Skip to content

fix(standard): redact the prepare record and report a failed prepare action - #236

Open
pierreleca wants to merge 6 commits into
ThinkFlowLab:mainfrom
pierreleca:fix/computer-port-prepare
Open

pierreleca wants to merge 6 commits into
ThinkFlowLab:mainfrom
pierreleca:fix/computer-port-prepare

Conversation

@pierreleca

Copy link
Copy Markdown
Contributor

Closes #235

What

Three repairs to what #224 added to agent.computer/v1. The prepare record now goes through the repository's redaction, as the relay's records do: a credential-shaped value in an action or a file name is written as [REDACTED]. A prepare action that fails now answers with an error that names it, and the actions after it are not run. On main the call answers as a success and goes on. The kit's late reference now finds the button in the observation it is sent, and its test fails on blank pictures. The port's page states both changes of prepare, and the kit's words say that late does not bound the number of presses. The kit's controls are recorded again, on the desktop that main builds since #220 (Pillow 13.0.0). The kit stays at version 2.

Touches

backend (the standard: the port's server) · catalogue (the kit's reference, its words and its record) · docs and site (the port's page).

Verified

  • The suites, each module in its own process: 3,016 backend and 661 route tests. The three modules that fail also fail on main in the sandbox used here.
  • Each new test fails on feat(standard): reach the desktop through acting calls alone once the task is given #224's code and passes here.
  • prove on the kit, on a desktop image that holds Pillow 13.0.0: the null control 0 of 9, the reference 9 of 9, a control that clicks without looking fails late.
  • A scripted run of an agent outside the catalogue that starts its own driver through prepare: 18 prepare calls on the kit's nine units, none in error, the same calls as before the change. It ran on the same port and kit code, before this branch moved onto today's main.
  • The wheel builds and its verify step passes; the generated pages are current.
  • Not verified: a run with a real model on this head.

Risk

An agent whose prepare action failed unseen now gets an error before its task. An action that looks like a credential is recorded as [REDACTED]; the desktop still receives it unchanged. Two limits are kept: a failed action's error holds the desktop's whole error output, and a step's action that raises is still no error. Undone by a revert: no stored data changes and nothing migrates.

… reference

The kit's own proof, taken again once the reference for late finds the
button in the observation it is given: the nine units' ids, the same
twice and unique; no reference found in the unit's prompt (9 of 9); the
null control passing 0 of 9 checks and the reference 9 of 9. Besides the
two control jobs' names, the leak check's counts follow the references
this run drew: 2 looked for and 7 too short, against 1 and 8 before.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Standard: on agent.computer/v1 a failed prepare action answers as a success, and its record keeps keys in clear

1 participant