Architecture · the multi-agent harness · the pipeline layer · being welded in now

The Harness Plumbing — Where Work Outlives Its Workers

Warp has judgment at the top and isolated engines at the bottom — and a missing middle. Tonight welds in the pipeline layer: the failure walk that moves a dead engine’s job to the next fit engine, the record that lets a job outlive the session that started it, and the preflight that refuses a doomed dispatch before a cent of spend. This page is that architecture — the layers, the pipes, and one job flowing through them.

Fleet building now (run wf_c30305bf-1f7) · #2 laws ratified · #3 #4 intents drafted for your pen

The machine, in five layers

Warp is strong at the bottom of this stack and at the top. The middle layer — where a job should outlive any single engine and any single session — has been missing. That layer is what tonight builds.

have it
You + the lead
Judgment lives here and only here: what to build, which engines fit, what a verdict means, what merges. The lead writes briefs and adjudicates — it should never have to babysit a dying run at 2 a.m. Today it does. That is the whole complaint.
tonight
The pipeline layer — the missing middle
Three organs, one layer. The failure walk (#2): a dead engine's job moves itself to the next fit engine. The pipeline record (#3): the job's whole history is an object on disk, not state in the lead's head. The preflight (#4): a doomed dispatch is refused before a cent of spend. All three sit above the doors, below the lead — routing without judging.
upgraded today
The four doors
Sealed room on the Mac · single-shot api lens · cloud sandbox · codex's own cloud runner. Isolation, hard clocks, custody of the work, one ledger row per run — and as of this morning, every door hands back a typed machine word on every exit path instead of prose. The pipeline layer exists only because that word now exists.
thin spot
The engine fleet
Codex, grok, Kimi-K3, GLM, deepseek — each described by a card of facts: what it runs, its clocks, what it refuses, who its fallback is. Cards are the pipeline layer's entire knowledge of the world. The thin spot: the flagship fallback chain points at native engines that have no cards and refuse every dispatch — tonight closes that hole.
the floor
The record floor
The ledger (one row per run), harvested reports, work branches. Everything above writes down what it did in a form a stranger can recompute. The pipeline record grows out of this floor — same discipline, one level up.

By type: the three organs · fences · who is building it tonight · your calls

One dispatch through the pipes

Follow a single job — a build brief headed for codex at 2 a.m. — through the whole harness, including the night it goes wrong. Every stop is a pipe that exists, shipped this morning, or is being welded in right now.

  1. A job is born: a brief and a fit list. The lead writes a self-contained brief and names the candidate engines — or says nothing and inherits the failed card's fallback chain. From this moment the job is data: everything downstream reads facts, never intentions.
    ↘ go deeper
    Fit list is caller-declared, defaulting to the card's fallback_engine chain (ratified call #2). A fully silent card set means honestly exhausted after one attempt — early surface, never a guess.
  2. Preflight (#4, built tonight): the doomed dispatch never launches. Before any engine is paid, the job's declared domain is checked against the target card's facts. A card that says this engine cannot take this work refuses or reroutes on the spot. A card that says nothing lets the job through — a preflight that guesses is a quieter failure than the one it prevents.
    ↘ go deeper
    Same pure decision module as the failure walk, pointed before spend (reroute-decide.py --preflight). Guard facts fail closed (the frozen security guard's law); capability facts fail open. The existing before-spend security swap in route.sh is untouched — one law per layer. The too-eager guard is a first-class twin arm: the proof that a silent card DISPATCHES ships alongside the proof that a filtered card refuses.
  3. Through a door: isolation, a clock, custody. A chaperone hands the brief through one of four doors. The engine works in a sealed worktree or a cloud runner — never the live checkout — under a hard timeout from its own card. Whatever it produces is committed to a bay branch by the door, not by the engine's good manners. A dead worker leaves no wreckage.
    ↘ go deeper
    Sealed room (run-sealed.sh, any engine with a card) · api lens (api-worker.sh, single-shot second opinions) · sandbox lane (Cloudflare containers, capacity builds) · codex cloud (their runner, our notary — diff custody, isolated clone, fetch-back). One ledger row per run, every lane, battery-proven.
  4. The typed word (#1 — shipped this morning): failure becomes a value. However the run ends — finished, starved, filtered, cancelled, door-sick — the door hands back one machine word from a closed set of seven, on every exit path. Prose never decides anything again. This is the socket the entire pipeline layer plugs into: you can only route around a failure that comes back as a value.
    ↘ go deeper
    bay-src/bin/door-result.py: {ok, category, reason, partial} with seven categories — complete, nothing_to_do, engine_failed, timeout, refused, cancelled, door_failed. The projection derives ok; downstream routes on the projection ONLY, never on reason text or report prose. All four doors emit it as of 0.6.119, live on this machine and published.
  5. The failure walk (#2, built tonight): the job moves itself. The word says the engine died. A pure decision module — facts in, decision out, recomputable by anyone — reads the word, the cards, and the attempt history, and names the next fit engine or stops honestly. A thin loop re-dispatches clean from the same starting point. The lead hears exactly one of two things: finished work, or one complete failure chain. Never a mid-flight death. Three handmade per-engine patches become one mechanism.
    ↘ go deeper
    What walks: engine_failed yes; timeout only to a strictly larger clock (same-size retry is the banned blind retry); refused only when a card fact explains the refusal and another family clears it; cancelled and door_failed never (an operator's stop is respected; a sick door surfaces — rerouting through it burns spend and hides bay breakage). Budgets: 3 attempts, each engine at most once, wall ceiling 2x. Partials: attempt N+1 starts clean, every partial's branch rides the record named out loud — your ratified amendment of the double-carry gate. The module gates spend, so it joins the dangerous-files list: every future edit demands a declaration and a non-author court.
  6. The pipeline record (#3, built tonight): the job's memory leaves the lead's head. Every attempt writes one line — engine, word, decision, the card facts read — into a record that is itself the job. A multi-step pipeline (build, court, cure, walk) becomes a recomputable object on disk. The crash that once lost a seven-gate deploy because it lived only in a session's context becomes structurally impossible.
    ↘ go deeper
    Grows from the reroute record's shape (staging/reroutes/<job>.jsonl — the intent names it the embryo on purpose). Design is being duelled tonight by two independent pedigrees, cross-refuted, then synthesized into a draft intent for your pen. Rubric: append-only, one writer, recomputable, cold-session resumable, zero settlement-side machinery — Quanta stays dry dock. Replay property: run the decision module over the recorded facts and get the identical decisions byte-for-byte.
  7. Cold resume: a session dies, the work does not. A brand-new session — no context, no memory of the job — reads the record and continues from the exact step where the last session died. The proof is the acceptance walk: kill a live pipeline after step one, hand a cold agent nothing but the record's path, watch it finish.
  8. The flywheel: recovery teaches prevention. A failure gets recorded with its cause. The cause becomes a fact on the engine's card. The next preflight reads that fact and refuses the repeat before spend. Recovery (#2, #3) feeds prevention (#4) through the cards — the harness gets harder to hurt every time it is hurt. No scores, no predictions, ever: facts recorded, facts read.

Fences that never flex

Code never decides meaning. The pipeline layer routes on typed facts; judgment stays with the lead. No door edits; a second routing truth is how the old dispatcher died.
No prediction. Card prose never enters fit; no success-likelihood numbers — a guess wearing a number is banned by name.
Security fail-closed. The frozen guard's word is the only authority; a filtered engine never resolves security work, at preflight or on the walk.
No cross-door rescue. A sealed job never silently becomes a sandbox job. Completed-but-lost work surfaces; rerouting it would double-carry.
Workers never ship. No engine pushes or opens a PR. Custody, review, and merge are deliberate acts above the pipeline layer.
Nothing grades itself. Every organ lands with a twin sabotage-proven to bite, and a non-author court on the diff.

Who is welding it in tonight

Codex cloud carries the heavy modules; grok builds every twin and the missing native cards, one seat, serialized; Kimi-K3 sits opposite every builder — refutation packets on each diff and one of the two rival designs for the record. Opus appears exactly twice, where judgment is load-bearing: synthesizing the two draft intents, and tie-breaking any court that splits. A cloud leg that dies reroutes once to the sealed room — the workflow practices the exact pattern it is building.

↘ go deeper — the court law
Three lenses per primitive: K3 hunts violations of the ratified laws; a non-author codex instance runs tight audits under the meta-adversary charter; grok runs the vacuity sweep — every twin's sabotage arm must bite, and an arm that cannot fail is itself a major finding. One cure round by default, a second only on continued refutation, then the surface is held for you — a second refutation indicts the frame, and nobody buys a third cleverer patch. Verdicts are testimony, never blended.

What only you can decide

Ratify the two draft intents

The pipeline record and the preflight land as panel-courted drafts with open calls listed. Their organs do not merge until you red-pen them. #2 is already your ratified word — its four calls are law tonight.

Any held court findings

A divergent or twice-refuted verdict is never merged on a shrug. If the morning packet carries held items, each arrives with both verdicts and its basis.

The spend-gate declaration

The decision module joins the dangerous-files list — every future edit to it demands a declaration and a non-author court. That addition rides the #2 pull request.

The lead sees exactly one of two things: a finished result, or one complete failure chain — never a mid-flight death.