Skip to content

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Releases]

v1.5.0 - July 15, 2026

Added

  • Added LeanResourceExceeded and LeanTimeout exceptions. The Lean worker exceeding its memory cap or time budget now raises a distinct, non-retryable exception instead of the generic AxleRuntimeError, so callers can record it as a deterministic outcome rather than retrying.
  • check, extract_decls, and extract_theorems now accept optional names and indices arguments (as theorem2sorry and similar tools already do) to select which declarations to process. The returned documents / messages are restricted to them, and elaboration is skipped for the proofs of unselected declarations. This is a speed feature and should be used when elaborating the whole file is too slow. Note that this changes the behavior of the tool: the returned Lean messages will be incomplete, and the per-document content field for extract_decls and extract_theorems is returned empty. See the check and extract_decls pages for more details. Regular users with no speed concerns can disregard these options.
  • extract_decls documents now include an unfolded_type_hash field: the type hash after unfolding module-local elaboration auxiliaries (e.g. foo.match_1), so types differing only by such an auto-generated name deduplicate where type_hash would not. See the extract_decls page for details. The behavior of the base type_hash field is unchanged.
  • verify_proof now accepts a verify_negation option (default false). When set, verify_proof additionally checks whether content proves the negation of formal_statement — i.e. whether it disproves the statement — and reports the result in a new negation field, which carries the same okay, tool_messages, and failed_declarations as the top-level result. The field is omitted unless verify_negation is set. Regular users can ignore this option.
  • disprove and extract_decls now accept a verbosity parameter (0=default, 1=robust, 2=extra robust), which affects the pretty-printed types output by the tools (the negated field in disprove, and the per-document type field in extract_decls). Higher verbosity levels make the pretty-printer more explicit, which helps when the default output re-elaborates ambiguously. This is the same behavior as existing tools with the verbosity parameter, e.g. have2lemma, sorry2lemma.
  • Added AxleClient.get_latest_environment(), which fetches the available Lean+Mathlib environments and returns the latest one.
  • Added instructions for citing AXLE in the documentation. See Citing AXLE.
  • Added Lean 4.32.0 support.

Fixed

  • Fixed a bug in tools that analyze term-mode goals (have2lemma, sorry2lemma, disprove) causing goal extraction to silently fail with include_whole_context=false. For example, this request failed in previous versions, leaving the content unchanged.
  • verify_proof is now module-aware. Previously, module dependencies would be treated as disallowed axioms when not using the default header. For example, this request was wrongfully rejected but now passes.

Changed

  • Various tools now skip proof elaboration for unselected declarations when names or indices is provided. This is a speed change; however, any outputs pertaining to the unselected declarations (e.g. Lean messages from their proofs) are unreliable and should not be used.

v1.4.0 - July 1, 2026

AXLE will be presented at the 3rd AI for Math Workshop at ICML 2026 as a contributed talk! Read the technical report on arXiv.

Changed

  • ignore_imports now defaults to true. When your code's imports don't match the environment's default header, AXLE substitutes the default header (reusing the cached environment) instead of raising an error. Setting ignore_imports=false no longer errors on a mismatch; instead AXLE processes your imports as written, which is significantly slower and may give incorrect results if a required dependency such as Mathlib.Tactic is missing (a warning is returned in these cases). See Import Mismatches for details.

  • Reworked the tool_messages and okay fields for a few tools. See Interpreting the okay field for details.

    • check now reports validation findings (sorry, disallowed axioms, unsafe definitions) as tool_messages warnings instead of errors. okay continues to reflect compilation only, and the offending declarations remain listed in failed_declarations.
    • repair_proofs now reports failed repairs (e.g. terminal tactics that fail to prove a sorry) as tool_messages errors instead of warnings, so okay is True only when the repaired code compiles and all repairs succeed.

  • Many tools now inspect values of opaques: opaque and partial def (and theorem in 4.30+).

    • merge now considers opaque bodies, not just types, when de-duplicating.
    • extract_decls now populates value fields for opaques.

  • In check and verify_proof, file-level validation errors now invalidate all contained declarations — e.g. use of open private results in every declaration being added to failed_declarations.

  • Reworded one of the verify_proof/check error messages to be more descriptive.

    • Before: Declaration '{name}' uses 'sorry' which is not allowed in a valid proof
    • After: Declaration '{name}' is incomplete (uses 'sorry' or has errors)

Added

  • Added Lean 4.30.0 and 4.31.0 support.
  • Added a relax_defeq_transparency repair pass to repair_proofs (on by default). Lean 4.29's backward.isDefEq.respectTransparency (default true) keeps isDefEq from unfolding reducible/instance definitions when unifying implicit arguments, breaking proofs that relied on it. Mathlib turns it off per-theorem. This repair prepends set_option backward.isDefEq.respectTransparency false in when the fix gets the proof further (all errors resolved, or the first error appears later in the source). On environments without the option, the repair is a no-op.
  • extract_decls and extract_theorems report four new per-declaration fields: type_depth, term_depth, wall_ms, and heartbeats. See the extract_decls page for more details.

Removed

  • Removed the http2 parameter from the AxleClient constructor, which was slowing the client down. The client now uses HTTP/1.1 exclusively (via aiohttp); the optional HTTP/2 transport and its httpx dependency have been dropped. Code that passed http2=... should remove that argument.

Fixed

  • Fixed extract_decls bug for opaques where value dependencies were misclassified as type dependencies.
  • Fixed disprove bug negating only the goal instead of negating the entire declaration type. The negated goal is now returned in a new negated field (a map from theorem name to negated goal) instead of being appended to each results message.
  • Browser-based clients (web apps, extensions, in-page demos) can call AXLE directly! The HTTP API now supports cross-origin (CORS) requests: OPTIONS preflights return 204 with Access-Control-Allow-* headers, and every /api/v1/ response carries Access-Control-Allow-Origin.
  • Fixed HTTP status codes on /api/v1/ endpoints: an unknown tool name now returns 404 (previously 200 with a user_error body), and methods other than POST return 405 with an Allow header (previously 422).
  • Fixed a rare bug resulting in lost executor slots. This bug used to cause a rare user-side failure or, more commonly, very high latency for some requests.

Thanks to @SSingh-07 on Github for submitting a few issues, which we have fixed in this update!!

v1.3.0 - June 3, 2026

Added

  • Added link shortening to the gateway. The web UI has been updated correspondingly. Try it out: https://axle.axiommath.ai/check#r=7d70453f-813f-4d19-8de9-44793dafa835
  • Added Claude web, desktop, and mobile support to the axiom-axle-mcp MCP server via a hosted endpoint at https://mcp.axiommath.ai/mcp. See the Quick Start for details. Thanks to Andrew Sutherland for suggestions on setting up this hosted instance.
  • Added three new fields to the info field of every response to identify the executor version your request was handled on: _executor_commit_sha, _executor_docker_image_id, and _executor_artifact_sha256.

Changed

  • Added a new option theorems_only (default true) to all tools that select over theorems/lemmas. These tools now have the ability to select over all declaration kinds: theorem2lemma, theorem2sorry, simplify_theorems, repair_proofs, have2lemma, have2sorry, sorry2lemma, disprove:

    • To use this feature, set theorems_only to false. For backwards compatibility (default), keep theorems_only set to true.
    • You can now sorry out any declaration body, simplify/repair any declaration containing a proof, and extract lemmas from any sorry locations and any have statement locations in any declarations, including definitions, opaques, instances, etc.
    • For theorem2lemma and disprove, the new setting is a no-op on non-theorem kinds.
    • Note that the value of theorems_only affects what the names and indices fields select over. When theorems_only is false, names and indices refer to all declarations, not just theorem kinds.

  • Reworked repair_proofs (the first of several planned changes):

    • Added two new passes to repair_proofs: remove_unknown_options, which strips unknown options both at the command-level and within proofs/terms, and enable_autoImplicit, which restores the autoImplicit option at the beginning of a theorem if an unknown identifier error occurs in a theorem's type signature.
    • Added command-level re-elaboration to repair_proofs, allowing repairs to stack (for example, when applying terminal tactics reveals another error to fix).
    • replace_unsafe_tactics now warns the user when replacing native_decide with decide +kernel fails. The tactic location is now left untouched.
    • apply_terminal_tactics now warns when no terminal tactics could be successfully applied at a given location in repair_proofs.
    • Fixed a bug in apply_terminal_tactics allowing malformed proofs with metavariables to be counted as successes in repair_proofs.

  • verify_proof now permits partial def and opaque. These checks were overly strict previously and do not raise soundness concerns.

  • merge now deduplicates other declaration kinds: axioms, opaques, inductives, classes, structures, etc. Previously, only theorems and definitions were eligible for deduplication.
  • extract_decls now names anonymous declarations (examples, anonymous instances) by their start position, line then column (e.g. _example_12_0), rather than a running counter (e.g. _example_0), so the placeholder is stable and remains unique across a file even when several share a line.
  • Added the merge_duplicates (default false) option to sorry2lemma, which merges extracted lemmas that are duplicates (either with other lemmas, or to the existing top-level theorem/lemma from which they are extracted) by definitional equality into a single lemma with all callsites pointing at it. Existing behavior can be retained with the default setting merge_duplicates=false.

Fixed

  • Added faster, more graceful retries on certain classes of connection errors. Minor change.

v1.2.1 - April 29, 2026

Deprecated

  • extract_theorems has been deprecated and will no longer be updated. Please use extract_decls instead, which supports all declaration kinds (def, theorem, lemma, abbrev, instance, structure, etc.).

Changed

  • The AXLE client now uses HTTP/2 by default. We don't expect any significant performance differences from this change, but feel free to file a bug report if this is not the case. Users may set the http2 parameter to false in the client constructor to revert back to the original HTTP/1.1 settings.

Added

  • Added a new option expand_scoped_notations to the normalize tool, which delaborates scoped notations into their expanded forms. See the normalize documentation page for details.

Fixed

  • Fixed a bug in the executors causing requests to hang, occasionally resulting in abnormally high latencies.

v1.2.0 - April 15, 2026

Added

  • Added two new fields in extract_theorems to be consistent with extract_decls (see below):
    • kind: always theorem for extract_theorems.
    • declaration_messages: same content as theorem_messages. theorem_messages is now deprecated and will be removed in a future update.
  • Added extract_decls, an upgraded version of extract_theorems that extracts all declaration kinds.
    • New kind field in each document. Possible values: theorem, def, abbrev, axiom, opaque, structure, class, class inductive, inductive, instance, example, unknown
    • Note: Not all fields are meaningful for all declaration kinds (e.g., proof_length/tactic_counts only apply to theorems/lemmas with tactic proofs.)
    • This tool should be used instead of extract_theorems as it is a strict superset of functionality. extract_theorems will be deprecated in a future update.

Fixed

  • Added "Last Used" and "Requests (24h)" columns to the API key console page for better visibility into API key usage.

v1.1.1 - April 8, 2026

Changed

  • [!] We are turning on the autoImplicit and turning off the pp.unicode.fun Lean options. AXLE will now automatically insert implicit variables when they are missing. This is a significant behavioral change, check your code! These settings are consistent with Lean's default. The previous options were remnants from internal use preferences.
  • [!] We have renamed mathlib_linter to mathlib_options, which now sets linter.mathlibStandardSet to true, autoImplicit to false, relaxedAutoImplicit to false, and pp.unicode.fun to true. Use this toggle to enable the stricter defaults that Mathlib uses by convention.

Added

  • Added Lean 4.29.0 support.
  • Added support for glob patterns in the permitted_sorries field for verify_proof. See the verify_proof documentation page under the permitted_sorries field for example use cases.

Fixed

  • Fixed a bug causing timeouts to be capped at 10 minutes. All requests now max out at 15 minutes (with documentation updated correspondingly).

v1.1.0 - April 1, 2026

Changed

  • [!] Removed document_messages from the response of extract_theorems — to replicate old behavior, run the content field of the resulting documents through the check tool. This change significantly improves the speed of extract_theorems.
  • [!] includeEndPos has been turned on for Lean messages. This changes the format from: -:4:38: error: Function expected at... to (when endPos is available): -:4:38-4:43: error: Function expected at... This change affects all tools with Lean messages.
  • Significantly reworked the Lean executor pool backend.
    • Latency has been decreased by 50% in most cases. For longer requests, the new executors can be more than 5 times faster!
    • Previously, the first request to each environment required a ~10s warmup. This is no longer the case, and so requests will be more faithful to their Lean timeout limits (not including queueing / waiting for available slots).
    • Eliminates a security risk involving persistent Lean workers.
  • Improved the Lean worker warm-up pipeline. Worker scale-up is also more aggressive than before. In the worst case, when all workers are completely occupied / offline, users should expect no more than a 2-3 minute delay before more worker capacity spins up.

Fixed

  • Removed redundant parsing resulting in occasional speedups in repair_proofs, normalize, etc. when content does not change.
  • Pruned missing executors from the gateway registry. Fixes a bug with autoscaling improperly triggering.

v1.0.2 - March 18, 2026

Added

  • Added explicit okay return value to repair_proofs

Changed

  • Improved error messages for unknown options in simplify_theorems, repair_proofs, normalize
  • Improved error messages for ignore_imports error (with links to relevant docs)
  • Improved the efficiency of merge, bringing down the time spent on large requests by 20-30%.

v1.0.1 - March 11, 2026

Added

Fixed

  • Increased request limits and fixed a typo in the documentation. Users with an API key are now limited to 20 active requests, and anonymous users are limited to 10 active requests.
  • Increased maximum timeout to 15 minutes (from 5 minutes).
  • Environments are now sorted by prefix (alphabetically) and then by version number (more recent versions first)
  • Fixed a bug with disprove failing to recognize implicit local variables. This bug was found by Bulhwi Cha on Lean Zulip.

v1.0.0 - March 4, 2026

Added

  • Initial release of AXLE Python client
  • Async client (AxleClient) with all 14 API tools:
    • verify_proof - Verify proofs against formal statements
    • check - Check Lean code for errors
    • extract_theorems - Extract theorems with dependencies
    • rename - Rename declarations
    • theorem2lemma - Convert theorem/lemma keywords
    • theorem2sorry - Replace proofs with sorry
    • merge - Combine multiple Lean files
    • simplify_theorems - Simplify proofs
    • repair_proofs - Repair broken proofs
    • have2lemma - Extract have statements to lemmas
    • have2sorry - Replace have statements with sorry
    • sorry2lemma - Extract sorries and errors to lemmas
    • disprove - Attempt to disprove theorems
    • normalize - Standardize formatting
  • CLI tool with commands for all tools
  • Helper functions for string manipulation
  • Configuration via environment variables
  • Type hints and PEP 561 compliance
  • Comprehensive documentation