v0.0.11

Added

• Context compaction (summarization) for mini_swe_agent_v2 — long-running agents no longer exhaust the model context window. When the previous LLM response's prompt_tokens reaches compaction_token_trigger (default 28000), the agent calls the model a second time (without tools) to summarize old turns and replaces the live history with [system, task, summary] + recent_turns (last compaction_keep_recent_turns=2 assistant turns kept verbatim). Repeated compactions naturally fold the previous summary back into old_turns. Configurable via agent.compaction_enabled / compaction_token_trigger / compaction_keep_recent_turns / compaction_summary_prompt; enabled by default.
• Full-trajectory artifact — when any compaction occurs, the adapter writes {log_dir}/{agent_id}_full_traj.json containing a segments list (alternating solver / summarizer blocks) so the unabridged pre-compaction history is preserved for analysis even though the live messages list has been shortened.
• LitellmModel.summarize_context(messages, summary_prompt) — separate completion call (no tools) used by the agent's compaction step; tracks cost via GLOBAL_MODEL_STATS and tags the resulting message with extra.summary=True.

Changed

• mini_swe_agent_v2 adapter config merge — switched from shallow dict.update to recursive_merge, so partial overrides (e.g. only agent.compaction_enabled) no longer clobber sibling keys from the default YAML.