Imported from Sheshiyer/wtfmedia (AGENTS.md). Install upstream with npx skills add Sheshiyer/wtfmedia. Copyright stays with the author.

Agent operating contract

This repository is wtfmedia.

1. Read README.md, PROJECT.md, .project/HANDOFF.md, and docs/AGENT-ONBOARDING.md before starting work.
2. Treat the Thoughtseed Labs vault as referenced knowledge, never as a runtime dependency or a place to copy private notes, transcripts, or seed corpora.
3. Preserve the existing tooling and deployment boundaries. Use the commands declared in PROJECT.md and keep generated output ignored.
4. Keep changes scoped to this repository. Do not edit vault registries, native client stores, Paseo, OmniRoute configuration, provider credentials, or external deployment state without a separate owner-approved task.
5. Never add secrets, .env material, native session identifiers, prompt or response bodies, or machine-local absolute checkout paths.
6. Record a bounded checkpoint in .project/HANDOFF.md when a reviewed change is ready for another client to pick up.

Current project facts for agents

• WTF OS and Ask WTF are the product center. Ask WTF must answer from retrieved transcript evidence or fall back truthfully; never claim the model cannot hallucinate.
• Production domain: https://wtfhq.in.
• Production chat contract: browser/web calls /api/chat with messages: [{ role, content }] and sourceMode of published, uncut, or both. The response body is streamed/plain text; source metadata is exposed through X-Sources and fallback state through X-Fallback.
• Cloudflare target resources are wtfmedia-edge, wtfmedia-web, wtfmedia-catalogue, WTFMEDIA_STATE, wtfmedia-catalogue-v1, wtfmedia-ops, wtfmedia-ingest, and wtfmedia-ingest-dlq.
• Current approved corpus receipt: 55 published transcript assets, 49 approved mapped uncut text assets, 55/55 published plus 49/49 uncut KV receipts, and 11,948 Vectorize records indexed by source_mode and video_id. D1 verifies all 49 mapped uncut assets; mapping is not trusted timeline alignment.
• Deferred sheet exceptions stay out of "fully ingested" claims: WTF is a Battery?, WEF - Economics, The Foundery, and the Brain Armstrong transcript-row mismatch.
• The latest ingest hardening requires a declared available D1 source asset and backing R2 object before vector staging. If that receipt is missing, fail closed as source_asset_unavailable.

Beta staging lane (v0.3.3-beta.1)

The beta track adds authenticated features that do not regress the public production path (v0.2.0). Key additions:

• AppRail restored from the alpha baseline with login flow wired through protected settings.
• Authenticated Ask WTF staging lane: server-side RAG, persisted chat answers, and protected chat history. Turns route through the server rather than client-side fetch.
• Consolidation policy panels: memory governance, RAG source health, and session history — exposed under the operator workspace behind the staging gate.
• Staging integration boundary: unverified operator release UI is hidden; beta sign-in routes through protected settings; the alpha baseline is separated from the beta track at the route and component level.
• Edge Frame.io fallback: the edge worker falls back to episode Frame.io links when direct URLs are unavailable.

The beta is a repository tag and merged PR, not a production deployment. All production boundary rules from the Boundaries section still apply.

Boundaries

Merging to main does not by itself prove Cloudflare production is running the same source commit. Verify production separately with Wrangler deployment receipts and live API probes. Do not deploy, rotate secrets, mutate DNS, enqueue live ingest, or broaden the corpus unless the user explicitly asks for that action in the current task.

This packet is active for repository work, but relocation, registry writes, session migration, provider changes, and production cutovers remain manifest-gated.

Temperance project rail

This repository is registered with Temperance Engine as a project rail. Host runtime (models, OmniRoute, OpenCode plugins) lives under ~/.temperance_engine and ~/.config/opencode; this repo owns planning and acceptance.

Auto next-wave

When an agent session starts in this cwd, enrich injects dispatch: NEXT-WAVE …. Do not wait for the user to say "temperance dispatch" or "proceed".

temperance-next-wave --cwd .
temperance-project-init --cwd . --check
temperance-batch --foreground --tasks .planning/next-wave-tasks.json --concurrency 4 --worktree

Manifest: .temperance/project.json (schema temperance.project.v1)