Tau v0.2.1
User-facing changes since v0.2.0, selected from the full commit descriptions.
Before upgrading
- Rust extension SDK users must update their source. The harness uses protocol
10.2 and SDK 0.10.0.Configuregains optional-secret absence metadata,
changing Rust struct construction. The seven packaged external projects remain
on SDK 0.8.0 / protocol 10.0; they are admitted with the existing
same-major warning and best-effort behavior, but do not gain the new metadata API. - Tool-policy lists now merge additively. The six tool name, group and tag
enable/disable selector lists append unique entries across configuration layers
and agent/group/role scopes. Use an explicit empty list to clear an inherited
field; omission inherits it. Review overlays that previously relied on a
nonempty list replacing an earlier list. - Credential and configuration refresh still requires a full harness restart.
An extension-only respawn retains its resolved credential snapshot. This release
does not activate a new binary or restart any running session for you.
GPT-6.1 Sol and cost estimates
- The ChatGPT/Codex catalog now includes
gpt-6.1-sol, with its context/output
limits, supported reasoning levels and standard short-context API-equivalent
rates. The shipped default model is unchanged; account access is not guaranteed. - Sol uses the existing private Function/Custom tool workflow and local-summary
compaction. Unverified private native/inline compaction, Lite, image tool
results, phase and verbosity remain unadvertised. The public Responses example
uses the existing text/Function-only adapter. - Official OpenAI Chat Completions rejects unsupported Sol tool declarations and
tool history locally rather than silently dropping them. Custom compatible
endpoints retain their configured contracts. - Fixed cost estimates were audited. The retired DeepSeek V4 Flash alias routes
to V4.1 Flash with conservative peak estimates; the unknown-model fallback is
labeled GPT-5.5-equivalent. These are estimates, not invoices: long-context,
tier, regional, tool, subscription and dynamic pricing remain outside the
scalar base-rate calculation.
Configuration and web tools
- Optional web credentials now behave as declared. Exa, Parallel and You retain
their anonymous configuration when optional keys are absent; Brave, Tavily and
Firecrawl leave their configured pools. Missing undeclared or required secrets
still fail, and both search and fetch must retain a usable provider or the
entire extension rejects configuration with a pool-specific diagnostic. - Hosted web activity distinguishes search, page opening and page searching
without showing queries, URLs or patterns in the activity label. Prompts explain
when an external fetch is a useful fallback rather than a mandatory duplicate
read. Cached/index access is not a guarantee of a current live page. - A complete annotated harness example is linked from the learning guides. It
includes GPT-6.1 Sol role selections and fictional project-path placeholders;
adapt it to your setup rather than treating it as the shipped default.
Agent responsiveness
- Completed delegates can receive follow-up prompts and peer messages while older
background tools continue running. Background results retain their ownership
and normal delivery; genuine replacement inference or foreground tools still
protect their active turn. - Ordinary agent messages admitted during an exact or plural tool wait become
eligible after 20 seconds, down from 120 seconds by default. Configured
overrides still apply. This does not cancel tools or impose a global delivery
deadline; the ordinary-message any-input wait remains five seconds.
Distribution and limits
Tau remains pre-1.0. The native distribution includes Tau and seven independently
versioned external projects, with nine executables including the Telegram gateway.
AMD64 and ARM64 each have individual and tau-full DEB, RPM and tar.gz packages.
tau-full DEB/RPM is an exact-dependency metapackage; its tarball combines payloads.
Native packaging checks compilation, ELF shape, source/version metadata, archive
integrity and a bounded version probe. It does not qualify package-manager
install/uninstall, supervised startup or broad distro/CPU/kernel portability.
The binaries require the conventional glibc loader, glibc 2.34 or newer, libgcc
and CA certificates; NixOS users should use the Nix route. Linux 5.12 remains the
documented mount-API minimum. Assets are checksummed, not signed or independently
attested, and bit-for-bit reproducibility is not claimed.