banner
agent debating
agent config
self test
self test 2

Inspiration

What it does

FKST Devbored — project writeup

Devpost-ready description. Repo: https://github.com/ChronoAIProject/fkst-devbored

Inspiration

The loop is all the rage — engineers at Anthropic and OpenAI now describe their day job as building the loop that builds the code. OpenAI's Symphony and the harness-engineering conversation crystallized the idea for us: the interesting problems live in the harness around the loop, not the model inside it. A loop that runs 24/7 without staying aligned to your judgment isn't autonomy — it's loop vibing. We wanted the smallest honest tool that lets a developer run Codex in a loop and still trust what they're looking at.

What it does

FKST Devbored is a lean harness for Codex dev loops with a no-build HTML console over it:

Board — the loop's work, grouped by signal authority: trusted marker facts vs. label hints, never conflated.
Evidence chain — every state transition as its actual carrier: three state:v1 markers, a separate result:v1 decision, a merge label rendered as a hint. Markers = fact, labels = hints, unknown ≠ zero.
Who decided — the recorded five-seat consensus with each seat's real argument prose.
Council sandbox — author a review team against the real council contract: every dropdown and bound comes from the contract file, invalid definitions can't export, decisions preview under the real agreement reducer.
Contract self-check — 55 human-readable cases run in the browser and headless (node demo/tools/selfcheck.mjs).
Engine snapshot — a loopback, read-only view of the substrate delivery ledger that renders offline and unknown honestly instead of fake zeros.

Three design choices keep it approachable: lean editable architecture, GitHub as the durable layer (issues/comments/labels hold business state, auditable by any human), and a light HTML front end with no build step — clone, two commands, running in under a minute.

How we built it

With the kind of loop the product visualizes. Codex implemented everything: 52 sessions on final-assembly day (25 headless codex exec build/review streams), each with a written spec, exclusive file ownership, and a hard test gate — recorded model gpt-5.6-sol. A PM froze interfaces between streams so they could run in parallel; independent reviews verified actual files with hostile probes (attacker-comment injection, forged states, contract mutations) before anything was accepted; a human held the gates. Codex also reviewed adversarially: it RETURNed our first scope dossier with 8 blocking fixes and found 5 latent semantic faults at round-close — all fixed and re-verified. Even the repo's name came from a recorded 6-seat council deliberation reduced with the repo's own validator (FKST-devbored won 5/6).

Challenges we ran into

Honesty is harder than features. The recurring bug class wasn't crashes — it was quiet overclaims: a board column promoting a label hint to a fact, a "live" badge on an offline database read, truncated arrays rendered as exact totals, a marker parser that accepted decisions as lifecycle states, a CAS comparator that disagreed with the engine's on version equivalence. Each was invisible on the happy-path fixture and wrong under other data. The fix was process: adversarial review lenses, hostile fixtures asserting exact failure codes, and a 51-case parity corpus derived from the engine's own tests.

Accomplishments we're proud of

A demo where every claim on screen is auditable — and a build trail to match: every stream's session ID is published in docs/SUBMISSION.md.
A contract validator whose self-check reads like English ("A Council definition with no reviewer seats is rejected") — 55/55 green, in-browser and headless.
A faithful browser port of the engine's transition-version comparator, proven against engine-written ordering keys.
Cold-clone to running console in under a minute, no build step.

What we learned

Harness engineering is about making the loop work the way you want — splitting work into small durable issues beats big contexts, and durable evidence is what lets a human check an agent's judgment and reformulate the loop when they disagree. Multi-session builds with frozen seams parallelize beautifully — and leave seams a cold-eyes review must catch.

What's next

Live read-only GitHub business reads behind the local BFF (trust-gated, truncation-honest), then launch-snapshot Workflow/Council configuration — councils authored in the sandbox, applied only at loop launch with explicit acceptance receipts. Remote writes stay behind explicit human authorization at every step. fkst-substrate — the provider-independent engine under it all — is public and forkable.

Built with

codex (gpt-5.6-sol) · fkst-substrate (Rust engine) · vanilla HTML/CSS/JS (no build) · Node (loopback BFF + headless checks) · GitHub as the durable state layer