Feynman changelog: what's new each month
Every stable Feynman release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 4 months covered; the current month updates daily.
November 2026 · so far
10 releases: 0.5.62 → 0.5.71
November brought stability improvements across parallel research workflows, security updates, and better handling of edge cases in sessions and subagents. The month focused on reliability, performance, and correctness in core research features.
- Cache warming default changed from "streaming" to "idle"; set cacheWarming to "streaming" to restore old behaviour and billing patterns.
Highlights
- Parallel researcher subagents no longer fail on output field validation errors.
- Claude sessions now work correctly after empty or reasoning-only replies.
- Streamed tool calls and code-heavy answers use significantly less CPU through optimized parsing and reuse.
- Resuming moved or deleted project folders now offers Fork to copy into the current folder.
- DuckDuckGo searches are now paced to avoid blocking after multiple searches in quick succession.
New
- Session resume dialog now offers Continue, Fork, or Cancel instead of only Continue when a project folder was moved or deleted.
- Paper search accepts citation-count sorting with sort: "citation".
- Cache warming now defaults to "idle" mode, keeping prompt caches warm for up to 30 minutes between runs.
Improved
- Tool arguments now automatically wrap single values into lists when a list is expected, instead of rejecting them.
- Long streamed tool calls and code-heavy answers parse arguments once per delta and reuse highlighting and rendered Markdown.
- Subagents with a model specified as provider/id now always run on that provider instead of silently falling back.
- PubMed and Europe PMC full-text results preserve inline text formatting including subscripts and superscripts.
- Subagent timeouts no longer run down while waiting for user replies.
- Side question answers in /btw now stream in an overlay without blocking main editor messages.
- Failed session writes no longer cause resumed sessions to lose earlier history.
- Anthropic default model updated to claude-opus-5-5.
Fixed
- Parallel researcher subagents no longer fail with unsupported output field errors.
- Claude sessions no longer fail after empty or reasoning-only replies.
- pi.exec() no longer crashes on very large command output or hangs on commands that ignore SIGTERM.
- Sessions no longer lose subagent tools after compaction when pi-subagents was reloaded or updated.
- Permission prompts and dialogs no longer hide behind the fleet inspector.
- Subagent replies cut off at the model's output limit now fail instead of passing as success.
October 2026
45 releases: 0.5.17 → 0.5.61
October brought major stability improvements, better error messaging, and support for Pi 1.x releases. The month focused on fixing login issues, handling rate limits gracefully, and making the setup experience smoother for new users.
- Azure OpenAI provider renamed from azure-openai-responses to azure; use feynman model login to update.
- /deepresearch no longer waits for approval after showing its plan; it runs immediately.
Highlights
- ChatGPT login, Git Bash on Windows, and API key validation now work reliably.
- Research steps recover automatically when paper indexes are rate-limited or down, switching to fallback sources.
- New users reach a working model on first launch instead of hitting dead ends.
- Web search and paper lookup tools are significantly smaller and faster, cutting token usage.
- Critical security fixes for MCP SDK, undici, and workflow sandboxing.
New
- GitHub Copilot users now get a research model by default.
- Paper lookups resolve by DOI, helping with recent arXiv papers.
- DeepSeek Search API, Ceramic, and three more web search providers available.
- feynman mcp adds and lists MCP servers.
- Custom User-Agent header for fetch_content to bypass AI-blocking sites.
- OpenAlex lookups explain why a name or title lookup failed.
- Web search and fetch_content work with models that sign in without an API key.
- Subagent tool definitions shrink by 82%, reducing token cost per request.
Improved
- Rate-limited paper sources fall back to alternatives instead of ending research.
- arXiv, PubMed, and OpenAlex requests respect rate limits and retry intelligently.
- Setup wizard now asks only for a model on first launch; packages and alphaXiv stay in feynman setup.
- PubMed titles, abstracts, and author affiliations preserve formatting and completeness.
- Paper index timeouts switch to alternate sources with explanatory notes.
- Failed model calls explain the cause and suggest fixes.
- Provider error hints appear once the run stops, not during retries.
- Web search summaries now read source snippets, working with providers that return only links.
Fixed
- Settings saved with UTF-8 BOM by Windows editors no longer block startup.
- LiteLLM and custom providers reading keys from environment variables work again.
- Word documents open without requiring LibreOffice.
- Parallel subagents no longer fail with workflow script errors.
- Tool calls containing backslash sequences like \usepackage no longer lose fields.
- Background subagents keep results across session restarts instead of showing "exited before writing a result".
September 2026
20 releases: 0.3.48 → 0.5.16
September brought major architectural changes as Feynman stabilised under new ownership, integrated more deeply with Pi, and expanded its science research capabilities across literature databases and computational biology tools.
- npm package moved from @companion-ai/feynman 0.3.47 to @advaitpaliwal/feynman 0.3.48 to @companion-ai/feynman 0.3.49+; migrate with npm uninstall -g @companion-ai/feynman && npm install -g @companion-ai/feynman.
- Feynman's own /thinking command removed; use Pi's built-in /thinking instead.
- arXiv free-text search removed from feynmansciencedatabasesearch; use arxiv: prefix with arXiv ID lookup only.
- Session-search optional package preset removed due to index incompatibility with Feynman's session location.
Highlights
- Feynman is now a cleaner Pi extension (v0.5.0+) that runs Pi's CLI unmodified instead of patching it, eliminating a sealed runtime workspace.
- Science workbench launched with feynman serve for local research, expanded literature search across OpenAlex, arXiv, PubMed, Europe PMC, bioRxiv/medRxiv, and Crossref.
- Protein and computational biology skills added with templates for AlphaFold2, ESMFold, DiffDock, ProteinMPNN, and other structure prediction and drug discovery tools.
- Research workflows now wait for background subagents before returning results, and telemetry now reports research runs without content.
- Repository and npm package returned to Companion-Inc/feynman and @companion-ai/feynman after a brief transition.
New
- Science workbench: feynman serve with local web app, demo data, and workbench-only Pi tools.
- Non-literature database search now covers OpenAlex, arXiv ID lookup, PubMed, Europe PMC full-text sections, bioRxiv/medRxiv, and Crossref.
- NVIDIA BioNeMo model endpoint support via feynmanmodelendpointcall tool.
- PaperRank command (feynman rank) for ranking and synthesising papers with options for citation expansion, full-text evaluation, and custom preferences.
- New commands: /commands, /capabilities, /watch, /jobs, and arxiv topic search in feynmansciencedatabasesearch.
- Pi session flags now pass through: --continue, --resume, --session, --fork, --no-session, and --export.
- feynman --mode rpc with --no-themes flag for use in ACP editors like Zed.
- Telemetry-visible research and workflow events via PostHog instead of OpenTelemetry spans.
Improved
- Settings are written once on first run; package entries from older releases are replaced with bundled versions.
- One-shot runs (feynman --prompt) now exit when they finish instead of staying open.
- Explicit prompts never wait on stdin, so scripts and CI runs do not hang.
- feynman --version answers immediately without verifying the bundled research runtime.
- Researcher and verifier subagents now run in the background by default with access to web and literature tools.
- System prompt and workflow prompts reduced by about 200 lines of duplicated tool-name rules.
- OpenAlex exact searches now respect the tool's limit (default 5, max 20) instead of returning 50-100 records.
- Error messages now name the failing service (for example "www.ebi.ac.uk request failed: 503") instead of generic messages.
Fixed
- macOS and Linux standalone bundles had nodemodules/.bin command links pointing at the build machine, now fixed.
- Windows patching of bundled Undici proxy runtime no longer fails with EPERM on non-elevated accounts due to symlink recreation.
- Crossref rate limiting now leaves margin over the pool's limit and properly retries with Retry-After header.
- Parallel Semantic Scholar and OpenAlex searches no longer fail with 429 Too Many Requests due to one-per-second limits.
- Dead network connections now fail in about 15 seconds instead of retrying silently for over five minutes.
- Extension error on every turn when Feynman runs inside another Pi setup no longer appears.
August 2026
30 releases: 0.3.17 → 0.3.46
August brought substantial improvements to research reliability, content handling, and provider support. The month focused on hardening cache operations, fixing edge cases in document parsing and tool execution, and expanding integrations with external services.
Highlights
- Research sessions now reliably handle concurrent requests to NCBI and other rate-limited services through paced queuing, preventing evidence loss from burst errors.
- Document parsing improved for edge cases including rotated pages, mixed fonts, table headers, and GPO operator stamps that were obscuring bill text.
- Research subagents now stop gracefully on context-window overflow instead of retrying oversized inputs, and properly report logical failures to the model.
- Browser cookie handling on Windows now supports modern Chrome and Edge formats with correct DPAPI decryption and large expiry values.
- Fetched content cache hardened with size and entry limits, symlink rejection, atomic file operations, and concurrent access tolerance.
New
- Response IDs now exposed in fetchcontent results, allowing direct passage lookup without guessing.
- Bocha added as a configured search route option via pi-web-access.
- Local current date now provided to parent research turns and researcher child sessions.
- Firecrawl search and authenticated browser fetch profiles now available as opt-in features.
- HTTP proxy support for routing search and fetch calls through explicit proxy servers.
- Bounded PDF extraction via pdf.maxPages setting, defaulting to 100 pages.
- Provider-specific search preferences configurable through openaiSearchProviders.
- Structured reasoning metadata now preserved through session persistence and model replay.
Improved
- Concurrent NCBI requests now share a single paced queue to prevent 429 errors; NCBIAPIKEY reaches E-utilities while remaining redacted from provenance.
- PubMed request timeouts now cover response body as well as initial headers.
- Inline data URI payloads replaced with bounded omission markers before reaching model context or cache.
- Pi auth.json and models-store.json files now retain administrator-managed modes and ACLs during updates.
- Research prompts beginning with dash now safely pass through end-of-options delimiters.
- Large rendered tool diffs no longer hit V8 argument-count limits.
- JPEG conversion and image resizing now correctly apply later EXIF orientation blocks after XMP metadata.
- Runtime repair extended to fifteen minutes for slow Windows and Node 25 clean installs.
Fixed
- Malformed custom agent definitions no longer block unrelated valid research agents from loading or running.
- Stopping a tool run now properly ends the active Pi loop before queued steering input can be drained into an aborted call.
- Research sessions with image context no longer incorrectly remain marked as having pending messages after delivery.
- Image-only steering and follow-up messages now removed from Pi's pending queue when delivered.
- JSONL session files lacking trailing newlines now repaired at the append boundary before new messages persist.
- Incomplete summaries from output-token limits no longer cause partial compaction checkpoints that fail research.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Feynman: pricing, features and alternatives · All changelogs