- serializeConversation now takes Message[] (after convertToLlm)
- Handles all custom message types via convertToLlm
- Includes thinking blocks as [Assistant thinking]
- Removes truncation of tool args and results (already token-budgeted)
Instead of passing conversation as LLM messages (which makes the model
try to continue it), serialize to text wrapped in <conversation> tags.
- serializeConversation() formats messages as [User]/[Assistant]/[Tool result]
- Tool calls shown as function(args) format
- Tool results truncated to prevent bloat
- Conversation wrapped in <conversation> tags in the prompt
- SUMMARIZATION_SYSTEM_PROMPT explains the task clearly
- Tells model to NOT continue conversation, ONLY output summary
- Updated prompts to reference 'messages above' for clarity
- Pass systemPrompt to completeSimple
- CompactionDetails type with readFiles/modifiedFiles
- extractFileOperations collects from tool calls and previous compaction details
- UPDATE_SUMMARIZATION_PROMPT for merging with previous summary
- generateSummary now accepts previousSummary for iterative updates
- compact() extracts files, passes previousSummary, returns details
- Only merges from !fromHook compaction entries (backward compatible)
Prepends: 'The user explored a different conversation branch before
returning here. Summary of that exploration:'
This helps the LLM understand the summary is background context from
a different path, not the current thread being continued.
The old messagesToText only extracted text content, losing all tool calls.
Now uses convertToLlm like compaction does, preserving the full conversation
including tool calls, which is essential for generating useful summaries.
getHookPaths(), getCustomToolPaths(), and getSkillsSettings() now return
copies of arrays instead of references to internal state. This prevents
callers from accidentally mutating settings without going through setters.
Fixes#361
Both now use consistent sections:
- Goal
- Constraints & Preferences
- Progress (Done/In Progress/Blocked)
- Key Decisions
- Next Steps
- Critical Context (compaction only)
Prompts instruct LLM to use EXACT format and preserve file paths/function names.
- BranchPreparation now uses AgentMessage[] instead of custom type
- Reuse getMessageFromEntry pattern from compaction.ts
- Add BranchSummarySettings with reserveFraction to settings.json
- Add getBranchSummarySettings() to SettingsManager
- Use settings for reserveFraction instead of hardcoded value
- Add ReadonlySessionManager interface to session-manager.ts
- Re-export from hooks/index.ts
- Add collectEntriesForBranchSummary() to extract entries for summarization
- Don't stop at compaction boundaries (include their summaries as context)
- Add token budget support to prepareBranchEntries()
- Walk entries newest-to-oldest to prioritize recent context
- Use options object for generateBranchSummary()
- Handle compaction entries as context summaries
- Export new types: CollectEntriesResult, GenerateBranchSummaryOptions
- Move compaction.ts to src/core/compaction/compaction.ts
- Extract branch summarization to src/core/compaction/branch-summarization.ts
- Add index.ts to re-export all compaction utilities
- Update all imports across the codebase
- Add test/utilities.ts with shared helpers (API_KEY, userMsg, assistantMsg, createTestSession)
- Add agent-session-tree-navigation.test.ts with e2e tests for tree navigation
- Add getChildren() method to SessionManager
- Add summaryEntry to navigateTree return type
- Update existing tests to use shared utilities
- Check response.stopReason instead of catching errors for abort detection
- Return result object from _generateBranchSummary instead of throwing
- Fix summary attachment: attach to navigation target position, not old branch
- Support root-level summaries (parentId=null) in branchWithSummary
- Remove setTimeout hack for re-showing tree selector on abort
- Add TreeSelectorComponent with ASCII tree visualization
- Add AgentSession.navigateTree() for switching branches
- Add session_before_tree/session_tree hook events
- Add SessionManager.resetLeaf() for navigating to root
- Change leafId from string to string|null for consistency with parentId
- Support optional branch summarization when switching
- Update buildSessionContext() to handle null leafId
- Add /tree to slash commands in interactive mode
Major changes:
- Replace monolithic SessionEvent with reason discriminator with individual
event types: session_start, session_before_switch, session_switch,
session_before_new, session_new, session_before_branch, session_branch,
session_before_compact, session_compact, session_shutdown
- Each event has dedicated result type (SessionBeforeSwitchResult, etc.)
- HookHandler type now allows bare return statements (void in return type)
- HookAPI.on() has proper overloads for each event with correct typing
Additional fixes:
- AgentSession now always subscribes to agent in constructor (was only
subscribing when external subscribe() called, breaking internal handlers)
- Standardize on undefined over null throughout codebase
- HookUIContext methods return undefined instead of null
- SessionManager methods return undefined instead of null
- Simplify hook exports to 'export type * from types.js'
- Add detailed JSDoc for skipConversationRestore vs cancel
- Fix createBranchedSession to rebuild index in persist mode
- newSession() now returns the session file path
Updated all example hooks, tests, and emission sites to use new event types.
- agentLoop now accepts AgentMessage[] instead of single message
- agent.prompt() accepts AgentMessage | AgentMessage[]
- Emits message_start/end for each message in the array
- AgentSession.prompt() builds array with hook message + user message
- TUI now receives events for before_agent_start injected messages
Hook commands now always execute immediately, even during streaming.
If a command needs to interact with the LLM, it uses pi.sendMessage()
which handles queueing automatically.
This simplifies the API and eliminates the issue of queued slash
commands being sent to the LLM instead of executing.
- Add BeforeAgentStartEvent and BeforeAgentStartEventResult types
- Add emitBeforeAgentStart to HookRunner
- Call in AgentSession.prompt() before agent.prompt()
- Hook can return a message to inject into context (persisted + visible)
- Add test hook demonstrating custom message rendering and before_agent_start
- addMessageToChat: exhaustive switch for all AgentMessage roles
- renderSessionContext: delegates to addMessageToChat, special handling for assistant tool calls and tool results
- export-html formatMessage: exhaustive switch for all AgentMessage roles
- Removed isHookMessage, isBashExecutionMessage type guards in favor of role checks
- Fixed imports and removed unused getLatestCompactionEntry
Two bugs:
1. createCustomMessage was returning role: 'user' instead of preserving
the hook message structure (role: 'hookMessage', customType, etc.)
2. rebuildChatFromMessages wasn't clearing the container before rebuilding
- Removed Attachment from agent package (now in web-ui/coding-agent)
- Agent.prompt now takes (text, images?: ImageContent[])
- Removed transports from web-ui (duplicate of agent package)
- Updated coding-agent to use local message types
- Updated mom package for new agent API
Remaining: Fix AgentInterface.ts to compose UserMessageWithAttachments
- Renamed AppMessage to AgentMessage throughout
- New agent-loop.ts with AgentLoopContext, AgentLoopConfig
- Removed transport abstraction, Agent now takes streamFn directly
- Extracted streamProxy to proxy.ts utility
- Removed agent-loop from pi-ai (now in agent package)
- Updated consumers (coding-agent, mom) for AgentMessage rename
- Tests updated but some consumers still need migration
Known issues:
- AgentTool, AgentToolResult not exported from pi-ai
- Attachment not exported from pi-agent-core
- ProviderTransport removed but still referenced
- messageTransformer -> convertToLlm migration incomplete
- CustomMessages declaration merging not working properly
HookMessage.content can be string or array, but LLM Message.content
must be an array. This was causing 'messages: at least one message
is required' errors when hooks sent string content.
- Rename HookAppMessage to HookMessage, isHookAppMessage to isHookMessage
- Remove entries array from SessionContext (use isHookMessage type guard instead)
- HookMessage.content now accepts string directly (not just array)
- Fix streamMessage type in AgentState (AppMessage, not Message)
- Rename CustomMessageComponent to HookMessageComponent
- Fix test hook to use pi.sendMessage
- Change from contextTransform (runs once at agent start) to preprocessor
- preprocessor runs before EACH LLM call inside the agent loop
- ContextEvent now uses Message[] (pi-ai format) instead of AppMessage[]
- Deep copy handled by pi-ai preprocessor, not Agent
This enables:
- Pruning rules applied on every turn (not just agent start)
- /prune during long agent loop takes effect immediately
- Compaction can use same transforms (future work)
- Add contextTransform option to Agent (runs before messageTransformer)
- Deep copy messages before passing to contextTransform (modifications are ephemeral)
- Add ContextEvent and ContextEventResult types
- Add emitContext() to HookRunner (chains multiple handlers)
- Wire up in sdk.ts when creating Agent with hooks
Enables dynamic context pruning: hooks can modify messages sent to LLM
without changing session data. See discussion #330.
- Extract execCommand to src/core/exec.ts, shared by hooks and custom-tools
- Rename CustomMessageRenderer -> HookMessageRenderer
- Rename registerCustomMessageRenderer -> registerMessageRenderer
- Renderer now receives HookMessage instead of CustomMessageEntry
- CustomMessageComponent now has setExpanded() and responds to Ctrl+E toggle
- Re-export ExecOptions/ExecResult from exec.ts for backward compatibility
- compaction-hooks-example.test.ts: get sessionManager/modelRegistry from ctx
- compaction-hooks.test.ts:
- Pass sessionManager/modelRegistry to HookRunner constructor
- Remove setSessionFile call
- Update tests to use session.sessionManager instead of event.sessionManager
Instead of using continue() which validates roles, prompt() now accepts
an AppMessage directly. This allows hook messages with role: 'hookMessage'
to trigger proper agent loop with message events.
- Add overloads: prompt(AppMessage) and prompt(string, attachments?)
- sendHookMessage uses prompt(appMessage) instead of appendMessage+continue
Following the same pattern as BashExecutionMessage:
- HookAppMessage has role: 'hookMessage' with customType, content, display, details
- isHookAppMessage() type guard for checking message type
- messageTransformer converts to user message for LLM context
- TUI checks isHookAppMessage() for rendering as CustomMessageComponent
This makes the API clean for anyone building on AgentSession - they can
use the type guard instead of knowing about internal marker fields.
