octet
← Changelog

octet 0.6.3

GitHub release Source notes

Released 2026-08-28.

v0.6.3 is a focused reliability, observability, and evidence release. It makes active work truthful in the terminal, closes bounded Responses/WebSocket and Harbor cancellation gaps, reconciles token accounting, and publishes the methodology behind the existing Terminal-Bench result. The command and binaries remain ygg and ygg-host; ygg --version reports ygg 0.6.3.

Highlights#

Truthful run lifecycle#

  • Every accepted interactive run opens with a model-adaptive shimmering Working state.
  • Thinking appears only after a real reasoning delta. Its fixed bold header shimmers while the latest explicit heading and plain Ctrl+O hint remain on the subdued detail row.
  • Visible assistant text is the liveness signal while it streams, so Working stays hidden until a completed turn leaves the owning run active.
  • Reasoning/Working handoff to a tool preserves the historical cache prefix and truncates only materialized transient rows instead of rebuilding long transcript history, keeping animation and composer input responsive.
  • Static transcript markers no longer repaint on a timer; intentional status shimmer changes style only at a bounded 80 ms cadence.
  • Normal provider completions with no visible text, media, or tool call now fail clearly instead of reporting a silent success.

Provider and cancellation correctness#

  • Codex Responses error envelopes accept the observed nullable provider code while still requiring a usable error message.
  • A connection-lifetime-poisoned preferred Responses WebSocket is retired before a bounded pre-generation HTTP fallback.
  • Automatic replay stops once text, reasoning, media, or tool generation has begun; ambiguous post-output failures remain terminal.
  • Dropping an active Responses/WebSocket result closes its owned transport.
  • Harbor's Docker adapter now owns an independently cleanable process group, performs TERM→KILL descendant cleanup before artifact finalization, and fails closed when process death cannot be verified.

Local observability and accounting#

  • ygg doctor provides read-mostly local checks for prerequisites, provider configuration, and model visibility without starting an agent or executable extension.
  • Optional ygg.telemetry.v1 JSONL records request timing/TTFT, retries, disjoint usage buckets, tool timing, compaction, and bounded run outcomes. It is off by default, owner-only, and does not record prompt, argument, result, or provider payload contents.
  • Provider, session, telemetry, and Harbor accounting now agree on disjoint uncached/cache-read/cache-write input. Cache-hit detail is never added to input a second time, and Harbor includes all durable usage operations and cache writes.
  • New standard-library analysis tooling reproduces Harbor distributions and native/finalized accounting differences.

Evidence that carries its caveats#

The repository now includes a compact checksummed Terminal-Bench 2.1 report.

The frozen v0.6.2 GPT-5.6 Sol/max campaign recorded:

  • 87.87% raw — 391/445;
  • 86.97% primary local audit — 387/445;
  • 86.52% strict sensitivity — 385/445; and
  • 97.75% primary/strict Pass@5 — 87/89.

The audit used GLM-5.3 Flash as a surrogate judge over all raw successes, followed by manual review. It is not official Terminal-Bench maintainer adjudication. The historical paid campaign was not rerun for v0.6.3.

Install#

The version-pinned installer supports GNU/Linux x86-64, macOS x86-64, and macOS Apple silicon:

curl --proto '=https' --tlsv1.2 -LsSf \
  https://github.com/skaft-software/ygg/releases/download/v0.6.3/install-ygg.sh | sh

Then verify:

ygg --version
ygg --help
ygg doctor

To compile the pinned tag with Rust 1.86 or newer:

cargo install --locked \
  --git https://github.com/skaft-software/ygg \
  --tag v0.6.3 \
  --bins \
  ygg-coding-agent

Packages and compatibility#

  • Core crates and first-party packages target Ygg 0.6.3.
  • Executable extension bundles use API 0.2 and require exact Ygg 0.6.3.
  • Matching Ygg Serve archives are published for the three supported targets.
  • Existing local configuration, credentials, sessions, and Serve data remain in place across installer upgrades.
  • ygg update --check and ygg update continue to use the installation channel that installed Ygg.

Known limits#

  • Linux musl binaries are not part of this release.
  • Theme selection remains disabled; the compiled model-aware default is the only runtime theme.
  • The Harbor termination proof covers the Docker adapter. It does not claim the historical v0.6.2 campaign was rerun or that non-Docker providers were tested.
  • Full raw benchmark/judge trajectory trees remain outside Git pending privacy and redistribution review; the compact package contains sanitized result/audit metadata, original hashes, and the retained judge checksum index.
  • The Codex comparison in the evidence report is raw aggregate only. Run dates, provider snapshots, available trajectories, and adjudication pipelines differ.
  • npm, Homebrew, PyPI, GHCR, universal Pi compatibility, universal MCP compatibility, and a performance site are not part of v0.6.3.