Feynman Files

v0.3.44 - 2026-08-26

Research continuity

• Stopping a tool run now ends the active Pi loop before queued steering or follow-up research input can be drained into an already-aborted model call. The queued input remains available for the next turn instead of producing a second spurious cancellation.

Compaction integrity

• Small-context local and proxy models now bound compaction reserve and retained-history budgets to the model's actual context window, preventing empty or short sessions from compacting continuously.
• Empty or structurally unusable summaries no longer replace research history. OpenAI Responses providers such as Grok also omit tool_choice when a compaction request has no tools.

Model reliability

• OpenAI-compatible structured reasoning deltas are accumulated without reparsing and reserializing the complete history for every streamed detail, preventing long reasoning streams from blocking the event loop while preserving same-model replay.
• OpenAI models reached through Amazon Bedrock now receive images returned by research tools as sibling user-image blocks instead of unsupported images nested inside toolResult.content. Text stays attached to its tool result, image-only results retain an explicit placeholder, and Anthropic Bedrock models keep their native nested-image shape.

Runtime reliability

• Non-interactive Pi extension handler work now has a cumulative 30-second budget. A handler that never settles reports its extension and event, expires its local context, safely absorbs late settlement, and lets later handlers and the research session continue.
• Timed-out model-tool and user-shell policy handlers still allow later policy handlers to run, then fail closed. User-shell interception returns an explicit non-executing result, so both TUI and RPC paths do not run the command after a policy timeout.
• Project trust, OAuth, and interactive dialogs preserve their supported behavior: documented UI prompts pause the remaining handler budget instead of resetting it, parent cancellation settles even a non-cooperative dialog, and OAuth callbacks remain outside the bounded event runner.

Document research

• Updated the bundled LiteParse runtime to 2.14.0. OCR rasterization now runs in bounded worker-sized rounds, and dense PDF layout deduplication avoids quadratic work. Existing parse, search, and screenshot interfaces are unchanged.

Validation

• Added executable abort-queue, Responses payload, structured-reasoning scale, small-context compaction, and summary-usability regressions across the maintained Pi runtime.
• Added executable extension-handler deadline coverage for downstream progress, timeout reporting, expired contexts, late rejection handling, fail-closed TUI/RPC user-shell interception, cumulative dialog timing, parent cancellation, project trust, and tool permission dialogs.
• Added exact source, root/nested runtime, package-tree, and runtime-archive verifier coverage for Bedrock tool-result images across bare, regional, and global OpenAI model IDs, with an Anthropic control.
• Re-ran installed document parsing, page-count, search, screenshot, package-artifact, runtime, and clean-consumer verification against LiteParse 2.14.0.