CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Added

• Worker isolation mode: CONDUCTOR_WORKER_ISOLATION=thread runs every worker as a thread instead of a multiprocessing.Process (default process is unchanged — spawn remains the start method). For environments where multiprocessing's fork+exec bootstraps (spawn children, resource_tracker) fail, e.g. Firecracker microVM guests. Thread-mode tradeoffs: no per-worker force-kill (shutdown is cooperative), CPU-bound workers share the GIL, and signal.signal becomes a no-op off the main thread. Implementation: the Windows-only Process→Thread shim moved to conductor.client.automator.worker_isolation (the private worker_manager._patch_conductor_use_threads_on_windows helper is removed — the Windows gate calls apply_thread_isolation() directly) and now also swaps the logging-relay Queue for a plain queue.Queue

• Canonical metrics mode: opt-in harmonized metric surface via WORKER_CANONICAL_METRICS=true -- details

• MetricsSettings gains clean_directory and clean_dead_pids for opt-in stale .db file cleanup (both default to False)

• SchedulerClient now carries the schedule lifecycle operations itself: pause(reason=), resume, delete, run_now, preview_next, reconcile (declarative tri-state sync) — with typed errors (ScheduleNotFound, InvalidCronExpression, ...). pause_schedule gains an optional reason= (stored by OSS Conductor servers; ignored by Orkes servers)

Changed

• Multiprocessing start method now defaults to spawn on all platforms (was fork everywhere except Windows). fork caused silently-restarting SIGSEGV worker subprocesses on macOS (exitcode -11) and fork-with-held-lock deadlocks on POSIX.
• Legacy metrics emit unchanged by default; no env var required
• metrics_collector.py is now a compatibility shim; from conductor.client.telemetry.metrics_collector import MetricsCollector continues to work
• get_schedule returns a typed WorkflowSchedule (or None for missing schedules) instead of a raw camelCase dict, matching its declared annotation and docs/SCHEDULE.md; dict-consumers should switch to attribute access or to_dict()

Fixed

• @worker_task workers are now picklable, making the decorator path work with the spawn/forkserver start methods (fixes TypeError: cannot pickle '_thread.lock' object and PicklingError: it's not the same object as ...; issues #264, #271): Worker.api_client is created lazily in the worker process, runtime state (locks, pending async tasks, background loop) is excluded from pickling and rebuilt in the child, and decorated functions are pickled as importable references resolved in the child
• TaskHandler.start_processes() no longer hangs the interpreter when a worker fails to start (e.g., unpicklable state under spawn); it now cleans up already-started subprocesses and raises with actionable guidance
• Worker processes killed by a signal now log a diagnostic hint (signal number, PYTHONFAULTHANDLER=1 guidance) instead of restarting silently
• Per-schedule pause/resume now work on both Conductor server families: the client sends PUT (the OSS Conductor dialect — the spec-generated GET failed there) and transparently falls back to GET on a 405 for Orkes servers
• Agent(model="claude-code/...") registration no longer crashes with SpawnSafetyError/PicklingError under the spawn start method (top-level or nested as a sub-agent); the tool worker is now built through the same picklable passthrough path already used by other frameworks
• AgentResult.tool_calls no longer drops or misnames tool calls. A tool task was identified by a call_ reference-name prefix, which is the LLM provider's tool-call id format — so an Anthropic-backed agent (toolu_…) recorded no tool calls at all. Tools are now identified by the task type they compiled to, and named from _agent_tool_name (set by the server on every tool kind), method, or the task definition. Consequences: http, api, mcp, agent_tool, human and pull_workflow_messages tools now appear at all; media and RAG tools report their own name instead of generate_image/llm_index_text; and a camelCase worker name is no longer lowercased, so assert_tool_used(result, "getWeather") matches
• AgentResult.events is now populated by run() and run_async() for natively-compiled agents, derived from the same execution already fetched for tool_calls rather than from an extra call. It was previously only filled in when an on_event callback or stream() was used, so the eight conductor.ai.agents.testing assertions that read events — assert_no_errors, assert_handoff_to, assert_events_contain, assert_max_turns and the rest — saw an empty list. Four of them check for absence and so passed vacuously, assert_no_errors on every eval case (expect_no_errors defaults to True); expect_handoff_to, used by the eval runner's own documented example, failed every time. Framework agents (LangChain, LangGraph, OpenAI Agents, Claude Agent SDK) also derive both fields from their execution's task list now: the whole agent runs inside one passthrough task, but the Claude Agent SDK injects a task per tool call, so those tool calls appear. LangChain and LangGraph publish their steps to the event stream rather than as tasks, so that detail still reaches on_event and stream() consumers only
• Tool arguments reported on AgentEvent.args and in AgentResult.tool_calls now agree, and both strip every key Conductor injects — previously _agent_tool_name, _allowed_commands, __humanTaskDefinition and __conductor_agent_ctx__ leaked into one surface or the other