Bello 0.7.2

Highlights

• Safer interrupted-run handling: eligible controller runs can continue only after protected state, consumed budgets and previous-worker cleanup have been checked. Uncertain side effects and final workspace updates are not blindly replayed.
• Updated Pi to 1.0.4 and Claude Agent SDK to 0.2.164, paired with managed Claude CLI 2.1.293. New primary-role configurations use GPT-6 Astra; saved model choices are preserved.
• Clearer completion after adversary review, with Windows snapshot, watcher-lifetime and drive-root fixes.
• Managed Codex 0.161 helpers for optional Smart Execution and log distillation on Linux x64, Windows x64 and macOS arm64. These helpers are separate from the ordinary user-managed host Codex CLI.

Verification

Qualified source: 41325958fe81eb2b6c4eee59a93b69f62c0b4789, tree 339952833f74ae93758f3d145ae0dede8c4e2f32. Release integration is tracked in PR #39.

The final-source local gate passed 4,580 Python tests, 133 Node tests and 24 real offline Pi tests. All eight source CI workflows passed, covering cross-platform tests, Windows sandbox and restoration, Claude on Windows, package builds and five clean-install configurations, native Linux/Windows, native macOS, archive installation, and production HTTPS installation.

Production pins identify actual archive and manifest hashes from the managed native helper release. Both Windows HTTPS configurations qualified the same pinned Windows archive. Native qualification uses real binaries, a synthetic loopback provider and four actual CPU ModernBERT inference cases plus two protected-output bypass controls; it does not claim live-provider quality or fresh compilation for cache-hit builds.

Compatibility and safety

Automatic crash continuation is conditional. Linux and Windows require native descendant-cleanup checks. macOS fails closed after a controller crash because owned process groups cannot prove cleanup of arbitrary detached descendants. Unknown side effects, unverified launchers, active children, approvals/reviews and unsafe final-apply phases require manual inspection instead of blind replay.

The native Windows qualification does not establish arbitrary public-path read confinement. Apple Developer ID signing/notarization is not claimed.

Bounded live Pi, Claude, Codex and mixed-engine fixtures passed on earlier source 0bbe6e35, not the final release bytes. One Pi post-cancellation worker notification remains unexplained despite successful continuation; that observation is not discarded.

Bello managed Codex 0.161.0 — Smart Execution and log distillation

Qualified native helper assets for Bello 0.7.2 release preparation. This is a GitHub-only native-helper prerelease, not the Bello Python package release.

Targets: Linux x86_64, Windows x86_64 and macOS Apple Silicon. Each archive contains the native Codex launcher, required platform helpers, capability manifest, source/build provenance and upstream license notices. Target-qualified checksums, manifests and BUILD-INFO are published alongside the archives.

Bello selects this private-cache helper for Smart Execution or native log distillation. Ordinary Codex execution with those features off continues to use the user's host Codex CLI; no global Codex installation is replaced.

Qualification requires selection, asynchronous execution/cancellation, persistent history and CPU ModernBERT proofs against the exact candidate and a cold installed cache, plus the bundled Linux sandbox proof. ModernBERT's six cases comprise four inference runs and two protected-input bypass controls. Proofs use synthetic local providers, not paid model calls or a model-quality benchmark.

Linux and Windows require full-tree guardian cleanup. macOS establishes normal process-group cleanup only; arbitrary detached descendants are not covered and automatic whole-run crash continuation remains fail-closed. The Windows proof does not claim confinement of every arbitrary public-path read.

Source tag target: 988af82. Exact per-target archive identities and qualified build/proof inputs are recorded in the attached metadata. Release operators retain the complete bounded CI receipts.

The normal public HTTPS download and installed-cache gate follows publication of these helper assets. Do not treat this prerelease alone as the completed stable Bello 0.7.2 release.

macOS binaries passed native signature verification; Developer ID signing and Apple notarization are not verified.

Qualified source checks: https://github.com/Makson179/Bello/actions/runs/37820605970 (Linux and both Windows environments), https://github.com/Makson179/Bello/actions/runs/37820606027 (macOS arm64).

Bello 0.7.1

Bello 0.7.1 adds Sonnet 5.5 support, verified native Windows Claude Code installation, and configuration editor improvements.

Highlights

• Add Claude Sonnet 5.5 through the pinned Claude Agent SDK 0.2.161, including exact model identity and supported effort levels.
• Support Windows SDK sdist installations with a managed, pinned official Claude Code 2.1.284 CLI. Downloads and cached binaries are checked for size, SHA-256, private permissions, Anthropic Authenticode signature, and version.
• Close a cache-integrity gap: verification hashes the binary bytes on every readiness check, including same-size modifications with a restored modification time.
• Use the default private per-user Windows cache and provide actionable, sanitized ACL diagnostics. Shared or unsafe cache ancestors remain rejected; security checks have not been weakened.
• Share validation between the configuration editor and startup preflight; report invalid active roles, subagent profiles, and provider readiness before launch.
• Improve model/effort transitions, autosave visibility, and Fast-mode wording. Fast denotes the OpenAI/Codex priority tier and is separate from Claude and Smart Execution.

Verification

• All required and targeted checks passed before merging PR #38, and all post-merge checks passed on the exact release commit.
• Full test matrix passed on Linux (Python 3.11–3.14), macOS (Python 3.13), and Windows Server 2022 / Python 3.11 and Server 2025 / Python 3.14. Platform skips and expected failures remain explicitly reported in CI.
• Both Windows versions passed real SDK/CLI installation, checksum, private ACL, Authenticode, version, cache-reuse, deterministic tamper detection, and signed-out Sonnet 5.5 metadata checks.
• Windows restricted-sandbox and offline Pi checks passed, including non-administrator enforcement and ACL restoration. Native Codex proofs passed on Linux and Windows.
• Claude metadata checks are signed out; native proofs use synthetic local providers. These checks do not claim authenticated live model generation.

Source: 3bf0076f6da8119770a1b289e3a21ada69e046a7 (PR #38). README, media, and native runtime pins are unchanged.

Publishing this stable release starts the repository's existing Publish workflow for the portable wheel, Windows wheel, and source distribution on PyPI.

Full changelog: v0.7.0...v0.7.1

Bello 0.7.0

Bello 0.7.0 — Smart Execution

• Add opt-in Smart Execution for independent tool calls across Codex, Claude Code, and Pi, including coder, completion-review, adversary, and their supported subagents. Runtime observers remain outside Smart Execution.
• Add asynchronous result delivery, waiting/cancellation handling, and regression coverage for terminal states.
• Deliver verified native Smart Execution bundles automatically for Linux x86_64, macOS arm64, and Windows x86_64; validate complete bundle contents before starting native execution.
• Improve Windows native setup and explicit root-read permission handling, while preserving private-file and metadata protections.
• Keep completion-review evidence in selective, read-only files instead of always inlining the full context.
• Bypass distiller processing for tool output up to 200 UTF-8 bytes. This does not replace or retrain the existing ModernBERT model.
• Improve Pi/OpenRouter transient backpressure handling and runtime/toolchain compatibility.

Claude Code subscriptions, Pi providers, mixed-model teams, and ModernBERT distillation were already available in 0.6.0.

Verification

Pi and mixed-engine smoke tests each passed their external verifier, 30/30. Actual cold HTTPS native installation and all 13 native async/selection cases passed on Linux, macOS, and Windows. The Windows archive also passed nine selection cases and thirteen async cases after fresh local installation.

Native assets: verified native helper release.

Package build

PyPI distributions are built from commit 21ce1e4, including the subsequent Windows journal fix, network-test stabilization, and refreshed README. The original Git tag is unchanged.

Native Codex 0.155.1 — Smart Execution helpers

Technical native helper assets for Bello Smart Execution and tool-output selection. This is not a Bello application release and does not change the latest Bello version.

These are Bello-patched, unsigned builds based on OpenAI Codex native version 0.155.1, upstream revision be2951ea34f0d295ed0becf97079f92fa5f6950e. Each archive includes the required companion executables, its selection manifest, build information, the modification patch, LICENSE, NOTICE, and third-party notices. They are installed in Bello's private cache rather than replacing a globally installed Codex CLI.

Assets: Linux x86_64, macOS arm64, and Windows x86_64. All three passed actual cold installation through Bello's production HTTPS downloader, all 13 credential-free native selection/Smart Execution cases, exact archive/member checks, and unchanged cache reuse. These checks used no account credentials or paid model calls. The Windows archive additionally passed all nine selection cases and thirteen async cases after fresh local installation; it contains the preserved, verified native executables without recompilation.

Platform	Archive SHA-256	Manifest SHA-256
Linux x86_64	ba27107574789187a147e3a3caa159e0f5d435bd071759be623fe5745c5bdc07	e48fdaddd736198010c8ae1cfd073fa8f511cb27ab6534e89d8405b2f8673407
macOS arm64	8614648cf79bf5582879b2c899cdb515e6da01687f4b65ee9d529cca25023029	792dbcd8df2672df020a63115dddd1af86e31ff5e1ca6c2008ff627be1057a1c
Windows x86_64	8993eea58f8d4f3466f72713eba62bd4842e702d92d8552ab79c3a2d4b7308b8	d6b859995d238de668cd9bb533d23d3c2a7c96b1940808295a197e2360830ccd

Bello 0.6.0

Bello 0.6.0

• Use subscription Codex through its native app-server, Claude Code through its native SDK, and API models through the Pi engine.
• Configure runtime supervision, completion review, adversarial testing, and log distillation independently.
• Mix models across the coder, reviewers, and subagent pools.
• Optionally compress tool output with the trained ModernBERT log distiller. Model weights download on first use; subscription Codex also downloads its compatible native helper on supported Windows, macOS, and Linux platforms.
• Install the Bello configuration advisor and delegation skills in Codex or Claude Code through the shared marketplace.
• Includes Windows native-distiller integration, Unicode transport fixes, Linux helper auto-installation, persistent scratch space for completion checks, and process-tree cleanup when cancellation races command startup.

Install or update

pipx install 'Bello==0.6.0' --force

For Claude and the log distiller:

pipx install 'Bello[claude,log-distiller]==0.6.0' --force

README · Runtime and providers · Native Codex helper · Marketplace

Native Codex selection helper 0.153.4 (Apple Silicon)

Technical component for Bello native log selection on macOS arm64. Modified OpenAI Codex rust-v0.153.4; source patch, Apache-2.0 license, upstream notices and dependency notices are included in the archive. This prerelease does not replace global Codex or release Bello to PyPI.

Bello 0.5.2

Bello 0.5.2 adds GPT-6 Astra (gpt-6-astra) to the model selector for Coder, Runtime supervisor, Completion reviewer, and Adversary.

• Supports Astra's Codex reasoning settings: low, medium, high, xhigh, max, and ultra.
• Preserves the existing model-specific effort limits and default model settings.
• Documents source-checkout configuration and explains why the desktop app may show fewer effort options than Bello.
• Adds regression coverage for configuration persistence, CLI overrides, model discovery, editor behavior, and agent RPC parameters.

Validation: 955 tests passed, 7 skipped; wheel and source distribution built and passed metadata checks. Three earlier live Astra smoke tasks exercised all six effort values across the four roles and completed successfully.

Upgrade with bello update or pipx upgrade bello.

Bello 0.5.1

Bello 0.5.1 fixes native Windows startup and bello doctor failures caused by normal Codex-managed junctions and active lock or temporary files in CODEX_HOME. It preserves authentication, configuration, persistent plugin data, and versioned plugin payloads while excluding reconstructible runtime state. Native Windows Server 2022 and 2025 regression coverage is included.

Bello 0.5.0

Bello 0.5.0 adds native Windows support, configurable coder and reviewer subagents, private advisory plans, an optional revision coder, bounded adversary retries, app-server transport recovery, atomic checkpoints, and lifecycle hardening. It also refreshes the README with the latest cost, quality, safety, and ProgramBench evaluation results.