v0.7.0

Added

• User message creation timestamps are now rendered into live model input for chat, jobs, and sub-agents using a local format like [Monday, 2026-03-09T09:22:40-07:00]
• Stored messages now carry a nullable created_at column so new conversations preserve message creation time without rewriting legacy rows

Changed

• Time awareness now comes from per-message timestamp rendering instead of a request-only current-time system block
• The stable system prompt no longer carries runtime timezone guidance; runtime.timezone now drives cron matching, Slack timestamp formatting, and rendered user-message timestamps

Fixed

• Memory harvester and cleaner model calls now use the same fallback-chain behavior as agent requests, including support for models: and recovery onto later providers
• Memory worker and top-level agent failures now log concise single-line errors instead of noisy tracebacks for expected model/runtime failures

v0.6.2

Fixed

• OpenAI reasoning models with tool calls now route through LiteLLM's Responses bridge when reasoning_effort is enabled, avoiding the gpt-5.4 /v1/chat/completions BadRequestError that required /v1/responses.

Follow-up

• Tracking native Responses API support for this path in #23.

v0.6.1

Fixed

• Anthropic models no longer receive reasoning_effort="none" when thinking: off, avoiding the LiteLLM 1.82.0 AttributeError that previously broke every call with the default thinking setting.

v0.6.0

Added

• Guided onboarding via operator setup — picks a provider, detects timezone, stores credentials in .env, creates the first admin user, and can start the runtime immediately with --run
• python -m operator_ai entrypoint for environments where the operator script is not yet on PATH
• Simple reasoning control via thinking: off|low|medium|high at the defaults and per-agent level
• Memory retention classes (candidate and durable) plus memory.candidate_ttl_days for retention-aware recall and expiry

Changed

• Runtime/process settings now live under runtime in operator.yaml (timezone, env_file, show_usage, and reject_response)
• Current time is injected just in time for live model requests in chats, jobs, and sub-agents while the stable system prompt now carries only the timezone contract
• Jobs and sub-agents inherit the resolved agent thinking level, runtime prompt, and skill filter more consistently
• Memory recall is stricter about scope and retention — public contexts no longer list user-scoped memories, and private contexts only expose the current user's memories
• Legacy memory tables are reset on first run of this release to support retention/expiry metadata; existing memories and memory worker state are discarded
• Harvester and cleaner prompts are stricter about rejecting short-lived conversational noise and preserving retention metadata

Fixed

• run_skill now expands environment variable references in arguments after env sanitization, while still preserving $OPERATOR_HOME
• Cross-provider fallbacks drop reasoning metadata before each model call so retries stay compatible across providers
• Config loading now raises ConfigError internally instead of terminating via SystemExit
• Slack thread timestamps now render in the configured runtime timezone
• Same-agent spawn_agent(...) calls once again inherit the current agent's full prompt/runtime context even when no explicit agent= override is provided