mathiasandClaude Opus 4.8 c40b46b661 fix(llm): send generous max_tokens on every request
The copied OpenAI-compatible client sent no max_tokens. Thinking models
(qwen3, deepseek-r1) spend their budget on the reasoning trace and return
EMPTY content when max_tokens is unset, which the summarizer treats as an
error. ADR-004 says change Tapir's copy rather than the hyperguild upstream,
so set a generous default (8192) leaving room for both reasoning and output.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-02 20:55:49 +02:00
2026-06-02 11:06:13 +00:00

tapir

Watches a user's YouTube/Vimeo subscriptions and, when a subscribed channel posts a new video, summarizes it into highlights and takeaways using a local-first AI stack with an optional, per-user BYO-AI fallback. Standalone-first; feeding a personal knowledge base ("brain") is one optional sink, not the reason Tapir exists. Written in Go.

Status

Pre-code. The repository currently holds the guardrail documentation — vision, decisions, architecture, data model, and behavior specs — committed before implementation so the design intent is version-controlled and the build has something to be checked against.

Read these first (the guardrails)

Doc What it is
VISION.md Product vision, principles, and the staged Definition of Success. Stage 0 ("useful to me") is the gate before any multi-user work.
DECISIONS.md Architecture Decision Records (append-only). Why Go, why no Supabase, standalone-first, captions-first, etc.
docs/architecture/architecture.md C4 context + container diagrams, key sequence diagrams, and the Clean Architecture layering (Mermaid).
docs/data-model.md Entities and the per-user isolation model (Stage 0 / Stage 1 scope).
docs/use-cases/ Gherkin .feature files — the BDD behavior spec that seeds the test suite.

Approach

  • Clean Architecture / ports & adapters. A provider- and sink-agnostic engine depends only on interfaces (VideoSource, Summarizer, Sink, SecretStore). YouTube, Vimeo, the AI router, the user store, and the brain sink are adapters. "Standalone vs homelab" is a wiring choice, not two codebases.
  • TDD/BDD. The .feature files are the living behavior spec; the use-case core is tested through fake adapters. Behavior is specified as executable scenarios, not prose that drifts.
  • Trunk-Based Development. Commit directly to main, one logical change per commit, every commit deployable (see ADR-009). CI is the quality gate.

Conventions

Reuses homelab conventions: Go, Dex for identity, ESO + 1Password for secrets, Postgres for persistence. No new auth or secrets system (ADR-002).

Next

First implementation step: scaffold the Go service (engine + interfaces + the copied llm package), captions-first, with the store and brain sink adapters — decomposable into independent units suitable for a Claude Code swarm. Tracked as the first build issue.

S
Description
Watches a user's YouTube/Vimeo subscriptions and summarizes new videos (highlights + takeaways) via local-first AI with optional BYO-AI fallback. Standalone-first; brain is one optional sink. Go.
Readme
4.5 MiB
Languages
Go 91.3%
Python 4.5%
templ 4%
Dockerfile 0.2%