Release Notes
Added
- Support for reasoning and adaptive-sampling models. Tribal resolves, per provider and model, which request fields a target accepts and shapes each request to match: OpenAI reasoning models (the o-series and the GPT-5 reasoning line) send the output cap as
max_completion_tokens and have caller sampling parameters dropped, while Anthropic adaptive models (Opus 4.7) have temperature dropped but keep their required max_tokens. Both tribal check --providers and ingest now succeed for these models. An unrecognised model keeps sending every parameter, so ordinary models are unaffected.
- Per-stage
temperature and max_tokens from configuration now reach the model. Previously they were silently dropped and only recorded in the system fingerprint.
- Configuration validation for sampling parameters: when set,
temperature must be within [0.0, 2.0] and max_tokens at least 1. Model IDs may no longer be empty or whitespace-only.
Changed
- Sampling parameters are now optional in configuration. An unset
temperature or max_tokens means "use the provider default" rather than a built-in number, and the four built-in per-stage sampling defaults have been removed.
- The system fingerprint records the effective post-reconcile sampling parameters, so it reflects what is actually sent. Fingerprints change for any model whose request shape the capability layer adjusts.
- Rewrote the README around the shipped onboarding surface (
tribal bootstrap, tribal check, tribal mcp-config, Docker Compose, and the skills), and added CONTRIBUTING.md.
Fixed
- Reasoning models are no longer rejected by the provider readiness probe (
tribal check --providers), which previously forced temperature and max_tokens values these models reject.
Install tribal 0.2.2
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/tribal-memory/tribal/releases/download/v0.2.2/tribal-installer.sh | sh
Install prebuilt binaries via Homebrew
brew install tribal-memory/tap/tribal
Download tribal 0.2.2
