Coding OS
Sign in
Browse docs▾

ADR-0012: A lighthouse consumer (streamos) breaks the dogfood monoculture

  • Status: Proposed (2026-06-20, audit group F) — strategy + first lighthouse named; the multi-week build is owner-driven
  • Deciders: Kourosh Ebrahimzadeh
  • Context tags: flywheel, dogfood, monoculture, metrics, learning, strategic-audit-2026-06

Context

The strategic audit (2026-06-19) named dogfood monoculture the master risk, upstream of every other finding. The evidence:

  • Essentially all recorded work is INFRA (the meta-repo building itself); near-zero FRONTEND, little BACKEND. The cognitive loop has never run on a real consumer application.
  • This makes every "intelligence" claim unfalsifiable: the flywheel shows 98.75% success not because the system is brilliant but because one agent does one kind of well-understood INFRA work. cos_route_model cannot give a differential recommendation (one model family, one domain, all CLEAR/COMPLICATED); learn_extract mints almost nothing (no variance to learn from); the formula composer sits dormant (interactive INFRA rarely triggers COMPLICATED+ dispatch).

The other remediations close the mechanism gaps but not the signal gap:

  • B1 (TASK-463) wired a real blocked outcome emit path, so non-success outcomes are now structurally possible.
  • B5/C1/D/G/461 removed dead weight, cut the token floor, and made coverage honest.

But wiring an emit path does not create variance — only diverse real work does. A flywheel that only ever turns on INFRA tasks stays degenerate no matter how clean its plumbing. The missing input is a real, non-INFRA application driving genuine FRONTEND/BACKEND tasks — with their reworks, blocks, and partial outcomes — through the same board, metrics, and learning loop.

Decision

Adopt streamos as the first lighthouse consumer. The monoculture is broken by running the full cognitive loop on a real app, not by more meta-repo work. The multi-week build is owner-driven; this ADR fixes the strategy, the choice, and the success criteria so the work is measured, not vibes.

  1. Lighthouse = streamos (~/Projects/<consumer-project>, Go + SvelteKit). It is already a registered consumer with a live .coding-os/coding-os.db and a seeded board — "half-wired" exactly as the audit found. It is a real product (BACKEND + FRONTEND), not INFRA, so its work exercises the dimensions the meta-repo never does.
  2. What "running the loop" means — real streamos feature/bug work flows through cos task-* (BACKEND + FRONTEND swimlanes), classified through the Cynefin gate, executed with the stack skills, verified, and closed with honest outcomes (the B1 blocked path + explicit partial when acceptance is only partly met). The point is the outcome distribution, not task count.
  3. Success criteria (the falsifiability test) — the lighthouse is working when, on streamos's own DB: (a) task_outcomes carries a non-degenerate mix (success / rework / blocked, not ~99% success); (b) ≥1 domain other than INFRA has enough samples for cos_route_model / learn_suggest to return a history-backed, differential recommendation; (c) learn_extract promotes at least one pattern from real non-INFRA work. Until then, "the system makes agents smarter" remains unproven.
  4. This is the demand-driver for depth-over-breadth (report §S4). Stack work is justified by a real consumer needing it (streamos's Go + SvelteKit stacks earn promotion to stable via stack-maturity.md), not by minting stack #27.

Consequences

  • Positive: the headline differentiator ("a cognitive OS that makes agents measurably better over time") gets its first falsifiable test bed. The flywheel finally has variance to learn from; the B1 emit path has real signal to carry.
  • Positive: streamos doubles as the proof that the consumer-distribution seam (ADR-0010) and cross-adapter seam (ADR-0011) are needed before a second external consumer — the lighthouse surfaces those pressures first.
  • Negative / cost: this is weeks of real product work, owner-driven, not a meta-repo refactor an agent finishes in a session. It also splits attention between building coding-os and building streamos — that tension is the point (dogfood), but it is real.
  • Deferred: no code lands in the meta-repo from this ADR. It is the strategic decision + measurable success criteria. Progress is tracked on streamos's own board, not here. Linked from memory strategic-audit-2026-06.

See also