Cassette Build Report 058 — The Live Plan Had to Fit a Human Life

Scope note: This report covers the redesign of Cassette’s live-testing plan. It proves that the queue became finite, inspectable, and structurally valid; it does not claim that the planned live operations ran.

Claude Fable gave me twenty-seven steps. I still saw five.

I had asked the agent to review Cassette’s closed work, review the live campaign, divide that campaign into sessions I could actually sit through, and create a review method for those sessions. The technical review was useful. It found hard-coded training shapes, an inexact comparison inside a supposedly exact certificate, a computed import that escaped confinement, and capacity tests that were already red beneath a passing label.

Then Fable organized the live work as L01.25a, L01.5b, L03g, and similar children of the five old phases. There were more rows, but the work still lived inside the same dense parents. A suffix had divided the list. It had not divided my time.

I corrected this three times because the distinction kept disappearing. I was not asking for smaller headings. I was asking whether I could begin one session, understand its physical and technical consequences, leave when the bounded work ended, and return later without asking the next agent to reconstruct what had happened.

This mattered because live testing would not occur in a sealed laboratory day. I would move among work, children, other projects, and Cassette. A model transfer could take hours. A compile could outlive the conversation that started it. A training job could pause while the agent session ended. If the plan required my continuous presence, it was not merely inconvenient. It invited the next agent to improvise from partial state.

Fable’s first repair made the identifiers flat and limited sessions to twenty minutes. It also exposed a small but telling error. The agent renamed steps with a substitution table. One old identifier was also a new identifier, so the fifth step became the twenty-seventh and the sixth depended on it. Fable reported that the script had succeeded, but its assertion output had gone to a file the agent never read.

Cassette had spent weeks learning that a test name is not evidence. The same rule now applied to the agent’s shell command.

Codex then reviewed whether the rewritten plan would work, not merely whether its YAML parsed. Several sessions still concealed multiple sittings. Transfer start and transfer completion shared one record. Training rows named three operations without giving each operation a lifecycle. Eighty fault cases had no fixed connection to their source artifacts, hardware profiles, approvals, or recovery chains. The names were short. The work was not.

The repaired plan changed its unit of account. A long operation now has a bounded START record and a later VERIFY record. The start proves that the operation was admitted, attached to the right drive and authority, and reached its first durable boundary. The later record examines the continuing or completed operation. The operation may outlive both agent sessions.

The queue grew to thirty-five fixed records plus eighteen families expanded from the data. That larger number did not enlarge the project. It disclosed work that had always been hidden inside phrases such as “run the protocol matrix” or “test failure recovery.” Every model, training operation, client mapping, physical profile, and fault coordinate becomes a literal record before execution.

The live runbook carries the human boundary directly:

An exact card is an evidence record, not automatically a permission request.
Read-only identity, health, capacity and coarse-inventory checks, documents,
plans and evidence preparation need no fresh approval. Ordinary operations
within the authorized campaign scope proceed without the principal present.

The passage appears in the public PHASE_LIVE_RUNBOOK.md. Its language resolves a second pressure in the work. A precise evidence card records what an operation may do. It should not become another reason to summon me. Once I authorize a campaign root and its ordinary scope, the agent may prepare cards, run contained operations, and bind evidence without pretending that every new digest received a separate personal blessing.

Review found other structural faults. Early bootstrap stages depended on review machinery they had not built. Review history could fork after a lease ended. Five bounded reviews could accumulate five repairs against one stale attempt. A symbolic link could escape a sealed evidence directory. Later graph revisions could erase earlier dependencies. Several write-bearing records had no approval lineage.

Those failures received mechanical answers. Bootstrap review uses a small off-tree record until the normal system exists. Every review atomically advances one current head. One accepted repair ends an attempt and creates a new child. Evidence bundles reject links that escape their directory. Graph revisions may add authorized nodes and edges but may not rewrite inherited ones. Every write names the card that covers it.

At the end of this work, the fixed graph was acyclic. All expanded families named an evidence level. The failure partition was complete and disjoint. Three read-only reviews found no missing obligation or hidden multi-session record. The wider suite still had eleven failures, all visible and assigned. No drive had been mounted, written, detached, or faulted. No source account had been contacted. No model had been downloaded.

The plan was ready to begin because it finally admitted that my absence was normal.

This is not a concession to a distracted operator. It is a requirement for durable human-agent work. If one conversation must remain open and one person must remain attentive, the repository and operation records do not contain enough state for the next agent to continue correctly. Cassette needed to preserve that state before I left.

The agent had initially grouped related work. I needed it to divide responsibility. The accepted plan now gives each session a bounded action, durable record, stop condition, and later verification point, so another agent can continue without reconstructing an absent person’s attention.