Repository navigation
PyPI: https://pypi.org/project/shinka-evolve/0.0.8/
Added
- Added configurable tournament parent selection in PR #217, with
tournament_top_k=8andtournament_num_draws=3defaults. Draw distinct
contestants uniformly from the selected island's highest-scoring correct,
finite-scored archive entries and mutate a highest scorer, breaking winner
ties uniformly. Empty eligible archives fall back to the island's eligible
population; pool and draw sizes shrink to available candidates. The default
parent selection strategy remainsweighted. - Expanded the bundled model catalog with OpenAI GPT-6, Claude 5.5/Fable 5.1,
Gemini 4 Argon, Gemini 3.5 Flash-Lite and 3.6–3.8 Flash, and DeepSeek V4 Flash
and V4 Pro. Added provider-specific reasoning and thinking controls, refreshed
pricing, and historical-price overlays for upstream-removed IDs in PR #214.
Also added Claude Fable 5.2 ($8 input / $40 output per million tokens) and
Gemini 4 Hydrogen ($1.50 / $7.50), with adaptive Claude thinking and Gemini
thinking-level controls.
Gemini 3.7 Flash and Gemini 4 Argon use introductory 2026 pricing. Existing
default model pools remain unchanged. Thanks @anantgar for Gemini 3.6 support. - Added optional Weights & Biases logging with resumable run IDs, per-individual
scores, timing and cost metrics, and final run summaries in PR #149. PR #177
adds monotonic population/island snapshots and a secured metadata boundary.
Thanks @anantgar. - Added Verilog/SystemVerilog as a first-class evolution target in PR #137,
including.svtask detection, syntax validation, EVOLVE-BLOCK marker support,
failure-artifact rendering, and a self-contained RTLLM example. Thanks @Tyronita. - Added startup pricing refresh from models.dev with conditional HTTP caching,
validated offline fallback, provider-qualified model discovery, and per-run
pricing snapshots in PR #141. The tutorial now includes a no-cost catalog
preflight. Thanks @rodmarkun. - Added evaluation output controls:
JobConfig.eval_verbosesuppresses framework
progress output, andDatabaseConfig.max_stdout_log_charslimits persisted
stdout_logmetadata to its tail while keeping full logs on disk. Quiet mode
(run_shinka_eval(..., verbose=False)) also suppresses result-save framework
output in PR #140. Thanks @marcopirazzini. - Added SLURM support for
numeric_threads_per_job, applying numeric-library
thread caps consistently across local and SLURM evaluation jobs in PR #140.
Thanks @marcopirazzini.
Changed
- Reused provider SDK clients and their HTTP connection pools across compatible
requests, with explicit credential/configuration isolation, process-safe sync
caching, event-loop-scoped async caching, and cache shutdown helpers in
PR #164. - Bounded stalled OpenAI-compatible endpoints in PR #144 (issue #143): request
timeout, retry count, and backoff budgets are configurable through
SHINKA_LLM_TIMEOUT,SHINKA_LLM_MAX_RETRIES, andSHINKA_LLM_BACKOFF_*
environment variables. OpenAI SDK retries now default to0
(SHINKA_OPENAI_MAX_RETRIES) so they no longer stack under ShinkaEvolve's own
retry layers. Thanks @dexhunter. - Reduced Python complexity-analysis overhead by parsing each candidate AST once
while preserving the existing metrics contract in PR #175. Thanks @dexhunter. - Relaxed the exact HTTPX dependency pin to a
>=0.27compatibility lower
bound in PR #181 (issue #180). Thanks @htmai-880. - Hardened the
@claudeagent workflow in PR #195: only owners, members, and
collaborators can trigger it, checkout credentials are not persisted, web
tools are denied, and actions are pinned to full commit SHAs.
Thanks @kazuki-hanai. - Removed the automatic Claude Code Review pull-request workflow.
Fixed
- Fixed RTLLM equivalence labels in PR #218: the default 16-cycle SAT check
reportsbounded; opt-inRTLLM_PROVE=1requires induction proofs with shared
and real clock semantics before reportingformal. Added solver-output
spoofing protection and compatibility with older Yosys versions. Thanks @prath47. - Fixed
AsymmetricUCBandThompsonSamplerwithasymmetric_scaling=False:
exponential_basedefaults to1.0, which initialised the reward accumulator
in log-space whileupdate()accumulated it linearly, so every arm stayed at
-inf. ForAsymmetricUCBmodel selection collapsed to uniform; for
ThompsonSamplerthe posterior kept learning throughalpha/beta, but the
adaptive reward range collapsed after the first decay. Exponential scaling now
only activates together with asymmetric scaling, and abandit_state.pkl
saved by an affected run is reset to the prior with a warning on load, since
its reward history cannot be recovered, in PR #191. Thanks @yurekami. - Fixed prompt evolution resume inserting another seed when only generation-zero
prompts exist, retaining stored prompt IDs, fitness history, and costs, in
PR #216. - Fixed custom prompt-model sampler settings being forwarded as raw provider
arguments, allowing configured token limits and reasoning settings to work
for both diff and full prompt mutations, in PR #216. - Fixed billed empty OpenAI replies being omitted from budgets across single,
batch, proposal, novelty, prompt, and meta-analysis calls in PR #211
(issue #210). Output exhaustion stops proposal retries; failed charges survive
resume, and recovered replies retain complete billed usage. Thanks @prath47. - Fixed sync and async meta-summary metadata alignment and generation ordering
after failed batch responses in PR #208. Thanks @lll-hhh. - Fixed duplicated API, embedding, novelty, and meta-analysis costs on initial
and spawned island copies in PR #215, preserving original spend and cumulative
cost plots. Thanks @guudewie. - Fixed population selection: archive inspirations stay unique across fallback
selection (PR #202), multi-criteria ranking ties cannot let worse newcomers
displace better programs (PR #199), and island migration selects all migrants
before moving programs to preserve quotas and elites (PR #200).
Thanks @flyingdoubleG. - Fixed chart-to-program selection and cumulative-max step timing in PR #201.
Thanks @guudewie. - Fixed proposal metadata and LLM request settings diverging by reusing each
proposal's sampled kwargs in PR #213. Thanks @yuiandyuki. - Fixed evaluated-program imports for dataclasses with postponed annotations
and pickling program-defined result types in PR #212. Thanks @yuiandyuki. - Fixed circle-packing validation and scoring: invalid packings retain raw data
and diagnostics while receiving the correct score in PR #205 (issue #197),
and the notebook validator rejects negative radii and non-finite values in
PR #209 (issue #203). Thanks @Yushu-Li, @iamgahr, @lll-hhh, and @maximusrufus. - Local OpenAI-compatible endpoints fail without proposal retries on HTTP
401/403/404, with a typed endpoint failure preserved infailure.json, in
PR #204 (issue #196). Thanks @ToolongTD and @iamgahr. - Fixed run completion reporting:
shinka_runexits 1 when evolution produces
no evolved program; seed-only runs and partial runs with evaluated candidates
retain exit 0. The final database summary separates generation progress, seeds,
island copies, and evolved programs, and labels row correctness as including
copies. - Fixed evaluation timeouts to start when evaluation begins, so slow LLM proposals
no longer consumeLocalJobConfig.timeand cause immediate candidate termination
in PR #198. Thanks @fortunto2. - Fixed patch application and code extraction: shifted SEARCH/REPLACE patches
preserve relative indentation and dedents (PR #190; thanks @yurekami),
extract_betweenmatches fence tags literally to avoid corrupting C++ rewrites
or raisingre.erroron Python 3.10 (PR #193; thanks @MaxFreedomPollard), and
empty SEARCH insertions preserve line boundaries and EVOLVE-BLOCK-END markers
for consecutive patches and block-comment languages such as Wolfram
(PR #185, issue #183; thanks @Atharva-Kanherkar). - Fixed
ThompsonSampler.decay()to avoid invalid-value warnings while arms
remain unsampled in PR #192. Thanks @yurekami. - Fixed provider response parsing: Anthropic dispatches by content-block type to
avoid crashes and truncation for redacted, thinking-only, and multi-block
responses (PR #188; thanks @Atharva-Kanherkar). OpenAI Responses finds assistant
text by content type without mistaking reasoning or tool text for an answer
(PR #142). - Fixed pricing and usage accounting: Anthropic and DeepSeek sync/async queries
agree on token costs, separate thinking tokens, and safely fall back to zero
cost for models missing from the catalog. Runtime models.dev refreshes preserve
pinned embedding-price overrides, and CSV reasoning flags handle whitespace
and mixed value types, in PR #166. - Fixed Gemini retry handling so unsupported structured-output requests fail
immediately, synchronous retries are paced, and sync/async calls share the
same default thinking budget, in PR #163. - Fixed beam-search parent sampling to stay within the selected island and
retain its active parent across samples in PR #194. Preserved zero-score
ranking, metric/timestamp tie-breaks, and stored program fields on beam
rollover. Thanks @tmkz4. - Fixed Rust candidate validation on stable toolchains by replacing the
nightly-only parser flag with isolated, non-mutatingrustfmtsyntax checks
in PR #179. Thanks @Atharva-Kanherkar. - Fixed Azure OpenAI LLM client construction to use the v1 base URL contract,
avoiding doubled/openai/v1/openairequest paths and 404 responses reported
in PR #148 (issue #147). - Fixed OpenAI retry handling so ShinkaEvolve respects provider
retry_after
hints from transient Cloudflare/API 5xx responses in PR #142. - Fixed provider request compatibility: omitted unsupported
temperaturefor
DeepSeek V4 thinking calls, Claude Opus 4.8 on Anthropic and Bedrock, and GPT-5
reasoning models on the Responses API. DeepSeek V4 receives thinking-mode
controls andreasoning_effort, and its retry logs include model kwargs. - Fixed Google GenAI client setup to detect broken IPv6 connectivity to Google
API hosts and prefer IPv4 when needed in PR #142. - Fixed spelling and grammar errors in the bundled prompts in PR #138.
Thanks @anantgar.