Xean 4.1.0

xean 4.1.0 runs mathematical exploration and verification on native Pi Durable. It requires Bun 1.4.2 or later on macOS or Linux.

Version 4 is deliberately narrow. It focuses on reasoning over prepared notes and checking the resulting arguments. In the reported trials, that scope was sufficient for many tested problems, at recorded costs we found manageable.

xean provides citation checking, but no literature-discovery or paper-preparation workflow. To give Explorer a source packet, the user or another system must find relevant papers and convert their texts into notes with provenance, exact hypotheses, and conventions. Explorer has no literature-retrieval tool. Optional Codex search can retrieve cited results during source verification, without automatically importing papers. Caller attestations admit established results at their stated scope, and applications still require checking.

This release adds joint blind reconstruction of candidates and generated dependencies, atomic batch verdicts, and optional init --verify-on-solution. That mode defers all verification, including source search, until Coordinator nominates a claimed solution. Default verification timing remains available. Pi handles persistence and recovery.

Earlier ad1f082 TSP and First Proof P05 campaigns exercised generated dependency reconstruction and received fresh independent gpt-6-astra/high PASS assessments with their complete original source packets. These are bounded model reviews, not formal certification or qualification of every later runtime change. The specific P10 replay and broader source-reliability tests remain pending.

Long responses can disconnect near 15 minutes (#24). Recovery preserves recorded progress, but retained reasoning and failed-response usage remain incompletely measured. The former 3.5 implementation is preserved separately as Yean.

Download xean-4.1.0.tgz and verify it with SHA256SUMS. Follow the README and FAQ and use fresh campaign databases. Keep earlier campaigns with their recorded runtimes. The source archive includes the frozen lockfile, pinned Pi artifacts, patches, and licenses. Release checks and packaged checks each pass 113 tests and 1,066 assertions.

Xean 3.3.0

Xean 3.3.0 continues useful partial model output automatically and gives
Coordinator clearer Explorer handoffs.

• Responses ending with max_messages or max_output_tokens retain reusable
  text and reasoning and continue within the same submission. Continuations
  have no fixed count and preserve the role's completed-response allowance.
  Context limits still apply. Tools from incomplete responses never execute.
• Explorer retains validated partial submissions and reports when context or
  response limits force a handoff. Coordinator assignments identify relevant
  note IDs and the remaining gap in the exact task.
• Pi moves to source commit ee409a85f6ad92bf8c6ab8ed735c6cd24f2e749e.
  Native document retention replaces the custom cache cap. Unobserved document
  trackers leave memory after use and reload from storage when needed.
• The test audit consolidates duplicated fixtures, uses Pi's in-memory storage
  for cases that do not require persistence, and strengthens recovery and
  final-response read checks.

The database schema and public submission formats remain unchanged from 3.2.0.
Compatibility checks reopened and resumed representative paused and completed
model-free databases. Transfer of an in-flight model generation remains
unqualified. Keep existing campaigns with their frozen runtimes.

The release candidate passed 236 tests and 2,072 assertions on macOS ARM64 with
Bun 1.4.2. Packaging reran an 86-test subset with production dependencies and
passed the installation, CLI, export, and license checks. The rewritten suite
detected all 14 selected injected faults, including a final-response read error
the previous suite missed. Measured runtime line and function coverage was
unchanged. These scripted checks establish protocol behavior rather than
mathematical research quality.

Download xean-3.3.0.tgz and verify it with SHA256SUMS. The source archive
includes the frozen lockfile, pinned Pi artifacts, declared patches, and
licenses. Follow its README with Bun 1.4.2 or later.

Full changelog and
installation and qualification limits.

Xean 3.2.0

Xean 3.2 unifies note submissions and worker assignments, inspects campaign
databases in place, and fixes Codex implementation calls rejected by strict
output schemas.

Start fresh campaigns. Xean rejects 3.0.0 and 3.1.0 databases. Keep existing
campaigns with their frozen runtimes. This release includes no migration.

• Caller submit and Explorer share notes and edits. The correct command is
  removed, and candidate belongs to each note or edit instead of the batch.
• Explorer's guidance and Literature's query become assignment in work plans
  and standalone inputs. Inspection reports use work[].assignment instead of
  work[].guidance. Use role review instead of the dedicated review command,
  then inspect to read its assessment.
• status, inspect, and export use a read-only Pi Session and a consistent
  SQLite read transaction, eliminating full database copies. Pi rejects writes
  from inspection callbacks. The newer Pi snapshot bounds scheduler ownership metadata by
  live work.
• Codex implementation calls use Pi's strict-schema conversion and result
  validation. An optional native Codex profile disables plugins for research
  while preserving native web search. Implementation workers select their
  profile separately.

Qualification passed 213 tests and 2,044 assertions on each tested platform:
macOS ARM64 with Bun 1.4.3, and Linux ARM64 and x86-64 with Bun 1.4.2.
Source-distribution checks passed 66 Pi and role tests. One live Codex CLI implementation screen used
gpt-5.6-luna with low reasoning, independently executed the generated artifact,
and reopened the completed role without credentials or another call. This live
check covered a standalone implementation role. Earlier full-campaign
qualifications retain their original revisions in the README.

Download xean-3.2.0.tgz and verify it with SHA256SUMS. The source archive includes
the frozen lockfile, pinned Pi artifacts, declared patches, and licenses. Follow
its README with Bun 1.4.2 or later. Fleet, Nix, and private Xean Lab are not required.

Full changelog and
installation and qualification limits.

Xean 3.1.0

Xean 3.1 adds an editable notebook built on Pi Documents, stops repeating verification of unchanged inputs, and moves campaign inspection to the CLI.

Start fresh campaigns. The database format is incompatible with 3.0.0. Keep existing campaigns with their frozen runtimes. This release includes no migration.

• Explorer edits notes in place using stable IDs and expected revisions. New notes and edits publish atomically, and mathematical changes invalidate affected checks.
• Completed assessments, including INCONCLUSIVE results, stay closed for unchanged inputs. Coordinator chooses repairs, source retrieval, other work, or idle.
• After eligible terminal failures, Xean publishes validated Explorer submissions and completed Verifier checks with the failed status and original error.
• xean status returns compact JSON. xean inspect and xean export supply details on demand. The Observe web interface and the prompt-evaluation runner are removed.
• Completed roles release idle conversation context, document-tracker caching is bounded, and Xean reuses more native Pi APIs.

Qualification passed 203 scripted tests on fresh production installs on macOS ARM64, Linux ARM64, and Linux x86-64, plus 16 additional SQLite reliability cases. The x86-64 full suite passed on tmpfs. On Btrfs, the tests needed longer timeout allowances than the default. A fresh openai-codex/gpt-5.6-luna check with a saved Pi login passed note reading, an elementary campaign, and credential-free reopening. Six fresh gpt-6-astra/max golden campaigns ran through an OpenAI-compatible gateway using SSE. At the qualification snapshot, two had passed internal acceptance, independent model review, and credential-free reopening, and four remained running. These results are not formal proof certificates. Public OpenAI, the direct OpenAI subscription route, Anthropic, Google, fresh login, and OAuth refresh remain unqualified live.

Download xean-3.1.0.tgz and verify it with SHA256SUMS. The archive includes the frozen lockfile, pinned Pi artifacts, declared patches, licenses, and source. Follow its README with Bun 1.4.2 or later. Fleet, Nix, and private Xean Lab are not required.

Full changelog and installation and provider limits.

Xean 3.0.0

Xean 3.0 rebuilds the mathematical research workflow on Pi Durable. Pi owns tasks, conversations, checkpoints, cancellation, and storage. Xean supplies research policy and mathematical verification.

• Sequential research: Coordinator chooses one worker, waits for its committed result or failure, then decides again. Campaigns preserve completed work across interruptions.
• Shared mathematical memory: notes provide an index, detailed summaries, and full text for proofs, conjectures, questions, and failed approaches.
• Staged verification: checks cover correctness, declared support, sources, task requirements, and blind reconstruction across the generated dependency chain. Reconstruction batches retain completed progress. INCONCLUSIVE remains unresolved, and independent review is a separate result.
• Standalone use: the TypeScript library, CLI, and Observe run with Bun. Model profiles reuse Pi authentication. Codex provides optional source checking, literature, independent review, and implementation work.

This is a breaking replacement. APIs, settings, and campaign formats have changed. Start new campaigns with 3.0 and keep older campaigns with their original runtimes. Migration and historical-format readers are not included.

Download xean-3.0.0.tgz, extract it, and follow the README. The archive includes pinned Pi packages, declared patches, licenses, and the frozen lockfile. Ordinary users do not need Fleet, Nix, or private Xean Lab.

Model-based verification can make mistakes, and research outcomes vary. Live qualification covers the documented OpenAI gateway and saved OpenAI Codex login configurations. Fresh OAuth login, the direct OpenAI subscription route, public OpenAI, Anthropic, Google, and successful current-build ChatGPT Web generation remain unqualified live. The README records platform evidence and optional-provider limits.

Xean 2.0.0

Xean 2.0.0 provides a stable, versioned distribution of the Pi-based implementation: durable campaign execution, mathematical exploration and verification, a separate CLI, and a read-only observer.

Compatibility: This is a replacement for 1.x. Public APIs, CLI contracts, and campaign formats have changed. Start new campaigns with 2.0.0 and keep existing campaigns on their original release and frozen settings. No migration is provided.

The release fixes blocked recovery at the call cap, truncated CLI output, missing observer process status, misleading observer liveness text, duplicate browser-provider retries, Bun selection through package scripts, and the closed-book literature default. Observer status and available log streams survive independent read failures. Shutdown and provider cleanup use less bookkeeping. Coordinator instructions clarify how exploration starts with no existing notes. Documentation and public role-replacement contracts were audited.

Download xean-2.0.0-source.tar.gz and its SHA-256 checksum, unpack it, and run bun run setup with Bun 1.4.2. The archive includes pinned Pi packages and patches. Run bun run xean --help to begin. Workspace packages are not published separately to npm.

Qualification:

• Clean installation, CLI examples, deterministic execution and reopening on macOS ARM64 and Linux ARM64.
• 89 tests and 1,092 assertions on both platforms, plus typechecking, formatting, dependency provenance, and documentation links.
• Fresh gateway and full solver acceptance smokes, Claude Explorer/verifier checks, ChatGPT Web Explorer execution, and Codex primary-source retrieval with Astra/max. Completed reopening preserved records without new calls.
• Observer HTTP, DOM, console, desktop/mobile rendering, mathematical typesetting, and escaped content checks.

Provider authentication and retrieval depend on the executing environment. Custom Codex providers that implement standalone search must declare supports_standalone_web_search=true for Responses Lite models. Native source qualification used Codex 0.157.1 with that capability enabled. Browser usage remains unmeasured and browser requests are not retried automatically. Model judgments remain fallible. Mathematical acceptance, independent review, and catalog closure remain separate. Power-loss behavior and x86_64 were not newly qualified.

See CHANGELOG.md and README.md for the release and installation contracts.