@fickydev/pigent 0.1.6 → 0.1.8

package/CHANGELOG.md

@@ -39,6 +39,12 @@
 - Added safe Pi runner failure replies with internal error persistence.
 - Added per-session locking to serialize Pi runs for the same agent/channel/chat/thread.
 - Added Telegram chat instructions to Pi prompt composition.
+- Added Telegram API retry/backoff for transient network, rate limit, and server errors.
+- Planned model selection support across profile, agent, and Telegram chat override levels.
+- Added profile and agent config-based Pi model and thinking level selection.
+- Added session-scoped model and thinking level overrides with `/model` and `/thinking` Telegram commands.
+- Planned Telegram inline-button model and thinking level pickers backed by configured model choices.
+- Added Telegram inline-button pickers for `/model` and `/thinking`, callback handling, and bot command menu registration.
 - Kept daemon process alive after startup so CLI runs do not exit after `pigent ready`.
 
 ### Changed

package/PLAN.md

@@ -61,7 +61,8 @@ Profile defines reusable behavior shared by agents.
 
 Profile may include:
 
-- model preference
+- model preference using `provider/modelId`
+- thinking level preference (`off`, `low`, `medium`, `high`)
 - base system prompt
 - response style
 - default skills
@@ -152,6 +153,8 @@ Each Telegram group/private chat can define:
 - default agent
 - allowed agents
 - custom instructions
+- per-agent model override via Telegram command
+- per-agent thinking level override via Telegram command
 - heartbeat settings
 - enabled/disabled status
 
@@ -170,6 +173,61 @@ telegramChats:
       This is engineering group. Be concise. Prefer TypeScript.
 ```
 
+## Model Selection
+
+Pigent should support model selection at multiple levels without exposing provider secrets to channel prompts.
+
+Model identifiers use Pi SDK `provider/modelId` format, for example:
+
+```text
+anthropic/claude-opus-4-5
+openai/gpt-4.1
+ollama/qwen2.5-coder:7b
+```
+
+Selection priority:
+
+```text
+session override > agent config > profile config > Pi default
+```
+
+A session is keyed by `agentId + channel + chatId + threadId`, so the same agent can use different models in private chats, group chats, and Telegram forum topics.
+
+Config support should come first:
+
+- `profiles/*.yaml` may define `model` and `thinkingLevel`
+- `agents/*/agent.yaml` may override `model` and `thinkingLevel`
+
+Telegram command support should persist session-scoped overrides after config support exists:
+
+- `/model` shows current model for the default agent in the chat
+- `/model <provider/modelId>` sets model for the default agent session in the current chat/thread
+- `/model default` clears model override for the default agent session in the current chat/thread
+- `/thinking <level>` sets thinking level for the default agent session in the current chat/thread
+- `/thinking default` clears thinking override for the default agent session in the current chat/thread
+
+Telegram should also provide inline-button pickers so non-technical users can choose from configured choices:
+
+- `/model` can reply with inline keyboard buttons for configured model choices plus `Use default`
+- `/thinking` can reply with inline keyboard buttons for `Off`, `Low`, `Medium`, `High`, and `Use default`
+- callback payloads should be short and action-scoped, such as `model:set:<choiceId>`, `model:default`, `thinking:set:medium`, and `thinking:default`
+- after selection, bot should answer the callback and update or send a confirmation message
+- daemon startup should register Telegram bot commands through `setMyCommands` so users see Pigent commands in the Telegram command menu
+
+Model choices should be explicitly configured instead of exposing every Pi model blindly:
+
+```yaml
+modelChoices:
+  - id: anthropic/claude-sonnet-4-5
+    label: Claude Sonnet
+  - id: anthropic/claude-opus-4-5
+    label: Claude Opus
+  - id: openai/gpt-4.1
+    label: GPT-4.1
+```
+
+The command and callback handlers should validate agent access and model availability through a Pi model resolver. They must not reveal API keys, auth values, raw environment variables, or provider secrets.
+
 ## Prompt Composition
 
 For each request, compose context from:
@@ -195,6 +253,7 @@ Initial entities:
 - telegram chat agents
 - messages
 - heartbeats
+- session model/thinking overrides
 - tasks/events later
 
 SQLite first. Keep schema types portable so PostgreSQL migration is straightforward.
@@ -319,6 +378,8 @@ drizzle/
 ### Milestone 6: Pi Runner
 
 - create Pi SDK sessions
+- resolve profile/agent model preferences
+- resolve chat model overrides
 - send prompts
 - return responses
 - persist inbound/outbound messages
@@ -362,6 +423,8 @@ Add Hono only after daemon works. Use it for:
 
 ## Open Questions
 
+- Should `/model <agentId> ...` support non-default agent sessions in the command itself?
+- Should model availability errors block `/model` saves or save with a warning?
 - Which Pi SDK session manager should be used for persistent sessions?
 - How exactly should per-agent skills/extensions be loaded into Pi runtime?
 - Should group instructions live in YAML, DB, or both?

package/TODO.md

@@ -39,6 +39,10 @@
 
 ## Config
 
+- [x] Add root `modelChoices` config for Telegram model picker buttons
+- [x] Add profile-level `thinkingLevel` config
+- [x] Add agent-level `model` override config
+- [x] Add agent-level `thinkingLevel` override config
 - [x] Define root config format
 - [x] Define `AgentConfig`
 - [x] Define `ProfileConfig`
@@ -66,6 +70,8 @@
 - [x] Define `messages` table
 - [x] Define `heartbeats` table
 - [x] Define `runtime_kv` table for offsets and daemon state
+- [x] Add `agent_sessions.model` for session model override
+- [x] Add `agent_sessions.thinking_level` for session thinking override
 - [x] Implement repositories
   - [x] `AgentRepository`
   - [x] `SessionRepository`
@@ -79,11 +85,15 @@
 - [x] Define `OutboundMessage`
 - [x] Define `ChannelAdapter`
 - [x] Define Telegram normalized types
+- [x] Normalize Telegram callback query updates
 - [ ] Implement `TelegramApi`
   - [x] `getUpdates`
   - [x] `sendMessage`
   - [x] error handling
-  - [ ] retry/backoff
+  - [x] retry/backoff
+  - [x] `answerCallbackQuery`
+  - [x] inline keyboard `reply_markup` support
+  - [x] `setMyCommands` Telegram command menu registration
 - [x] Implement `TelegramPollingAdapter`
   - [x] polling loop
   - [x] offset tracking
@@ -121,6 +131,12 @@
 
 ## Pi Integration
 
+- [x] Confirm SDK model selection supports `createAgentSession({ model, thinkingLevel })`
+- [x] Implement Pi model resolver for `provider/modelId`
+- [x] Apply profile-level model selection
+- [x] Apply agent-level model override
+- [x] Apply session model override
+- [x] Apply thinking level selection
 - [x] Confirm SDK APIs needed for session creation
 - [ ] Confirm persistent session manager approach
 - [ ] Confirm per-agent system prompt injection
@@ -162,6 +178,19 @@
 - [ ] `/sessions` list active sessions for chat
 - [ ] `/reset-session <agentId>` clear chat session
 - [ ] `/heartbeat status` show heartbeat state
+- [x] `/model` show current model for default chat agent session
+- [x] `/model <provider/modelId>` set model for default chat agent session
+- [x] `/model default` clear model for default chat agent session
+- [x] `/thinking` show thinking level for default chat agent session
+- [x] `/thinking <level>` set thinking level for default chat agent session
+- [x] `/thinking default` clear thinking level for default chat agent session
+- [x] `/model` inline button picker for configured model choices
+- [x] `/thinking` inline button picker
+- [x] Handle model picker callback and persist selected session model
+- [x] Handle thinking picker callback and persist selected session thinking level
+- [x] Add `Use default` model/thinking buttons
+- [ ] `/model <agentId> <provider/modelId>` set model for an explicit agent session
+- [ ] `/thinking <agentId> <level>` set thinking level for an explicit agent session
 - [x] `/help` show bot commands
 
 ## Tests
@@ -171,6 +200,10 @@
 - [ ] Unit test routing rules
 - [ ] Unit test session key generation
 - [ ] Unit test heartbeat `NOOP` behavior
+- [ ] Unit test model selection priority
+- [ ] Unit test Telegram `/model` command parsing
+- [ ] Unit test Telegram model picker callback parsing
+- [ ] Unit test Telegram thinking picker callback parsing
 - [ ] Repository tests against temp SQLite DB
 - [ ] Integration test fake Telegram update to fake Pi runner
 

package/agents/assistant/agent.yaml

@@ -1,6 +1,8 @@
 id: assistant
 name: Assistant
 profile: generic-assistant
+model: null
+thinkingLevel: null
 workspace: ~/.pigent/workspaces/assistant
 systemPromptFile: ./SYSTEM.md
 skills: []

package/drizzle/migrations/0001_session_model_overrides.sql

@@ -0,0 +1,2 @@
+ALTER TABLE `agent_sessions` ADD `model` text;--> statement-breakpoint
+ALTER TABLE `agent_sessions` ADD `thinking_level` text;