Outcomes and honest boundaries first; machinery only when you ask for it.
Your question: What is this, what is real today, and why might I care? Start here, then cross the lane whenever something catches your eye.
Baseline is a public workbench for agentic collaboration and Valheim community tooling. The same body of work looks different depending on why you arrived, so start from the lane that sounds like you. Choose your lens: https://djcdevelopment.github.io/baseline/ Explore everyth...
Three million building pieces become 1,833 structures and 1,411 photographs framed by arithmetic rather than a human camera operator.
The public hub for the repository map, durable decisions, evidence, corpus, and discovery surfaces of the Valheim project fleet.
Historical 2026-08-06 evidence of a guild tracker becoming a player picker, in-game objective, and human-review evidence; current authority is comfy-quest.
Reads a Valheim world file and answers the questions that stop being walkable once a server gets big: where is everyone building, and who owns this.
An aggregates-only telemetry API with privacy tests that fail if a player ID, name, or position ever shows up — plus the whole stack, runnable on your machine.
A localhost MCP server that talks to the running game mod: apply a config profile, run a bounded lab test, read back what happened.
An arcane tome in-game that teaches you exactly what quests are possible, letting you practice writing them before trying them out for real.
Find every structure in a world save, place a camera at each one from its own geometry, and photograph them unattended into a ranked gallery.
The in-game spellbook and web Tome no longer repeat scaffold-era claims that only kills or combat can bind. All eight pages now distinguish stable canonical actions from diagnostic/query witnesses, name compatibility aliases and action coalescing where relevant, and regenerate from one authoritative prose source.
The interior lane has a public story: five vantage recipes, the throne that threw the camera, four skies over one hall, and the instruments that curate without consenting on anyone's behalf
Retro for the isolate-boundary and AM4 black-screen session. The through-line is not the Gateway fix but that four independent guards had never been able to return a failure: an identity endpoint that hardcoded the value it was meant to discriminate, an API contract file no test read, a C9 acceptance criterion satisfied by a file existing, and a compose launch whose failure mode is exit 0 against blank-rooted world mounts. ADR 0019 turns that into a rule -- a guard is not accepted until it has been observed to fail, acceptance criteria state a property rather than an existence, contract files are read by a test or deleted, receipts carry negative controls, and commands whose failure mode is a zero exit get rendered before they are run. The retro is candid about operator error: a composition fired at a retained boundary the plan forbids, cutover legs disarmed to force a green join (a regression Derek had corrected an hour earlier), four gateway recomposes that silently downgraded a pinned release image, two false fix-did-not-work claims from measurement errors, and three attempts to extract a C9 verdict from an artifact Derek had already rejected.
The final-cutover plan recorded C9 as machine/artifact complete with only Derek's one-word smooth/rough/mixed verdict outstanding, and the remaining-cost table listed it as 1 operator verdict. That was wrong for six days. The retained side-by-side clip's own receipt records events 4 for the OMEN panel and events 5 for the i5 panel across two 20-second views -- near-static frames with a counter overlay rather than observable movement. Derek had already declined to call it on exactly that basis; the plan recorded the absence of a verdict as a pending reviewer action instead of an insufficient artifact, so every reader since has been told C9 was one word from done. Corrected the C9 execution-status row, the C9 acceptance and exit criteria, the 2026-08-02 checkpoint, the remaining-cost table, and the immediate-next-build paragraph. C9 is now blocked on producing an artifact that shows motion, or on a live two-client window, and the retained-boundary list no longer protects the artifact from being re-produced.
The turnkey vision draft sat untracked at the repo root describing three friction removers as though all three shipped. lab_reload did not exist at all (it does now); the failure-boundary story credited LabEventRing, which is a buffer that prints nothing; and lab_export conflicts with fieldlab ADR-0018, which replaced outbox payloads with a durable EventLog row -- the bridge consumer explicitly rejects schema-1 outbox payloads. It also said 7 Schools of Magic, omitting World, while every source in the repo says 8, and placed diagnostics in an F5 overlay when F5 is Valheim's console and F6 is the lab panel. Rewritten as docs/quest-lab-turnkey-vision.md with a status column per friction remover and the honest gap named: the quest lane is proven by unit tests against the real contract and has never been run inside Valheim. START-HERE gained a ComfyQuestLab row and the full AM4 route list, and now flags that /questpicker and /steward appear in every nav bar with no route in this repo's Gateway source. 7-tools drift fixed in START-HERE and HANDOFF-2026-07-29.
The README and CHANGELOG both still claimed lab_setup was not here yet, while ComfyQuestLab.cs registers it as a console command that raises the practice gallery and points at the tome. Found while fact-checking a strategy document written by another agent, which listed turnkey lab_setup packaging as complete - the doc was right and the repo was under-claiming. Stale docs are the same failure in either direction: a README that undersells a shipped command costs a newcomer the one instruction they needed. Also noted for follow-up and NOT changed here: the ontology document says 7 Schools of Magic where every source in the repo says 8, probably from misreading the in-game roster line about seven of the eight categories not being wired; and the README's no download claim now contradicts workbench.json, which describes downloading a zip - that contradiction is unresolved and someone should settle which is true.