| Age | Commit message (Collapse) | Author |
|---|
|
Security improvement: subagents now require explicit user confirmation via
Telegram inline buttons, preventing the agent from bypassing approval.
Changes:
- Add InlineKeyboardMarkup, InlineKeyboardButton, CallbackQuery types
- Add parseCallbackQuery for handling button presses
- Add sendMessageWithKeyboard and answerCallbackQuery API functions
- Add PendingSpawn registry for tracking unconfirmed spawn requests
- Add spawnSubagentToolWithApproval that sends approval buttons
- Add handleCallbackQuery to process approve/reject button clicks
- Add approveAndSpawnSubagent and rejectPendingSpawn functions
Flow:
1. Agent calls spawn_subagent → creates pending request
2. User receives message with ✅ Approve / ❌ Reject buttons
3. Button click (outside agent loop) spawns or cancels
4. Pending requests expire after 10 minutes
|
|
- Add global subagent registry to track running handles by ID
- Modify executeSpawnSubagent to use spawnSubagentAsync and return immediately
- Add check_subagent tool for querying status or getting results
- Export subagentTools convenience function with both tools
- Update Telegram.hs to use subagentTools instead of just spawnSubagentTool
Ava can now spawn subagents in the background and continue
conversations while checking on progress via check_subagent.
|
|
Based on Anthropic's effective harnesses research.
New modules:
- Omni/Agent/AuditLog.hs: JSONL audit logging with SubagentId linking
- Omni/Agent/Tools/AvaLogs.hs: Tool for Ava to query her own logs
- Omni/Agent/Subagent/HARDENING.md: Design documentation
Key features:
- SubagentHandle with TVar status for async execution and polling
- spawnSubagentAsync, querySubagentStatus, waitSubagent, cancelSubagent
- User confirmation: spawn_subagent requires confirmed=true after approval
- Audit logs stored in $AVA_DATA_ROOT/logs/{ava,subagents}/
- CLI: ava logs [--last=N] [<subagent_id>]
- read_ava_logs tool for Ava self-diagnosis
Tasks: t-267, t-268, t-269, t-270, t-271
|
|
- Wrap ava binary with Python (requests, httpx, pandas, etc.)
- Add CLI tools: curl, pandoc, ffmpeg, imagemagick, csvkit
- Add local whisper-cli fallback when OPENAI_API_KEY unavailable
Amp-Thread-ID: https://ampcode.com/threads/T-019b2dc2-36e0-75e1-90c1-622901fc9847
Co-authored-by: Amp <amp@ampcode.com>
|
|
|
|
- Add LineBuffering to stdout/stderr in Ava.hs
- Add StandardOutput/StandardError=journal to systemd service
|
|
|
|
- Add Omni.Agent.Paths module for configurable AVA_DATA_ROOT
- Create ava Linux user in Users.nix with SSH key
- Add systemd service in Beryllium/Ava.nix with graceful shutdown
- Update Skills.hs and Outreach.hs to use configurable paths
- Add startup logging of resolved paths in Telegram.hs
- Create migration script for moving data from _/var/ava to /home/ava
- Add deployment documentation in Beryllium/AVA.md
In dev: AVA_DATA_ROOT unset uses _/var/ava/
In prod: AVA_DATA_ROOT=/home/ava via systemd
Amp-Thread-ID: https://ampcode.com/threads/T-019b2d7e-bd88-7355-8133-275c65157aaf
Co-authored-by: Amp <amp@ampcode.com>
|
|
Enables orchestrator to spawn specialized subagents for focused tasks:
- WebCrawler: web search + page reading (haiku, fast)
- CodeReviewer: code analysis tools (sonnet, thorough)
- DataExtractor: structured data extraction (haiku)
- Researcher: combined web + codebase research (sonnet)
Key features:
- spawn_subagent tool with role-based tool selection
- Per-subagent resource limits (timeout, cost, tokens)
- Structured output with citations (claim, source_url, quote)
- Separate API keys for OpenRouter vs Kagi
- Efficiency-focused system prompts
Defaults: 200k tokens, $1.00 cost cap, 600s timeout, 20 iterations
|
|
Wrap parser chains in parentheses to fix compilation errors from
incorrect use of </ operator.
|
|
The deployer compared its in-memory stateServices map to decide if a
service needed restarting. When the deployer restarted, this state was
lost, causing it to think services were 'already at' the desired path
when they were actually running old code.
Changes:
- Add getRunningStorePath to Systemd module to read actual store path
- Update deployService to query systemd instead of stale in-memory state
- Add DerivingStrategies extension to Deployer.hs
|
|
|
|
- Systemd: use list comprehension instead of if-then-else
- Manifest: use </> operator, replace case with maybe
- Deployer: use newtype, use flip removeService
- Caddy: use newtype for single-field types
|
|
- Create Omni/Agent/Skills.hs with skill loader and tools
- Skills follow Claude Skills format (SKILL.md + scripts/references/assets)
- Directory structure: _/var/ava/skills/{shared,<user>}/
- Three tools: skill, list_skills, publish_skill
- Users can publish private skills to shared
- Integrate skills tools into Telegram bot
- Create skill-creator meta-skill at _/var/ava/skills/shared/skill-creator/
|
|
Added retry with backoff, parallel proccessing, editing pages down to main
content, summarization with haiku. It's so much faster and more reliable
now. Plus improved the logging system and distangled the status UI bar from the
logging module.
|
|
- Rename Omni/Bot.hs to Omni/Ava.hs
- Delete Omni/Bot.scm (unused Guile version)
- Remove cost limit (was 10 cents, now 0)
- Increase max iterations from 10 to 50
|
|
- Add Omni/Deploy/ with Manifest, Deployer, Systemd, Caddy modules
- Manifest CLI: show, update, add-service, list, rollback commands
- Deployer: polls S3 manifest, pulls closures, manages systemd units
- Caddy integration for dynamic reverse proxy routes
- bild: auto-cache to S3, outputs STORE_PATH for push.sh
- push.sh: supports both NixOS and service deploys
- Biz.nix: simplified to base OS + deployer only
- Services (podcastitlater-web/worker) now deployer-managed
- Documentation: README.md with operations guide
|
|
python_exec, http_get/post, outreach_*, feedback_list, and read_file
now require isBenAuthorized check, matching email/hledger pattern.
|
|
- Add feedback table with migration in Core.py
- Add FeedbackForm and FeedbackPage UI components
- Add /feedback GET/POST routes and /api/feedback JSON endpoint
- Add admin feedback view at /admin/feedback
- Create Omni/Agent/Tools/Feedback.hs with feedback_list tool
- Wire feedback tool into Telegram agent
|
|
- Add emailSendTool to Email.hs for sending approved drafts
- Add sendApprovedEmail function that checks draft status
- Use Network.Mail.Mime.simpleMail' with SMTP.sendMail
- Integrate with Outreach module to verify approval and mark sent
- Add tests for new tool
|
|
- Import Omni.Agent.Tools in Telegram.hs
- Add readFileTool to Ava's tool list
- Add podcastitlater context section to system prompt with key file paths
|
|
The haiku-based splitting was modifying message content. Replace
with deterministic paragraph-based splitting that preserves the
original text exactly.
|
|
- Create Omni/Agent/Tools/Outreach.hs with tools
- Drafts stored in _/var/ava/outreach/{pending,approved,...}
- Add Telegram commands: /review, /approve, /reject, /queue
- Integrate outreach tools into agent's tool list
Amp-Thread-ID: https://ampcode.com/threads/T-019b202c-2156-74db-aa4a-e0a2f4397fbb
Co-authored-by: Amp <amp@ampcode.com>
|
|
- Create Omni/Agent/Tools/Http.hs with http_get and http_post tools
- Support headers, query params, JSON body, 30s timeout
- Return structured JSON with status, headers, body
- Add 9 unit tests including real HTTP calls to httpbin.org
- Wire tools into Telegram agent's tool list
Completes t-265.2
|
|
- Create Omni/Agent/Tools/Python.hs with python_exec tool
- Execute Python snippets via subprocess with 30s default timeout
- Return structured JSON with stdout, stderr, exit_code
- Add 8 unit tests covering print, imports, errors, timeout
- Wire tool into Telegram agent's tool list
Completes t-265.1
|
|
- Replace openssl s_client with HaskellNet/HaskellNet-SSL for proper
IMAP client support (better protocol handling, no manual parsing)
- Add HaskellNet deps to Haskell.nix with doJailbreak for version bounds
- Fix lost messages: sendMessageReturningId now throws on API errors
instead of returning Nothing (which was incorrectly treated as success)
- Auto-retry markdown parse errors as plain text
- Hardcode benChatId for reliable email check loop startup
|
|
- New Omni/Agent/Tools/Hledger.hs with 5 tools:
- hledger_balance: query account balances
- hledger_register: show transaction history
- hledger_add: create new transactions
- hledger_income_statement: income vs expenses
- hledger_balance_sheet: net worth view
- All tools support currency parameter (default: USD)
- Balance, register, income_statement support period parameter
- Period uses hledger syntax (thismonth, 2024, from X to Y)
- Shell escaping fixed for multi-word period strings
- Authorization: only Ben and Kate get hledger tools
- Max iterations increased from 5 to 10
- Transactions written to ~/fund/telegram-transactions.journal
|
|
|
|
Memory changes:
- Add thread_id column to conversation_messages for topic support
- Add saveGroupMessage/getGroupConversationContext for shared history
- Add storeGroupMemory/recallGroupMemories with 'group:<chat_id>' user
- Fix SQLite busy error: set busy_timeout before journal_mode
Telegram changes:
- Group chats now use shared conversation context (chat_id, thread_id)
- Personal memories stay with user, group memories shared across group
- Memory context shows [Personal] and [Group] prefixes
- Add withTypingIndicator: refreshes typing every 4s while agent thinks
- Fix typing UX: indicator now shows continuously until response sent
|
|
- Parse message_thread_id from incoming messages
- Include thread_id in sendMessage API calls
- Pass thread_id through message queue system
- Replies now go to the correct topic in supergroups
|
|
OpenRouter's chat completion API doesn't properly pass audio to models.
Switched to calling OpenAI's /v1/audio/transcriptions endpoint directly
with the whisper-1 model.
Requires OPENAI_API_KEY environment variable.
|
|
Fixes 'database is locked' errors when multiple threads access the
memory database simultaneously (incoming batcher, message dispatch,
reminder loop, main handler).
|
|
Batches incoming messages by chat_id with a 3-second sliding window
before processing. This prevents confusion when messages arrive
simultaneously from different chats.
- New IncomingQueue module with STM-based in-memory queue
- Messages enqueued immediately, offset acked on enqueue
- 200ms tick loop flushes batches past deadline
- Batch formatting: numbered messages, sender attribution for groups,
media stubs, reply context
- Media from first message in batch still gets full processing
|
|
- Switch from gemini-2.0-flash-001 to gemini-2.5-flash
- Put audio content before text prompt (model was ignoring audio)
- Strengthen prompt to return only transcription
|
|
- Add Messages.hs with scheduled_messages table and dispatcher loop
- All outbound messages now go through the queue (1s polling)
- Disable streaming responses, use runAgentWithProvider instead
- Add send_message tool for delayed messages (up to 30 days)
- Add list_pending_messages and cancel_message tools
- Reminders now queue messages instead of sending directly
- Exponential backoff retry (max 5 attempts) for failed sends
|
|
Amp-Thread-ID: https://ampcode.com/threads/T-019b1894-b431-777d-aba3-65a51e720ef2
Co-authored-by: Amp <amp@ampcode.com>
|
|
|
|
- Add RelationType with 6 relation types
- Add MemoryLink type and memory_links table
- Add graph functions: linkMemories, getMemoryLinks, queryGraph
- Add link_memories and query_graph agent tools
- Wire up graph tools to Telegram bot
- Include memory ID in recall results for linking
- Fix streaming usage parsing for cost tracking
Closes t-255
Amp-Thread-ID: https://ampcode.com/threads/T-019b181f-d6cd-70de-8857-c445baef7508
Co-authored-by: Amp <amp@ampcode.com>
|
|
When the bot is added to a group, check if the user who added it is
in the whitelist. If not, send a message explaining and leave the group
immediately. This prevents unauthorized users from bypassing DM access
controls by adding the bot to a group.
|
|
|
|
Add parse_mode=Markdown to sendMessage and editMessage API calls
|
|
OpenAI's SSE streaming sends tool calls incrementally - the first chunk
has the id and function name, subsequent chunks contain argument fragments.
Previously each chunk was treated as a complete tool call, causing invalid
JSON arguments.
- Add ToolCallDelta type with index for partial tool call data
- Add StreamToolCallDelta chunk type
- Track tool calls by index in IntMap accumulator
- Merge argument fragments across chunks via mergeToolCallDelta
- Build final ToolCall objects from accumulator when stream ends
- Handle new StreamToolCallDelta in Engine.hs pattern match
|
|
Pre-filter now sees last 5 messages so it can detect when user
is continuing a conversation with Ava, even without explicit mention.
- Fetch recent messages before shouldEngageInGroup
- Update classifier prompt to understand Ava context
- Handle follow-up messages to bot's previous responses
|
|
- Update to Dec 2024 OpenRouter pricing
- Use blended input/output rates
- Add gemini-flash, claude-sonnet-4.5 specific rates
- Fix math: was off by ~30x for Claude models
|
|
Use Gemini Flash to classify group messages before running the
full Sonnet agent. Skips casual banter to save tokens/cost.
- shouldEngageInGroup: yes/no classifier using gemini-2.0-flash
- Only runs for group chats, private chats skip the filter
- On classifier failure, defaults to engaging (fail-open)
|
|
- Remove mention-based filtering, bot sees all group messages
- Add response rules to system prompt for group chats:
- tool invocation = always respond
- direct question = respond
- factual correction = maybe respond
- casual banter = stay silent
- Empty response in group = intentional silence (no fallback msg)
- Add chat type context to system prompt
|
|
- Only respond in groups when @mentioned or replied to
- Add ChatType to TelegramMessage (private/group/supergroup/channel)
- Add getMe API call to fetch bot username on startup
- Add shouldRespondInGroup helper function
|
|
- Fix Provider.hs to strip leading whitespace from OpenRouter responses
- Fix FunctionCall parser to handle missing 'arguments' field
- Use eitherDecode for better error messages on parse failures
- Switch to claude-sonnet-4.5 for main agent
- Use gemini-2.0-flash for conversation summarization (cheaper)
- Add read_webpage tool for fetching and summarizing URLs
- Add tagsoup to Haskell deps (unused, kept for future)
|
|
Refactor Telegram.hs into submodules to reduce file size:
- Types.hs: data types, JSON parsing
- Media.hs: file downloads, image/voice analysis
- Reminders.hs: reminder loop, user chat persistence
Multimedia improvements:
- Vision uses third-person to avoid LLM confusion
- Better message framing for embedded descriptions
- Size validation (10MB images, 20MB voice)
- MIME type validation for voice messages
New features:
- Reply support: bot sees context when users reply
- Web search: default 5->10, max 10->20 results
- Guardrails: duplicate tool limit 3->10 for research
- Timezone: todos parse/display in Eastern time (ET)
|
|
- Add TelegramPhoto and TelegramVoice types
- Parse photo and voice fields from Telegram updates
- Download photos/voice via Telegram API
- Analyze images using Claude vision via OpenRouter
- Transcribe voice messages using Gemini audio via OpenRouter
- Wire multimedia processing into handleAuthorizedMessage
Photos are analyzed with user's caption as context.
Voice messages are transcribed and treated as text input.
|
