Seven harnesses took the transcript. None of them took the approvals.

The loss manifest is the part worth reading twice, and not for its size. Across seven separate targets the dropped totals sat in a band three items wide, 54 to 57 ^[5][2]. If target design governed preservation, picking a better destination would buy something measurable. It buys three records. The author's reading is that the source session's own event vocabulary decides what can enter the intermediate model at all ^[7], and the documented design agrees: the validated event timeline in the middle is deliberately narrow, wide enough for ordered conversation events and closed to client-resident configuration ^[8].

That is where the approval story ends. Approval gates and audit settings sit on the client side of that boundary, which is why the transcript arrives and the controls do not ^[2]. It is not a gap that a later release closes by adding a field.

Private thinking makes the size of the hole legible. The frozen source carried 11 thinking blocks ^[3], and every one of the seven targets dropped 9 of them ^[5], roughly 82 percent ^[1]. Tool-use and tool-result blocks numbered 20 and 20 ^[3], so the record of what the agent touched survives better than the record of why it decided to. For a developer resuming work, that trade is fine. For the question the author says keeps arriving at engineering leadership, "three months from now, where can we see why this agent made that change" ^[10], it is close to a total loss.

Note what the manifest itself is: a count of omissions and transformations carrying no content ^[8]. It can tell you that 57 things did not survive. It cannot tell a reviewer which 57, which makes it a migration receipt rather than an audit record.

The contrast the author draws is with Slack Code, and it is Slack's own account. Slack says a coding-agent mention can create a channel that gathers the relevant people, the diffs and the planning documents, then archives as a searchable record ^[11], and that important changes can be routed to a person for approval inside that channel ^[12]. The mechanism, not the product, is the transferable part: the durable record and the approval event live somewhere that already has an administrator, instead of being stood up again inside whichever harness a developer adopts next quarter.

Two caveats on the numbers. Freezing the session was not fastidiousness; the live file kept growing during the first attempt, so each conversion was being measured against a different source ^[14]. And the run happened in a Python 3.12 environment on macOS while the README claims Python 3.11 or later and Linux ^[13], so this is one sample, taken outside the stated support boundary. The shape holds even if the decimals do not: seven destinations produced one loss profile, and the thing an auditor would ask for was in none of them.