Skip to content

Bello 0.8.1

Latest

Choose a tag to compare

@Makson179 Makson179 released this 06 Nov 04:29
51f309e

Bello 0.8.1

Goal mode

  • Successful turns continue automatically. Completion review runs only when the coder reports readiness.
  • goal_runtime / --no-goal-runtime runs Goal without runtime supervision. That means no runtime, cheap-triage or Goal progress review model calls, and runtime-off protection. As with --no-runtime, this also allows network access inside the sandbox.
  • Optional adversary checkpoint: goal_adversary_after_returns / --goal-adversary-after-returns N.
  • Periodic Goal progress review against GOAL_MAP.md, with optional, non-binding advice. A failed progress review is skipped and the run continues. A delivered progress advisory cancels a readiness review queued behind it.
  • /resume continues the coder. An idle Goal coder with nothing pending gets one continuation message after the 40-minute idle window.

Completion and watchdog

  • Outside Goal, automatic final review waits for 40 minutes of continuous coder idle (was 120 s).
  • Status notifications no longer reset the idle timer, and an unknown state no longer causes an endless wait.
  • Native Codex idle state is confirmed by a metadata-only status read.

Runtime supervision

  • The proof/event-hash delivery gate is replaced by informational staleness notes.
  • Old inspections get freshness labels, and evidence is verified before a defect is reported.
  • persistent_decision semantics are stricter, and restarts need stronger evidence.
  • The runtime prompts are rewritten.
  • New BELLO_RUNTIME_TRIAGE_EFFORT.

No Need For Context

  • New defaults: 244,800-token threshold and a 24,800-token absolute recent tail (retain_tokens, --context-retain-tokens). Percentage mode becomes legacy.
  • While No Need For Context is on, retain_tokens must be at most half of threshold_tokens; a lowered threshold with the default tail is rejected with a clear message.
  • A resumed thread may differ from the saved policy only when the new policy is off.
  • Native Codex patch: no estimate-based refusal; the trigger is the measured provider input+output, and accounting resets after compaction. This is in the helper release native-codex-0.161.0-nnc-tokens-20261105.

Claude Code backend

  • One pre-query retry for connect, initialize and context-usage timeouts.
  • Errors now show the CLI's complete error text, unredacted, plus a private stderr log. An error result from the CLI (for example an API error) is reported in full instead of "the turn failed".
  • Fix: a cancelled context preparation could swallow the next stop or turn start.

Recovery

  • The launcher check after preflight now actually runs.
  • Codex runs with No Need For Context on and Smart Execution off can now be recovered (the restore plan selected the wrong launcher).

Validation evidence

  • Test commands that redirect with 2>&1 (also with ; echo exit=$? or | tail) count as test evidence; before, completion review could not see such passing runs.

Native Codex

  • The app-server handshake reports Bello 0.8.1 (it said 0.5.2).

Windows

  • pipx and virtual-environment installs no longer fail with "process guardian rejected the worker" (#41).
  • prepare --network names the conflicting WFP fields (#42, partial).

Linux

  • Immutable-read recognition works with Ubuntu 26.04 uutils.

macOS

  • git works inside the command sandbox (it exited 128 on /etc/gitconfig).

Controller

  • Changed-path filtering no longer starves the event loop on very large trees.
  • Timed-out Git inspection processes are killed and reaped. This fixes an observed out-of-memory failure.

Configuration

  • New editor rows and advisor schema fields for the options above.
  • Saved configurations keep their explicit values.