ENV,
-                    ...(process.env.CLAUDE_CODE_DISABLE_1M_CONTEXT !== undefined
-                        ? { CLAUDE_CODE_DISABLE_1M_CONTEXT: process.env.CLAUDE_CODE_DISABLE_1M_CONTEXT }
+                    CLEMENTINE_1M_CONTEXT_MODE: oneMillionModeValue,
+                    ...(oneMillionDisableValue !== undefined
+                        ? { CLAUDE_CODE_DISABLE_1M_CONTEXT: oneMillionDisableValue }
                         : {}),
                 },
-            }),
             // Avoid ambient Claude Code user/project/local settings and plugins by
             // default. Those can silently attach hundreds of tools. Explicit MCP
             // servers above still work; "all integrations/full tool surface" keeps
@@ -3421,13 +3427,14 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                                         responseText = responseText || ('Claude says the account credit balance is too low. I paused background jobs for a few hours so they stop draining/retrying, but interactive chat will also fail until credits are available again.');
                                     }
                                     else if (looksLikeOneMillionContextError(errorText)) {
+                                        process.env.CLEMENTINE_1M_CONTEXT_MODE = 'off';
                                         process.env.CLAUDE_CODE_DISABLE_1M_CONTEXT = '1';
                                         if (sessionKey) {
                                             this.sessions.delete(sessionKey);
                                             this.exchangeCounts.set(sessionKey, 0);
                                             this._compactedSessions.delete(sessionKey);
                                         }
-                                        responseText = responseText || ("Claude rejected the 1M context beta for this account. I've disabled 1M context for this process and reset the session. To persist the fix across restarts, run `clementine config doctor --fix`, then `clementine restart`.");
+                                        responseText = responseText || ("Claude rejected 1M context for this account. I've switched this process to 200K recovery mode and reset the session. To persist the fix across restarts, run `clementine budgets safe`, then `clementine restart`.");
                                     }
                                     else if (lower.includes('rate') && lower.includes('limit')) {
                                         hitRateLimit = true;
@@ -3552,13 +3559,14 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                         responseText = responseText || ('Claude says the account credit balance is too low. I paused background jobs for a few hours so they stop draining/retrying, but interactive chat will also fail until credits are available again.');
                     }
                     else if (looksLikeOneMillionContextError(e)) {
+                        process.env.CLEMENTINE_1M_CONTEXT_MODE = 'off';
                         process.env.CLAUDE_CODE_DISABLE_1M_CONTEXT = '1';
                         if (sessionKey) {
                             this.sessions.delete(sessionKey);
                             this.exchangeCounts.set(sessionKey, 0);
                             this._compactedSessions.delete(sessionKey);
                         }
-                        responseText = responseText || ("Claude rejected the 1M context beta for this account. I've disabled 1M context for this process and reset the session. To persist the fix across restarts, run `clementine config doctor --fix`, then `clementine restart`.");
+                        responseText = responseText || ("Claude rejected 1M context for this account. I've switched this process to 200K recovery mode and reset the session. To persist the fix across restarts, run `clementine budgets safe`, then `clementine restart`.");
                     }
                     else if (errStr.includes('rate') && (errStr.includes('limit') || errStr.includes('rate_limit'))) {
                         hitRateLimit = true;
@@ -4814,6 +4822,7 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                         throw new Error(errText);
                     }
                     if (looksLikeOneMillionContextError(errText)) {
+                        process.env.CLEMENTINE_1M_CONTEXT_MODE = 'off';
                         process.env.CLAUDE_CODE_DISABLE_1M_CONTEXT = '1';
                         throw new Error(errText);
                     }
@@ -5162,6 +5171,7 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                                 throw new Error(exitText);
                             }
                             if (looksLikeOneMillionContextError(exitText)) {
+                                process.env.CLEMENTINE_1M_CONTEXT_MODE = 'off';
                                 process.env.CLAUDE_CODE_DISABLE_1M_CONTEXT = '1';
                                 throw new Error(exitText);
                             }

package/dist/brain/adapters/markdown.js

@@ -33,8 +33,13 @@ export async function* parseMarkdown(filePath) {
         mtime = statSync(filePath).mtime.toISOString();
     }
     catch { /* ignore */ }
+    const frontmatterExternalId = typeof parsed.data?.externalId === 'string' && parsed.data.externalId.trim()
+        ? parsed.data.externalId.trim()
+        : typeof parsed.data?.external_id === 'string' && parsed.data.external_id.trim()
+            ? parsed.data.external_id.trim()
+            : null;
     yield {
-        externalId: `md-${hint}-${contentHash(body)}`,
+        externalId: frontmatterExternalId ?? `md-${hint}-${contentHash(body)}`,
         content: body,
         rawPayload: raw,
         metadata: {

package/dist/brain/connector-recipes.d.ts

@@ -3,8 +3,8 @@
  *
  * Each recipe is a blueprint for a one-click "auto-seed feed" that turns an
  * authenticated tool source (Claude Desktop connector, Composio toolkit, or
- * local MCP server) into a scheduled data feed that writes into the brain's
- * ingest folder.
+ * local MCP server) into a scheduled data feed that writes distilled notes
+ * into the brain's ingest folder.
  *
  * A feed materializes as:
  *   1. A CRON.md job entry with `managed: connector-feed` frontmatter
@@ -12,8 +12,8 @@
  *
  * The cron prompt tells the Claude Code agent to use the integration's MCP
  * tools to pull records, compare them with current memory when appropriate,
- * then call `brain_ingest_folder` to commit them — which writes markdown files
- * and runs the distillation pipeline in one step.
+ * then call `brain_ingest_folder` to commit them — which writes distilled
+ * markdown notes and indexes them in one step.
  *
  * Field syntax in prompt templates:
  *   {{fieldKey}}   — user-supplied value

package/dist/brain/connector-recipes.js

@@ -3,8 +3,8 @@
  *
  * Each recipe is a blueprint for a one-click "auto-seed feed" that turns an
  * authenticated tool source (Claude Desktop connector, Composio toolkit, or
- * local MCP server) into a scheduled data feed that writes into the brain's
- * ingest folder.
+ * local MCP server) into a scheduled data feed that writes distilled notes
+ * into the brain's ingest folder.
  *
  * A feed materializes as:
  *   1. A CRON.md job entry with `managed: connector-feed` frontmatter
@@ -12,8 +12,8 @@
  *
  * The cron prompt tells the Claude Code agent to use the integration's MCP
  * tools to pull records, compare them with current memory when appropriate,
- * then call `brain_ingest_folder` to commit them — which writes markdown files
- * and runs the distillation pipeline in one step.
+ * then call `brain_ingest_folder` to commit them — which writes distilled
+ * markdown notes and indexes them in one step.
  *
  * Field syntax in prompt templates:
  *   {{fieldKey}}   — user-supplied value
@@ -33,16 +33,16 @@ const COMMIT_INSTRUCTIONS = `When you have the records collected, call the \`bra
 - \`slug\`: "{{slug}}"
 - \`records\`: an array of \`{title, externalId, content, metadata}\` objects (one per item). \`externalId\` should be the source provider's stable id so re-runs dedup. \`metadata\` can include any fields you want preserved (url, modifiedAt, author).
 
-That tool writes each record to \`{{targetFolder}}/\` and runs the brain's distillation pipeline. You do NOT need to use Write — brain_ingest_folder handles file creation. Finish by reporting a one-line summary like "Ingested N new records, M unchanged".
+That tool runs the brain's distillation pipeline and writes the final notes to \`{{targetFolder}}/\`. You do NOT need to use Write — brain_ingest_folder handles note creation and indexing. Finish by reporting a one-line summary like "Ingested N new records, M unchanged".
 
 If the tool returns an error, include the error text in your summary.`;
-const MEMORY_DELTA_INSTRUCTIONS = `Before committing, call \`memory_recall\` for the feed slug/topic and use the returned chunks as the current memory state for this source. Keep records that are new, materially changed, or contain a new finding. Drop exact duplicates and rows that add no useful information. The ingestion pipeline will write markdown and embeddings; do not call \`memory_write\` for these feed records.`;
+const MEMORY_DELTA_INSTRUCTIONS = `Before committing, call \`memory_recall\` for the feed slug/topic and use the returned chunks as the current memory state for this source. Keep records that are new, materially changed, or contain a new finding. Drop exact duplicates and rows that add no useful information. The ingestion pipeline will write markdown, chunk it, and index it for recall; do not call \`memory_write\` for these feed records.`;
 // ── Recipes ────────────────────────────────────────────────────────────
 export const RECIPES = [
     {
         id: 'tool-backed-memory-seed',
-        label: 'Any tool: call and seed memory',
-        description: 'Call a selected tool from this connector, compare results with current memory, and ingest new or changed findings.',
+        label: 'Seed memory from this tool',
+        description: 'Pick one tool, fetch records from it, compare them with current memory, and save only new or changed findings.',
         icon: '🔌',
         integration: '*',
         requiredTools: [],
@@ -52,36 +52,36 @@ export const RECIPES = [
                 label: 'Memory topic',
                 placeholder: 'customers, calls, leads, deals, meetings...',
                 required: true,
-                help: 'Used for recall, deduping, and the generated feed slug.',
+                help: 'Used to search current memory and name this feed.',
             },
             {
                 key: 'toolName',
                 label: 'Tool to call',
                 required: true,
-                help: 'Pick the exact tool this feed should call when it runs.',
+                help: 'Pick the exact tool this feed should call each time it runs.',
             },
             {
                 key: 'callGoal',
-                label: 'What to pull',
+                label: 'What should Clementine fetch?',
                 placeholder: 'Fetch updated HubSpot contacts modified since the last run...',
                 required: true,
                 help: 'Describe the records to fetch, filters to apply, and any pagination bounds.',
             },
             {
                 key: 'variablesJson',
-                label: 'Variables JSON',
+                label: 'Tool variables (JSON)',
                 placeholder: '{"listId":"123","limit":100,"updatedAfter":"last_run"}',
-                help: 'Optional arguments, IDs, ranges, filters, or query variables the tool should use.',
+                help: 'Optional. Use {} if the tool needs no arguments.',
             },
             {
                 key: 'recordStrategy',
-                label: 'Record strategy',
+                label: 'How to save each result',
                 placeholder: 'One record per contact. Use email as stable id. Summarize lifecycle stage, owner, last activity, and new changes.',
-                help: 'Tell the agent how to convert the tool output into memory records.',
+                help: 'Tell Clementine what counts as one memory record and which field is the stable id.',
             },
             {
                 key: 'slug',
-                label: 'Slug override',
+                label: 'Memory bucket name (optional)',
                 placeholder: 'hubspot-contacts',
                 help: 'Optional. Leave blank to derive one from the connector and topic.',
             },
@@ -111,16 +111,16 @@ Tool source:
 
 Goal: ${v.callGoal || `Call ${v.toolName} and ingest useful returned data into memory.`}
 
-Variables JSON:
+Tool variables JSON:
 \`\`\`json
 ${(v.variablesJson || '{}').trim() || '{}'}
 \`\`\`
 
-Record strategy:
+How to save each result:
 ${v.recordStrategy || 'Convert the tool response into one memory record per returned entity or event. Use the provider stable id when available; otherwise use a deterministic hash of the source, topic, and meaningful record key.'}
 
 Steps:
-1. Call exactly this selected tool: \`${v.toolName}\`. Use the Variables JSON and the Goal above as the tool-call inputs. If the tool schema needs differently named arguments, map the provided variables to that schema. Do not switch to a different external tool unless this tool returns a clear instruction that another tool is required to read the selected records.
+1. Call exactly this selected tool: \`${v.toolName}\`. Use the Tool variables JSON and the Goal above as the tool-call inputs. If the tool schema needs differently named arguments, map the provided variables to that schema. Do not switch to a different external tool unless this tool returns a clear instruction that another tool is required to read the selected records.
 2. If the tool supports pagination or modified-since filters, prefer new/updated records and stop after ${limit} records. If no modified-since filter is available, fetch the most relevant ${limit} records.
 3. Normalize the tool result into candidate records. Preserve stable ids, URLs, timestamps, owners/authors, status fields, and provider metadata. Skip empty or purely administrative records.
 4. ${MEMORY_DELTA_INSTRUCTIONS}