Files
claude-mem/src/sdk/prompts.ts
T
Alex Newman 3ea0b60b9f feat: Mode system with inheritance and multilingual support (#412)
* feat: add domain management system with support for multiple domain profiles

- Introduced DomainManager class for loading and managing domain profiles.
- Added support for a default domain ('code') and fallback mechanisms.
- Implemented domain configuration validation and error handling.
- Created types for domain configuration, observation types, and concepts.
- Added new directory for domain profiles and ensured its existence.
- Updated SettingsDefaultsManager to include CLAUDE_MEM_DOMAIN setting.

* Refactor domain management to mode management

- Removed DomainManager class and replaced it with ModeManager for better clarity and functionality.
- Updated types from DomainConfig to ModeConfig and DomainPrompts to ModePrompts.
- Changed references from domains to modes in the settings and paths.
- Ensured backward compatibility by maintaining the fallback mechanism to the 'code' mode.

* feat: add migration 008 to support mode-agnostic observations and refactor service layer references in documentation

* feat: add new modes for code development and email investigation with detailed observation types and concepts

* Refactor observation parsing and prompt generation to incorporate mode-specific configurations

- Updated `parseObservations` function to use 'observation' as a universal fallback type instead of 'change', utilizing active mode's valid observation types.
- Modified `buildInitPrompt` and `buildContinuationPrompt` functions to accept a `ModeConfig` parameter, allowing for dynamic prompt content based on the active mode.
- Enhanced `ModePrompts` interface to include additional guidance for observers, such as recording focus and skip guidance.
- Adjusted the SDKAgent to load the active mode and pass it to prompt generation functions, ensuring prompts are tailored to the current mode's context.

* fix: correct mode prompt injection to preserve exact wording and type list visibility

- Add script to extract prompts from main branch prompts.ts into code.yaml
- Fix prompts.ts to show type list in XML template (e.g., "[ bugfix | feature | ... ]")
- Keep 'change' as fallback type in parser.ts (maintain backwards compatibility)
- Regenerate code.yaml with exact wording from original hardcoded prompts
- Build succeeds with no TypeScript errors

* fix: update ModeManager to load JSON mode files and improve validation

- Changed ModeManager to load mode configurations from JSON files instead of YAML.
- Removed the requirement for an "observation" type and updated validation to require at least one observation type.
- Updated fallback behavior in the parser to use the first type from the active mode's type list.
- Added comprehensive tests for mode loading, prompt injection, and parser integration, ensuring correct behavior across different modes.
- Introduced new mode JSON files for "Code Development" and "Email Investigation" with detailed observation types and prompts.

* Add mode configuration loading and update licensing information for Ragtime

- Implemented loading of mode configuration in WorkerService before database initialization.
- Added PolyForm Noncommercial License 1.0.0 to Ragtime directory.
- Created README.md for Ragtime with licensing details and usage guidelines.

* fix: add datasets directory to .gitignore to prevent accidental commits

* refactor: remove unused plugin package.json file

* chore: add package.json for claude-mem plugin with version 7.4.5

* refactor: remove outdated tests and improve error handling

- Deleted tests for ChromaSync error handling, smart install, strip memory tags, and user prompt tag stripping due to redundancy or outdated logic.
- Removed vitest configuration as it is no longer needed.
- Added a comprehensive implementation plan for fixing the modes system, addressing critical issues and improving functionality.
- Created a detailed test analysis report highlighting the quality and effectiveness of the current test suite, identifying areas for improvement.
- Introduced a new plugin package.json for runtime dependencies related to claude-mem hooks.

* refactor: remove parser regression tests to streamline codebase

* docs: update CLAUDE.md to clarify test management and changelog generation

* refactor: remove migration008 for mode-agnostic observations

* Refactor observation type handling to use ModeManager for icons and emojis

- Removed direct mappings of observation types to icons and work emojis in context-generator, FormattingService, SearchManager, and TimelineService.
- Integrated ModeManager to dynamically retrieve icons and emojis based on the active mode.
- Improved maintainability by centralizing the logic for observation type representation.

* Refactor observation metadata constants and update context generator

- Removed the explicit declaration of OBSERVATION_TYPES and OBSERVATION_CONCEPTS from observation-metadata.ts.
- Introduced fallback default strings for DEFAULT_OBSERVATION_TYPES_STRING and DEFAULT_OBSERVATION_CONCEPTS_STRING.
- Updated context-generator.ts to utilize observation types and concepts from ModeManager instead of constants.

* refactor: remove intermediate error handling from hooks (Phase 1)

Apply "fail fast" error handling strategy - errors propagate and crash loud
instead of being caught, wrapped, and re-thrown at intermediate layers.

Changes:
- Remove try/catch around fetch calls in all hooks - let errors throw
- Add try/catch ONLY around JSON.parse at entry points
- Delete error-handler.ts and hook-error-handler.ts (no longer needed)
- Update worker-utils.ts: functions now throw instead of returning null
- Update transcript-parser.ts: throw on missing path, empty file, malformed JSON
- Remove all handleWorkerError, handleFetchError imports

Philosophy: If something breaks, we KNOW it broke. No silent failures.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* refactor: remove intermediate error handling from worker service (Phase 2)

Apply "fail fast" error handling strategy to worker service layer.

Changes:
- worker-service.ts: Remove try/catch from version endpoint, cleanup,
  MCP close, process enumeration, force kill, and isAlive check
- SessionRoutes.ts: Remove try/catch from JSON.stringify calls, remove
  .catch() from Chroma sync and SDK agent calls
- SettingsRoutes.ts: Remove try/catch from toggleMcp()
- DatabaseManager.ts: Remove .catch() from backfill and close operations
- SDKAgent.ts: Keep outer try/catch (top-level), remove .catch() from
  Chroma sync operations
- SSEBroadcaster.ts: Remove try/catch from broadcast and sendToClient

Philosophy: Errors propagate and crash loud. BaseRouteHandler.wrapHandler
provides top-level catching for HTTP routes.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* refactor: remove error swallowing from SQLite services (Phase 3)

Apply "fail fast" error handling strategy to database layer.

SessionStore.ts migrations:
- ensureWorkerPortColumn(): Remove outer try/catch, let it throw
- ensurePromptTrackingColumns(): Remove outer try/catch, let it throw
- removeSessionSummariesUniqueConstraint(): Keep inner transaction
  rollback, remove outer catch
- addObservationHierarchicalFields(): Remove outer try/catch
- makeObservationsTextNullable(): Keep inner transaction rollback,
  remove outer catch
- createUserPromptsTable(): Keep inner transaction rollback, remove
  outer catch
- getFilesForSession(): Remove try/catch around JSON.parse

SessionSearch.ts:
- ensureFTSTables(): Remove try/catch, let it throw

Philosophy: Migration errors that are swallowed mean we think the
database is fine when it's not. Keep only inner transaction rollback
try/catch blocks.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* refactor: remove error hiding from utilities (Phase 4)

Apply "fail fast" error handling strategy to utility layer.

logger.ts:
- formatTool(): Remove try/catch, let JSON.parse throw on malformed input

context-generator.ts:
- loadContextConfig(): Remove try/catch, let parseInt throw on invalid settings
- Transcript extraction: Remove try/catch, let file read errors propagate

ChromaSync.ts:
- close(): Remove nested try/catch blocks, let close errors propagate

Philosophy: No silent fallbacks or hidden defaults. If something breaks,
we know it broke immediately.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* feat: serve static UI assets and update package root path

- Added middleware to serve static UI assets (JS, CSS, fonts, etc.) in ViewerRoutes.
- Updated getPackageRoot function to correctly return the package root directory as one level up from the current directory.

* feat: Enhance mode loading with inheritance support

- Introduced parseInheritance method to handle parent--override mode IDs.
- Added deepMerge method for recursively merging mode configurations.
- Updated loadMode method to support inheritance, loading parent modes and applying overrides.
- Improved error handling for missing mode files and logging for better traceability.

* fix(modes): correct inheritance file resolution and path handling

* Refactor code structure for improved readability and maintainability

* feat: Add mode configuration documentation and examples

* fix: Improve concurrency handling in translateReadme function

* Refactor SDK prompts to enhance clarity and structure

- Updated the `buildInitPrompt` and `buildContinuationPrompt` functions in `prompts.ts` to improve the organization of prompt components, including the addition of language instructions and footer messages.
- Removed redundant instructions and emphasized the importance of recording observations.
- Modified the `ModePrompts` interface in `types.ts` to include new properties for system identity, language instructions, and output format header, ensuring better flexibility and clarity in prompt generation.

* Enhance prompts with language instructions and XML formatting

- Updated `buildInitPrompt`, `buildSummaryPrompt`, and `buildContinuationPrompt` functions to include detailed language instructions in XML comments.
- Ensured that language instructions guide users to keep XML tags in English while writing content in the specified language.
- Modified the `buildSummaryPrompt` function to accept `mode` as a parameter for consistency.
- Adjusted the call to `buildSummaryPrompt` in `SDKAgent` to pass the `mode` argument.

* Refactor XML prompt generation in SDK

- Updated the buildInitPrompt, buildSummaryPrompt, and buildContinuationPrompt functions to use new placeholders for XML elements, improving maintainability and readability.
- Removed redundant language instructions in comments for clarity.
- Added new properties to ModePrompts interface for better structure and organization of XML placeholders and section headers.

* feat: Update observation prompts and structure across multiple languages

* chore: Remove planning docs and update Ragtime README

Remove ephemeral development artifacts:
- .claude/plans/modes-system-fixes.md
- .claude/test-analysis-report.md
- PROMPT_INJECTION_ANALYSIS.md

Update ragtime/README.md to explain:
- Feature is not yet implemented
- Dependency on modes system (now complete in PR #412)
- Ready to be scripted out in future release

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

* fix: Move summary prompts to mode files for multilingual support

Summary prompts were hardcoded in English in prompts.ts, breaking
multilingual support. Now properly mode-based:

- Added summary_instruction, summary_context_label,
  summary_format_instruction, summary_footer to code.json
- Updated buildSummaryPrompt() to use mode fields instead of hardcoded text
- Added summary_footer with language instructions to all 10 language modes
- Language modes keep English prompts + language requirement footer

This fixes the gaslighting where we claimed full multilingual support
but summaries were still generated in English.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

* chore: Clean up README by removing local preview instructions and streamlining beta features section

* Add translated README files for Ukrainian, Vietnamese, and Chinese languages

* Add new language modes for code development in multiple languages

- Introduced JSON configurations for Code Development in Greek, Finnish, Hebrew, Hindi, Hungarian, Indonesian, Italian, Dutch, Norwegian, Polish, Brazilian Portuguese, Romanian, Swedish, Turkish, and Ukrainian.
- Each configuration includes prompts for observations, summaries, and instructions tailored to the respective language.
- Ensured that all prompts emphasize the importance of generating observations without referencing the agent's actions.

* Add multilingual support links to README files in various languages

- Updated README.id.md, README.it.md, README.ja.md, README.ko.md, README.nl.md, README.no.md, README.pl.md, README.pt-br.md, README.ro.md, README.ru.md, README.sv.md, README.th.md, README.tr.md, README.uk.md, README.vi.md, and README.zh.md to include links to other language versions.
- Each README now features a centered paragraph with flags and links for easy navigation between different language documents.

---------

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
2025-12-22 20:14:18 -05:00

230 lines
6.9 KiB
TypeScript

/**
* SDK Prompts Module
* Generates prompts for the Claude Agent SDK memory worker
*/
import { logger } from '../utils/logger.js';
import type { ModeConfig } from '../services/domain/types.js';
export interface Observation {
id: number;
tool_name: string;
tool_input: string;
tool_output: string;
created_at_epoch: number;
cwd?: string;
}
export interface SDKSession {
id: number;
sdk_session_id: string | null;
project: string;
user_prompt: string;
last_user_message?: string;
last_assistant_message?: string;
}
/**
* Build initial prompt to initialize the SDK agent
*/
export function buildInitPrompt(project: string, sessionId: string, userPrompt: string, mode: ModeConfig): string {
return `${mode.prompts.system_identity}
<observed_from_primary_session>
<user_request>${userPrompt}</user_request>
<requested_at>${new Date().toISOString().split('T')[0]}</requested_at>
</observed_from_primary_session>
${mode.prompts.observer_role}
${mode.prompts.spatial_awareness}
${mode.prompts.recording_focus}
${mode.prompts.skip_guidance}
${mode.prompts.output_format_header}
\`\`\`xml
<observation>
<type>[ ${mode.observation_types.map(t => t.id).join(' | ')} ]</type>
<!--
${mode.prompts.type_guidance}
-->
<title>${mode.prompts.xml_title_placeholder}</title>
<subtitle>${mode.prompts.xml_subtitle_placeholder}</subtitle>
<facts>
<fact>${mode.prompts.xml_fact_placeholder}</fact>
<fact>${mode.prompts.xml_fact_placeholder}</fact>
<fact>${mode.prompts.xml_fact_placeholder}</fact>
</facts>
<!--
${mode.prompts.field_guidance}
-->
<narrative>${mode.prompts.xml_narrative_placeholder}</narrative>
<concepts>
<concept>${mode.prompts.xml_concept_placeholder}</concept>
<concept>${mode.prompts.xml_concept_placeholder}</concept>
</concepts>
<!--
${mode.prompts.concept_guidance}
-->
<files_read>
<file>${mode.prompts.xml_file_placeholder}</file>
<file>${mode.prompts.xml_file_placeholder}</file>
</files_read>
<files_modified>
<file>${mode.prompts.xml_file_placeholder}</file>
<file>${mode.prompts.xml_file_placeholder}</file>
</files_modified>
</observation>
\`\`\`
${mode.prompts.format_examples}
${mode.prompts.footer}
${mode.prompts.header_memory_start}`;
}
/**
* Build prompt to send tool observation to SDK agent
*/
export function buildObservationPrompt(obs: Observation): string {
// Safely parse tool_input and tool_output - they're already JSON strings
let toolInput: any;
let toolOutput: any;
try {
toolInput = typeof obs.tool_input === 'string' ? JSON.parse(obs.tool_input) : obs.tool_input;
} catch {
toolInput = obs.tool_input; // If parse fails, use raw value
}
try {
toolOutput = typeof obs.tool_output === 'string' ? JSON.parse(obs.tool_output) : obs.tool_output;
} catch {
toolOutput = obs.tool_output; // If parse fails, use raw value
}
return `<observed_from_primary_session>
<what_happened>${obs.tool_name}</what_happened>
<occurred_at>${new Date(obs.created_at_epoch).toISOString()}</occurred_at>${obs.cwd ? `\n <working_directory>${obs.cwd}</working_directory>` : ''}
<parameters>${JSON.stringify(toolInput, null, 2)}</parameters>
<outcome>${JSON.stringify(toolOutput, null, 2)}</outcome>
</observed_from_primary_session>`;
}
/**
* Build prompt to generate progress summary
*/
export function buildSummaryPrompt(session: SDKSession, mode: ModeConfig): string {
const lastAssistantMessage = session.last_assistant_message || logger.happyPathError(
'SDK',
'Missing last_assistant_message in session for summary prompt',
{ sessionId: session.id },
undefined,
''
);
return `${mode.prompts.header_summary_checkpoint}
${mode.prompts.summary_instruction}
${mode.prompts.summary_context_label}
${lastAssistantMessage}
${mode.prompts.summary_format_instruction}
<summary>
<request>${mode.prompts.xml_summary_request_placeholder}</request>
<investigated>${mode.prompts.xml_summary_investigated_placeholder}</investigated>
<learned>${mode.prompts.xml_summary_learned_placeholder}</learned>
<completed>${mode.prompts.xml_summary_completed_placeholder}</completed>
<next_steps>${mode.prompts.xml_summary_next_steps_placeholder}</next_steps>
<notes>${mode.prompts.xml_summary_notes_placeholder}</notes>
</summary>
${mode.prompts.summary_footer}`;
}
/**
* Build prompt for continuation of existing session
*
* CRITICAL: Why claudeSessionId Parameter is Required
* ====================================================
* This function receives claudeSessionId from SDKAgent.ts, which comes from:
* - SessionManager.initializeSession (fetched from database)
* - SessionStore.createSDKSession (stored by new-hook.ts)
* - new-hook.ts receives it from Claude Code's hook context
*
* The claudeSessionId is the SAME session_id used by:
* - NEW hook (to create/fetch session)
* - SAVE hook (to store observations)
* - This continuation prompt (to maintain session context)
*
* This is how everything stays connected - ONE session_id threading through
* all hooks and prompts in the same conversation.
*
* Called when: promptNumber > 1 (see SDKAgent.ts line 150)
* First prompt: Uses buildInitPrompt instead (promptNumber === 1)
*/
export function buildContinuationPrompt(userPrompt: string, promptNumber: number, claudeSessionId: string, mode: ModeConfig): string {
return `${mode.prompts.continuation_greeting}
<observed_from_primary_session>
<user_request>${userPrompt}</user_request>
<requested_at>${new Date().toISOString().split('T')[0]}</requested_at>
</observed_from_primary_session>
${mode.prompts.system_identity}
${mode.prompts.observer_role}
${mode.prompts.spatial_awareness}
${mode.prompts.recording_focus}
${mode.prompts.skip_guidance}
${mode.prompts.continuation_instruction}
${mode.prompts.output_format_header}
\`\`\`xml
<observation>
<type>[ ${mode.observation_types.map(t => t.id).join(' | ')} ]</type>
<!--
${mode.prompts.type_guidance}
-->
<title>${mode.prompts.xml_title_placeholder}</title>
<subtitle>${mode.prompts.xml_subtitle_placeholder}</subtitle>
<facts>
<fact>${mode.prompts.xml_fact_placeholder}</fact>
<fact>${mode.prompts.xml_fact_placeholder}</fact>
<fact>${mode.prompts.xml_fact_placeholder}</fact>
</facts>
<!--
${mode.prompts.field_guidance}
-->
<narrative>${mode.prompts.xml_narrative_placeholder}</narrative>
<concepts>
<concept>${mode.prompts.xml_concept_placeholder}</concept>
<concept>${mode.prompts.xml_concept_placeholder}</concept>
</concepts>
<!--
${mode.prompts.concept_guidance}
-->
<files_read>
<file>${mode.prompts.xml_file_placeholder}</file>
<file>${mode.prompts.xml_file_placeholder}</file>
</files_read>
<files_modified>
<file>${mode.prompts.xml_file_placeholder}</file>
<file>${mode.prompts.xml_file_placeholder}</file>
</files_modified>
</observation>
\`\`\`
${mode.prompts.format_examples}
${mode.prompts.footer}
${mode.prompts.header_memory_continued}`;
}