KUMA 0.3.4

This optional patch improves SDK resilience and evidence boundaries.

• Keep file snapshots within the tracked repository across Windows junctions and file replacement/growth races.
• Detect modern KUMA credentials before evidence egress.
• Preserve complete UTF-8 text when captured logs reach a byte limit.
• Reject malformed persisted operations, deeply nested response JSON and overflowing timeout values with safe SDK errors.
• Use English labels for invalid Agent Profile sections.

Case difficulty defaults, Judge wire contracts and GET-only operation recovery remain unchanged. Explicit external storage directories remain supported. No automatic installation or additional paid retries are introduced.

Install or update:

python -m pip install --upgrade "kuma-defuzex==0.3.4"

The release also includes the previously completed optional storage_path feature: evaluation artifacts and recovery ledgers can use an explicit external directory while file tracking stays rooted in repo_path. Omitted storage_path preserves existing locations. Path separation does not provide sandbox isolation. Concurrent ledger initialization is serialized.

KUMA 0.3.3

Optional patch release.

• Improve compatibility with bounded optional Judge capability advertisements while keeping Evidence validation strict.
• Check built-in Official Case/Judge configuration before Case generation.
• Add difficulty D0/D1/D2: zero problems, one obvious low-intensity problem, or two subtler/composed problems requiring stronger recognition, recovery and verification. D1 remains the default and preserves previous request identity. max_steps remains an upper bound; difficulty does not change Judge severity or promise a measured failure rate.
• Clarify container upgrades and installation environment checks.

Custom providers and saved Cases remain unchanged. No automatic upgrade or paid retry is introduced. Install once PyPI publishing completes with: python -m pip install --upgrade kuma-defuzex==0.3.3

Source: PR #119. Full notes: https://github.com/DefuzeX-AI/KUMA-DefuzeX/blob/v0.3.3/docs/releases.md

KUMA 0.3.2 — Evidence completeness and safe retries

Optional patch release for KUMA 0.3.x.

• Fix unified diffs without a final newline and exact per-file patch binding.
• Preserve Trace and response Evidence under limits; negotiated file-observation summaries expose retained/omitted facts without paths or contents.
• Safely omit sensitive filenames, improve bounded credential-prose and Authorization-example handling, and provide safe local projection diagnostics.
• Allow a new explicit Judge attempt after a validated retryable terminal failure, while preserving GET-only and lost-response recovery. No automatic paid retry.

File completeness requires a supporting service. Known file loss or incomplete capture fails before POST on unsupported services; historical unknown records retain compatibility with an explicit local warning. Original files are never modified. Credential detection is not a guarantee against unknown secrets.

Install the package once the trusted publishing workflow completes:

python -m pip install --upgrade kuma-defuzex==0.3.2

Python imports and CLI remain kuma. Requires Python 3.10+.

Full behavior and compatibility notes: https://github.com/DefuzeX-AI/KUMA-DefuzeX/blob/v0.3.2/docs/releases.md

Fixes #98, #108, #109, #111, #112, #113, #114 and #115 through #117.

KUMA 0.3.1

KUMA 0.3.1 is an optional, backward-compatible patch for 0.3.0 users.

What's new

• Optional detailed Judge assessment separates task completion, artifact quality and behavioral integrity, with explicit claim evidence and uncertainty.
• collect_public_messages() captures bounded completed public replies, including earlier messages, with redaction and explicit capture gaps.
• Run.submit(public_messages=..., runtime_actors=...) lets callers declare the actors of actual captured OTel spans. Setup/reviewer spans do not inherit the target Agent's identity; declarations are not attestations.
• Assessment expectations and exact Evidence references survive retry/recovery. Known operations remain GET-only; lost responses reuse the same request identity.
• Corrected stale documentation about the already released 0.3.0 APIs while preserving the streamlined bilingual homepage.

Compatibility and limits

Detailed assessment requires explicit service support. Ordinary legacy requests retain their existing behavior; explicitly supplied messages fail clearly on unsupported services instead of being discarded. Existing clients do not need a forced update. This release does not automatically instrument Agents, upload observations, promote capture completeness, collect reasoning or guarantee model accuracy.

Install

python -m pip install --upgrade kuma-defuzex==0.3.1

For OpenTelemetry capture, install kuma-defuzex[otel]==0.3.1. Python 3.10 or newer is required. PyPI publication is performed by the release workflow; if the version is not immediately visible, wait for that workflow to complete before retrying installation.

See the assessment guide and release notes for the exact API, limits and retry behavior.

KUMA 0.3.0

KUMA 0.3.0

Install or upgrade

python -m pip install --upgrade kuma-defuzex==0.3.0

Requires Python 3.10 or newer. Install in your Agent's environment and restart the Agent after upgrading. PyPI availability follows completion of the release publishing workflow.

What's new

• Local observe() sessions with detached redacted export, readable timelines and explicit atomic saving; no Case, account, Judge or automatic upload required.
• Explicit authenticated cloud observation storage, separate from local capture and evaluation.
• Bounded OpenInference model/tool projections and pinned, real offline OpenAI, LangChain and LangGraph examples.
• External Run/invocation labels, validated server receipts and measured client-stage timings; unavailable server/model timing remains unknown.
• Successful identical Judge requests within the same Run reuse the original operation instead of starting another evaluation.

Scope and compatibility

Cloud storage and official correlation require matching server contracts, permissions and advertised capabilities. Installing the SDK does not activate server features. Capture remains bounded and depends on instrumentation; framework support is limited to the documented matrix. The 0.2.x to 0.3.0 upgrade reminder never automatically installs software or blocks an Agent.

Local examples and disposable three-layer integration were verified without paid model calls. A positive production cloud-observation lifecycle with a user credential was not exercised.

See the observation guide, framework examples, and Run correlation guide.

KUMA 0.2.8

Bounded operation polling

• Reduces repetitive status requests using geometric fallback backoff capped at 8 seconds. Valid server poll_after_ms revisions remain authoritative.
• Reserves a small positive budget for one final useful poll without extending the total operation deadline or busy-spinning. Network latency and scheduling can still exhaust that budget.
• Timeout preserves the original operation and idempotency identities for GET-only recovery; no additional billable start is introduced.
• Includes real BackendClient deadline and retry regressions. Thanks to @li872 for the original PR #97 implementation and tests, incorporated through #99. Fixes #89.

Optional patch from 0.2.7. Install in the Python environment used by your Agent, then restart it:

python -m pip install --upgrade kuma-defuzex

Publication to PyPI runs automatically after this Release. See polling and deadline semantics for exact limits.

KUMA 0.2.7

English runtime messages

• SDK-authored errors, field correction messages, update reminders and example prompts now use English.
• Official Judge invalid-result failures display Service is busy. Please try again later.
• Stable error codes, retryability, request IDs, privacy safeguards and billing behavior are unchanged.
• Chinese user content and Agent Profile headings remain supported. Historical Chinese service messages are recognized only as input, not echoed.
• Current bilingual documentation and installation examples are updated; previous community contributions are preserved.

This is an optional patch update from 0.2.6. Update in the Python environment used by your Agent, then restart it:

python -m pip install --upgrade kuma-defuzex

Package publication runs automatically after this Release; wait for the PyPI workflow to complete before installing 0.2.7.

中文说明

SDK 自产报错、字段提示、更新提醒及示例运行提示统一英文；不翻译用户内容，不改变错误代码、重试或计费。中文 Agent Profile 标题兼容保留。这是可选补丁，不自动安装。

Change: #95

KUMA 0.2.6 — Optional safety-baseline selection

• Explicit strategy="safety-baseline" validates seven Basic Safety groups,
  then uniformly selects one for one Case and one Run. Versions and catalog
  release come from validated discovery. Missing/ambiguous groups or unsupported
  capabilities fail before a Case POST; no silent fallback or seven-Case batch.
• An explicit Agent Profile group wins. Default auto, disabled automatic
  matching, custom Providers, Judge lifecycle and request recovery are unchanged.
  Use the original request ID for recovery; a new create_run may sample again.
• Documentation now describes actual multipart Evidence metadata, recovery
  headers, current errors and live catalog coordinates. Thanks to @YiWang24 for
  PR #88; integrates the safety-baseline work from PR #62 and README PR #54.
• 显式 strategy="safety-baseline" 验证七个基础安全组后等概率选一个，只生成一个
  Case、返回一个 Run。版本以目录为准；缺组、多版本歧义或能力不足会在 Case POST
  前拒绝，不静默回退、不批量生成七个 Case。Profile 显式选组优先。
• 默认 auto、禁用的自动匹配、custom Provider、Judge 与恢复行为不变。恢复请用
  原请求 ID，重新 create_run 可能重新抽组。文档同步实际 Evidence、错误和目录行为。
• Optional patch from 0.2.5: the new mode is opt-in and defaults do not change.
  Availability follows the official Release/PyPI; no automatic installation.
  这是可选补丁，默认行为不变、不自动安装；是否发布以正式 Release/PyPI 为准。

python -m pip install --upgrade "kuma-defuzex==0.2.6"

Install after the automatic PyPI publication completes. / 自动 PyPI 发布完成后即可安装。

Changes: #90 (integrates #62), #54, #88.

KUMA 0.2.5 — Official Judge error display

Official Judge error display / 官方 Judge 错误展示

• Internal Judge invalid-result failures now show service_busy with the fixed message 服务忙，请稍后再试, without internal diagnostic details. Older service responses are supported.

• The original retryable flag and safe response request_id are retained. This patch does not add automatic retries, create another task, or fabricate a report; committed Run history is retained.

• Case generation, input validation, authentication, permissions, and quota errors are unchanged.

• Scope: this is an error-presentation fix, not a fix for the underlying Judge/model invalid-result cause or a claim of improved Judge success rates. A failed Judge still fails and does not produce a report.

• 官方 Judge 内部结果失败统一显示 服务忙，请稍后再试，不显示内部诊断详情，兼容旧服务响应。

• 保留原 retryable 与安全请求 ID；不增加自动重试、不创建新任务、不伪造报告，已提交的 Run history 保留。

• 不改变 Case 生成、输入校验、鉴权、权限和额度错误。

• 范围说明：仅修错误展示，不代表 Judge/模型无效结果的根因已修复，也不承诺成功率改善。实际失败仍然失败，不会生成报告。

Optional patch update from 0.2.4 / 从 0.2.4 升级为可选补丁。

python -m pip install --upgrade "kuma-defuzex==0.2.5"

Install after the matching PyPI publication completes. Python imports and CLI remain kuma.
请在对应 PyPI 发布完成后安装；Python 导入名和命令仍为 kuma。

Changes: #87
Report: #86

KUMA 0.2.4

• Interrupted HTTP protocol/body reads become safe, retryable network_error
  errors rather than raw HTTPException. A safe X-Request-ID already received
  stays attached; no header means no invented ID. Responses close on failure.
  Existing retry limits and POST idempotency keys remain unchanged; user
  interrupts still propagate, and authenticated redirects remain blocked.
• CLI output cleanup failures no longer mask the original stable save error.
  Cleanup is best-effort: an OS-denied unlink may leave a temporary file.
• Offline regressions cover executed Strategy Group metadata, saved Cases, and
  GET-only recovery without another Case POST. No new strategy behavior.
• HTTP 协议或正文读取中断现在返回安全的 network_error，保留当前响应的合法
  请求 ID；不改变重试上限或 POST 幂等键，不吞用户中断、不放开认证重定向。
  CLI 临时文件清理失败不再覆盖原保存错误；操作系统拒绝删除时可能保留临时文件。
  补充策略组元数据、保存 Case 和 GET-only 恢复的离线回归，不增加策略功能。
• Thanks to @Moshiii (PRs #80/#81) and @luochen211 (PR #69). 感谢社区贡献。
• Optional patch update from 0.2.3; no automatic installation. Availability
  follows the official Release. 这是可选补丁，不自动安装，以正式 Release 为准。

python -m pip install --upgrade kuma-defuzex