The repository as a continuous factory

The claim is that a class of work — automated, repetitive, collaborative, auditable — belongs in the repository as a scheduled loop rather than as a human sitting down to do it. The reference implementation is markdown-plus-front-matter workflows hardened into CI actions: a maintainer-shaped assistant that wakes daily, reads its own memory, and works a task list. One dormant project with a multi-year backlog went to burn-down plus three major releases; one narrow refactoring workflow landed 112 merged pull requests out of 142 opened.

The vocabulary that comes with it is the useful part. A repo is blocked, flowing or idle, and blocked-on-human is a legitimate state rather than a failure to route around. When work jams, the move is to add a quality gate to the factory, not to hand-fix the jammed instances — which makes the maintainer a designer of flow rather than a fixer of items. And the safety/productivity tradeoff inverts relative to interactive coding: “like a rail track — the better the quality of the rails, the faster your train can go”. Guardrails become throughput infrastructure rather than a tax.

The full loop is signals → triage → planning → coding → shipping → back to signals. The argued reason the promised productivity does not appear is that organizations automate the coding slice and leave triage, planning and validation synchronous and human — and triage in particular is almost never represented explicitly in any system at all. Meanwhile the observed bottleneck, in every account, is review: teams can produce faster but not review faster, with congestion collapse predicted by straightforward theory of constraints.

Brownfield is most of enterprise software, and it is a city rather than a ball of mud. The consistent advice is to generate an evidence-backed map of what the code actually does — dead code, duplication, complexity hotspots, dependency graphs — before migrating anything, because documentation has rotted and tribal knowledge cannot be verified. Unguarded agents “cause havoc through over-optimization”, producing code that passes every test while quietly undermining implicit architectural constraints. One social side-effect is worth its own line: AI-generated analysis, not written by the cleverest person in the room, turns out to be more contestable, and broke the bike-shedding bottleneck.

Also: GitHub Agentic Workflows; the “Impact of Automated Repository Maintenance Assistance” report; Agent Zoo / Agent Factory; Hud; Adam Tornhill, Your Code as a Crime Scene; SonarQube; Ink & Switch’s “malleable software”; Pi.

The four desk roles — intake, worker, PR review, verify — are Continuous AI loops over repository state, each running on its own cadence. What the talks supply is vocabulary Assay did not have: safe outputs, flow states, rails-enable-speed, the flow-designer stance. Its operating rules already encode the same positions, arrived at independently: BLOCKED-ON-HUMAN is the blocked-on-human state under another name, and “a red check is your work item, never a wait state” is the anti-jam rule stated from the worker’s side. The signals-to-triage gap is intake-desk’s founding argument arriving from outside, and shift-left is the same instinct as authoring verify rows with the brief rather than after it.

The review bottleneck is where Assay has spent the most. Reviewer verdicts, scan lanes and the post-merge verify loop all exist because review, not generation, is where the factory jams — and because a reviewer that is a different principal from the implementer is the only version of the check that means anything. The lifecycle is designed around that constraint rather than around generation throughput.

First, the flow-designer distinction is not measured. There is no counter separating “systemic issue filed plus guard added” from “instances hand-fixed” — which is exactly the ratio the factory framing says predicts whether a factory improves. Second, brownfield intake: forensic-sweep output becoming the fact lines on a brief, so a brownfield brief is evidence-backed by a generated map rather than tribal memory. That is a proposed authoring mode with nothing implemented, and the social finding — that generated analysis is more contestable than a senior engineer’s opinion — is the argument for it that Assay had not articulated.

The production-diagnosis half of the loop sits outside the current scope: the desks operate pre-merge, so an ops desk is a design seed, not current work. One part of that direction is rejected outright — proactive agents messaging humans is exactly the noise the desks-run-silent rule exists to suppress; state belongs in claims, boards and filed issues. On self-extension the diagnosis is adopted and the prescription is not: agents genuinely do reason better in code, so desk verbs should stay composable, scriptable and predictable — clean exit codes, machine-readable output — rather than accreting interactive cleverness. But the verbs are an authority boundary, not a usability layer, and end-user runtime extension collides directly with the trust gate.

• Whether anyone reports the jam ratio as an actual metric — gates added versus instances hand-fixed — rather than as an anecdote.
• Whether a second group publishes repository-assistant numbers at that scale (112 of 142 merged, a dormant repo restarted) for a cadence loop they did not build themselves.
• Whether triage gets an explicit representation anywhere outside intake-desk-shaped systems, since it is claimed to be the universally missing layer.