The scarce resource is the per-IP timedtext caption fetch (ADR-014); the pilot's
candidate set was mostly Shorts/clips/livestreams, each burning a fetch (a "none"
result is a completed fetch — it costs budget even when it yields nothing).
NewVideos now enriches candidates with one cheap Data API videos.list call
(contentDetails.duration + snippet.liveBroadcastContent — the quota API, a
DIFFERENT limit from the timedtext 429) and drops, before returning: videos
shorter than TAPIR_MIN_VIDEO_SECONDS (default 60) and any live/upcoming
broadcast. Dropped videos are never persisted, so the list declutters too.
Degrade-open: MinVideoSeconds=0 disables it (no quota call); a videos.list error
returns candidates unfiltered so discovery never breaks on a metadata hiccup. The
paste-a-URL path (VideoByID) is not filtered — an explicit request is honoured.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The first friendly-pilot live run produced zero summaries: koala/phi4-mini hit
three silent failure modes — 8k context overflow on long transcripts (HTTP 400),
intermittent malformed JSON (highlights as a bare string), and no fallback wired
at all (summarizer.New(primary, nil)).
Keep phi4-mini as the fast primary and add resilience around it:
- Ordered endpoint chain (summarizer.NewChain): phi4-mini → koala/phi4-14b
(local) → berget/mistral-small (worst-case external). All reached through the
one LiteLLM gateway by alias.
- A parse failure now advances the chain like a transport error — the old
Primary→Fallback shape returned the parse error without trying anyone else.
- Tolerant parse: highlights/takeaways coerce string→[]string, absorbing the
common small-model quirk without spending a fallback round-trip.
- Transcript truncation (TAPIR_MAX_TRANSCRIPT_CHARS=18000) prevents the overflow
rather than recovering from it; validated to fit phi4-mini's 8k window.
- Bounded completion budget (TAPIR_SUMMARY_MAX_TOKENS=1500) — the old 8192 budget
itself contributed to the overflow.
Local-first guarantee preserved by ordering: external endpoint is tried only
after every local one fails. TAPIR_CLOUD_FALLBACK_MODEL="" disables it entirely
for client/NDA deployments.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
homelab-integration.md gains a "Scheduled discovery" section documenting
TAPIR_DISCOVERY_INTERVAL and TAPIR_FETCH_RATE and the load-bearing
single-replica constraint (in-process scheduler → replicas: 1 is required;
>1 double-runs discovery). VISION Stage 0 carries a pointer to ADR-018's
gate-clock reset so nothing in docs implies the window started before
unprompted use was possible.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add GET /oauth/youtube/connect and /oauth/youtube/callback, mounted inside
the login + registration guard so CurrentUserID is always set and every
connection binds to the authenticated tapir user.
- connect: generate a per-user CSRF state (single-use, short TTL, bound to
the user), redirect to Google consent with access_type=offline and
prompt=consent so a refresh token comes back.
- callback: verify the state belongs to this user, exchange the code via the
existing auth.Exchange, persist the refresh token under a PER-USER ref
(web.YouTubeTokenRef = "youtube/<userID>/refresh_token") so tenants never
collide, then UpsertConnection (provider=youtube, status=active). Any
failure renders a clean error page and leaves no half-written state.
Reuses auth.Exchange and adds auth.AuthCodeURL (offline + consent) rather
than the CLI's listener/terminal flow (ADR-006: web flow, not CLI). The
ConnectHandler depends on a narrow web.Connections port, not the concrete
store. Wired in cmdServe only when YT client credentials are present;
TAPIR_YT_CONNECT_REDIRECT_URL configures the callback URL. Per-user token-ref
scheme documented in docs/homelab-integration.md.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
main.go dispatches `tapir auth` (interactive OAuth → persist refresh token via
SecretStore) and `tapir run` (wire YouTube source + local summarizer + store
sink, build engine, run the dedup-aware loop). Config-driven so live creds plug
in at demo time; SIGINT stops the loop cleanly. Block kept minimal so Worker E's
list/show cases union cleanly at merge.
Add .env.example documenting every TAPIR_* var and a README demo runbook. Pin
the summarizer alias-as-config decision and record the max_tokens fix in
docs/homelab-integration.md (clears two `confirm` items).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Implement ports.Summarizer in internal/adapters/summarizer. It routes through a
local Primary endpoint first and an optional BYO Fallback, owning the routing
itself (not delegating to llm.Router) so it can record AIProvider, AIModel, and
FallbackUsed on domain.Summary. Prompt asks for JSON {summary, highlights,
takeaways}; the parser tolerates thinking-model fences/reasoning and rejects an
empty summary.
The summarizer is the single egress point for content toward an AI model, so it
enforces the local-first guarantee from ai_routing.feature: with no BYO
configured (nil fallback) there is no external endpoint, so content reaches the
local stack and nowhere else. Tests assert all four scenarios via a fake client.
Model alias is config (TAPIR_SUMMARIZER_MODEL, host/name) — not hardcoded;
docs/homelab-integration.md notes it stays `confirm` and that thinking models
need an explicit max_tokens or they return empty content.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Conductor verified against the live koala k3s cluster before launching the
adapter workers: LiteLLM is in ns ai-stack, ClusterIP 10.43.159.89:4000,
off-cluster NodePort 30401. The doc's 31234 was actually llama-swap. Also
flag that sk-local-123 now 401s — use LITELLM_MASTER_KEY from the vault.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Lists the egress the build assumes (Go module proxy, toolchain download,
raw.githubusercontent for golangci-lint, github for non-proxied modules) and the
runtime assumes (LiteLLM gateway, brain-mcp, YouTube/Vimeo APIs, per-user BYO-AI
hosts only on opt-in), so a locked-down koala act_runner or dev env knows what to
allow or which GOPROXY to set. Notes the claude.ai sandbox allowlist is separate
and unrelated.
mathiasandClaude Opus 4.8