pi-schema: opt-in coercion, compile(schema, { coerce: true }) with coerce() and
parse(), the superset of TypeBox's Value.Convert rules and pi-ai's old coercion;
unresolvedRefs option; url and json-pointer-uri-fragment formats.
pi-ai: Tool takes any JSON Schema; validation through pi-schema; Type, Static
and TSchema are no longer exported (breaking). pi-agent-core and pi-durable:
tools on pi-schema, no TypeBox dependency. Coding agent: extensions keep Type,
Static, TSchema and StringEnum through an extension view of pi-ai; built-in
tools stay on TypeBox.
Replaces the WeakMap workaround. streamProxy and the agent and coding-agent test
mocks reimplemented AssistantMessageEventStream as plain EventStream subclasses;
they now use createAssistantMessageEventStream().
pi-ai's AssistantMessageEventStream sets AssistantMessage.durationMs with a
monotonic clock for every API implementation. The agent loop and the durable
tool task set ToolResultMessage.durationMs around execute(). Durable task
records carry startedAt/endedAt stamped by the Session. The coding agent's bash
renderer shows Took from the recorded duration, also after a reload.
closes#10549
This adds codemode and MCP support to pi. Codemode (packages/codemode) runs
model-written JavaScript in a QuickJS wasm VM inside a worker. From there the
script calls pi's tools as async functions, uses a per-session store, and can
read the model catalog and run classifiers. The codemode tool in coding-agent
is compatible with Codex's exec tool, and codemode.mode works like Codex's
tool modes (on or only). Large results are written to a JSON temp file. MCP
(packages/mcp) is a standalone client for stdio and streamable HTTP servers,
with OAuth.
The core only gets general mechanisms; codemode, tool_search and MCP are
built-in extensions that use them. Other extensions can replace these
built-ins. Tools declare an exposure: direct, model-only, codemode, deferred
or hidden. prepareLoadout() lets an orchestrating tool change which tool
declarations the model sees. ctx.executeTool() runs nested calls through
validation and the tool_call/tool_result hooks. Those calls emit events with
parentToolCallId and are recorded as bounded nestedCalls on the parent
result, which compaction and HTML export use. Tools can also return
structuredContent with an outputSchema, and can report errors with isError
instead of throwing. MCP servers come from mcp.json (global, or per project
after the project is trusted) or from pi.registerMcpServer(). They are
managed with /mcp and pi mcp list|login|logout, and their runtimes load on
first use.
Closes#10040
Make SessionManager projections authoritative for provider requests, add append-only context edits and actionable turn boundaries, and preserve existing queue scheduling during continuation and recovery.
This change makes system prompt text and tool changes part of the transcript rather than silently rewriting its starting conditions. This lets Pi record when instructions changed or tools became available, restore that state after resuming or navigating branches, and preserve cached prompt prefixes where the upstream supports it.
Run next-turn preparation only when the agent loop continues, allowing threshold compaction before the next provider request. Restore interactive progress when the same run resumes.
Closes#6879
Important to notice:
this changes when prepareNextTurn & prepareNextTurnWithContext are invoked:
the now run only when a next turn actually happens, and after shouldStopAfterTurn.
* add usage info to branch summary entries
* add usage to compaction entries
* allow custom tools to report llm usage in tool results
* allow observing and patching usage in tool_result hooks
* agent-harness: save usage in entries
for compaction, branch summaries and tool results
* fix(ai,agent): stop salvaging malformed tool-call argument JSON
Finalized tool-call arguments must strict-parse (allowing only lossless
string-escape repair). When the streamed JSON is truncated or malformed,
the raw JSON is preserved on ToolCall.malformedArguments, arguments
become {}, and the agent loop refuses to execute the call with an error
tool result instead of running it with silently salvaged partial
arguments. Partial parsing is now only used for streaming previews.
* fix(agent): fail tool calls from length-truncated assistant messages
Reverts the malformed-arguments tracking approach in favor of handling
truncation in the agent loop: when an assistant message stops with
"length", all tool calls in it are potentially borked (streamed
arguments are salvage-parsed), so none are executed. Each gets an error
tool result telling the model to re-issue the call, and the loop
continues. No new fields on ToolCall and no provider changes.
fixes#6284
* fix(typebox): migrate to v1 with extension compat
Replace AJV-based validation with TypeBox-native validation, keep legacy extension imports working (including @sinclair/typebox/compiler), and restore coercion for serialized/plain JSON schemas.
This change closes#3112.
* fix(typebox): use canonical imports and harden coercion
Switch first-party code to canonical typebox imports while retaining legacy extension aliases in the loader.
Remove obsolete runtime codegen guards, expand serialized JSON-schema coercion coverage, and update related tests and fixtures.
Fixes#3112.
---------
Co-authored-by: Mario Zechner <badlogicgames@gmail.com>
* feat(agent,coding-agent): add per-tool executionMode field to AgentTool and ToolDefinition
Add optional executionMode?: ToolExecutionMode to AgentTool and
ToolDefinition interfaces. Propagate through wrapToolDefinition and
createToolDefinitionFromAgentTool. No behavioral change yet — agent
loop will read this field in a follow-up.
* feat(agent): support per-tool executionMode override for sequential execution
When a tool defines executionMode='sequential', the agent loop
forces sequential execution of all tool calls in that batch,
even if the global config is parallel.
* feat(coding-agent): re-export ToolExecutionMode from @mariozechner/pi-agent-core
Makes the type available to extensions that want to set
executionMode on tool definitions.
* feat(coding-agent): add tic-tac-toe extension example with executionMode: sequential
Demonstrates per-tool executionMode: the agent plays via move/play
tool calls that share a cursor. Without sequential execution, play
can resolve before earlier moves finish, landing on the wrong cell.
Add AgentTool.prepareArguments and ToolDefinition.prepareArguments hook
that runs before schema validation in the agent loop. This lets tools
silently accept legacy argument shapes from resumed old sessions without
polluting the public schema.
The built-in edit tool uses this to fold legacy top-level oldText/newText
into edits[] when resuming sessions that predate the edits-only schema.
- AgentTool/ToolDefinition: typed prepareArguments returning Static<TParameters>
- agent-loop: prepareToolCallArguments() runs before validateToolArguments()
- edit tool: prepareEditArguments folds legacy fields, validateEditInput is strict
- Documented in extensions.md with edit-tool example
Breaking changes:
- Settings: 'hooks' and 'customTools' arrays replaced with 'extensions'
- CLI: '--hook' and '--tool' flags replaced with '--extension' / '-e'
- API: HookMessage renamed to CustomMessage, role 'hookMessage' to 'custom'
- API: FileSlashCommand renamed to PromptTemplate
- API: discoverSlashCommands() renamed to discoverPromptTemplates()
- Directories: commands/ renamed to prompts/ for prompt templates
Migration:
- Session version bumped to 3 (auto-migrates v2 sessions)
- Old 'hookMessage' role entries converted to 'custom'
Structural changes:
- src/core/hooks/ and src/core/custom-tools/ merged into src/core/extensions/
- src/core/slash-commands.ts renamed to src/core/prompt-templates.ts
- examples/hooks/ and examples/custom-tools/ merged into examples/extensions/
- docs/hooks.md and docs/custom-tools.md merged into docs/extensions.md
New test coverage:
- test/extensions-runner.test.ts (10 tests)
- test/extensions-discovery.test.ts (26 tests)
- test/prompt-templates.test.ts
Breaking change: replaces queueMessage() with two separate methods:
- steer(msg): interrupt mid-run, delivered after current tool execution
- followUp(msg): wait until agent finishes before delivery
Also renames:
- queueMode -> steeringMode/followUpMode
- getQueuedMessages -> getSteeringMessages/getFollowUpMessages
Refs #403
- 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
- 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
