The bounded session

The plan card says the first piece of the export work is one session, and /coldstart is how you open it. A bounded session is the agreement that makes that piece finishable: one observable outcome, a stated inside and outside, the decisions that stay yours, and the checks that could establish the result.

Inside that agreement, ordinary choices may be delegated when current evidence settles them.

Prepare one finishable problem, delegate the evidence-settled work inside it, and keep product intent and consequential judgment with the human.

The ordinary alternative

You tell the agent to improve exports and watch what it does. That works, and for a one-file fix it is the right amount of ceremony.

At this size it gives the agent too many jobs at once. "Improve exports" permits redesigning filtering, adding PDF output, introducing a worker queue, migrating old order statuses, or exposing fields that merely happen to exist in the database. You find out which one it chose after it chose.

The agreement, for this export

Every row is a field of the session entry the planning pass wrote. There is no separate agreement document to keep beside the work — the plan card is the agreement, and /coldstart states it as the session's scope before touching anything.

That agreement delegates execution without handing over every decision the work might reveal. The agent can read the app's current patterns, follow the serializer already in use, write the endpoint and run the six checks. What it does not do is stop to ask you before each reversible edit the plan already settled.

The approved customer fields are the exception, and the reason is worth being exact about: the database may contain more fields, but their existence is not permission to export them. Planting that decision before execution is what keeps you responsible for what leaves the application.

The two exclusions are excluded for different reasons. Historical order IDs are useful adjacent work that is not part of the approved result, so they get a later owner. The background-job idea is rejected for this change because it would introduce a worker, retries and monitoring — a rejection on the record, not a silent downgrade to "optimization".

When there is no route to open

/coldstart does not improvise one. A new project may have no pointer at all, a pointer may name no active section or session, the plan it names may be missing, or that session may already be closed. In each case the command names the piece of state that is missing and sends you back to the planning pass or to the close that never ran, and stops there. It does not assemble a plausible session out of filenames, recent commits or the last conversation, because a route invented that way reads exactly like one you agreed to.

Prepare a result, not an ambition

A finishable problem has one coherent outcome that can be observed and checked. It does not have to be one prompt or one feature: an audit, an investigation, a migration phase, a refactor, a research pass or a cleanup can all be finishable. What matters is that the session can say what should become true and where its responsibility ends.

The one-sentence goal is a useful sizing test. "Add CSV export and redesign filtering" is two halves wanting different evidence and different finish tests, which is a signal to split — a signal, not a grammatical law. Some coherent jobs contain the word "and", and some oversized jobs hide behind a short sentence.

The checks also have to be capable of disagreeing with the plan. A check written to confirm what the session already intends to do only decorates it.

Put human attention at the judgment seams

Continuous supervision spends human attention on reversible mechanics. Fully unguided work fails the opposite way: the agent invents product intent because no boundary told it to stop.

A seam is a point where a decision with real consequences sits inside a change that looks ordinary. A prepared session separates three kinds, and only one of them is mechanical.

At a recognizable seam, Coldstart can bring the pending action back for a human decision, and it can refuse selected actions that violate the session's declared mode. Only one of the two modes restricts: mode: prep refuses writes outside the section being planned, and any command that is not provably read-only. mode: build — the one this session runs under — adds no restriction of its own. It is not unrestricted: the risk floor still applies under it, and that is a separate mechanism from the mode contract. A clean call produces no such decision. That silence means one thing only: no rule the product carries matched this action. It is not proof that the action is harmless or intended.

A contextual seam has no such handle. A change to what customers can export may sit inside ordinary-looking application code. If the existing query behaves differently from the documented filter contract, the agent first asks whether current evidence settles the difference — adapting an implementation to a settled contract can stay ordinary work. Changing the exported result is a product and exposure decision, so you own it even though no mechanism recognizes the seam.

Give discoveries a durable disposition

Real work rarely follows a plan exactly. What matters is that each discovery gets an owner rather than a silent decision.

The agent may keep working on independent parts, but it must not build on a guessed answer.

Suppose it discovers that a large account cannot finish the export within the agreed synchronous limit. A background export may well be worth preserving — but adding a worker, retries, monitoring and new failure modes would change the responsibility and the evidence this session was prepared against. The session enforces the agreed limit and queues a separate design. It does not quietly become a queue-architecture project.

Verification earns the close

The plan names verification before execution so "the code is written" cannot quietly become the finish condition. Each observation has to be specific enough to fail, and an unrun check remains unrun. Running those six checks and deciding whether the result can close is a separate responsibility, and the next page is where it happens.

Limits

Preparation narrows responsibility. It does not guarantee success, eliminate mistakes or make the boundary impossible to miss. The agent can misread evidence, a plan can be stale, a check can be weak, and some judgment cannot be recognized mechanically. You stay responsible for intent, exposure, consequential scope and irreversible decisions.

Coldstart has not demonstrated a general productivity, quality, or error-rate gain from this method. For a small and obvious change, the planning and closeout ceremony can cost more than it saves. The heavier workflow earns its place only when recurring drift, risky seams, adjacent work, or multi-session handoffs make the boundary valuable.

Next

Running the named checks and deciding whether the work can finish is the next stage: verified close.