The v1.36.0 release note carried four things a reader cannot use. The dashboard's
seven in-browser tools and the installtype fix stay; the withheld judgement verbs,
the consequentialHint limitation and the two silent-failure notes are removed from
the release feed and moved to where an implementer looks.
- The four verbs are withheld because an agent that can approve can approve its
own work. That is settled design, so it now sits in docs/architecture/identity.md,
beside the approval-path conflict it belongs with, stated as the constraint it is:
attribution, not capability, is what releasing them depends on.
- The browser-tool surface had no home in docs/architecture, so it has one now:
docs/architecture/dashboard-agent-tools.md holds what a tool call can be
attributed to, why confirmation is a signal rather than a control, and the two
ways the surface fails silently.
Audited 1.35.0, 1.36.0 and 1.37.0 for the same class. 1.37.0 is clean. 1.35.0 carried
two build notes: a bullet recounting that a guard had never been written because it
failed, and a clause about a manager "finally" having a production writer. Both are
gone; the capabilities they were attached to remain.
RELEASING.md now carries the rule so the next release reads it: release notes state
what a reader gets, and design constraints, withheld capabilities and failure modes
belong in docs/architecture/, because a note about an absence reads as a feature to
someone who never had the thing.
docs/news/** is regenerated by task news, never hand-edited. task check passed before
this commit: the chain runs it, so a failing gate commits nothing.
2.1 KiB
The dashboard's tools for an in-browser agent
The dashboard registers in-browser tools an agent can call through the page's model
context. Four are read-only — list_tasks, get_task, recent_events,
daemon_health — and three recover stuck work: retry_task, stop_task,
cancel_task. Each wraps the same ui/src/lib/api.ts client the dashboard itself
calls, so an agent and a person exercise one path rather than two.
The four judgement verbs are withheld for the reason identity.md gives: an agent
that can approve can approve its own work.
What a tool call can be attributed to
A page can invoke its own tools, so an agent's call and a page script's call reach the server identically — the difference does not exist on the wire, and no header recovers it. A tool call is therefore attributed to the signed-in identity the request presented, with its source recorded as a hint; it never records that an agent acted, and it never records that a human approved. An agent's action is attributable only on a surface where the agent presents its own credential.
Confirmation is a signal, not a control
The three recovery tools require an explicit confirm: true, which makes a
consequential call deliberate rather than incidental. It is not a security control:
an agent can trivially satisfy it, and the browser cannot enforce
consequentialHint at all — Chrome accepts it at registration and omits it from
getTools(), so no client can see what the browser does not surface.
The boundary that holds is the session. These tools cannot exceed what the operator signed in at the provider is already permitted to do.
Two ways the surface fails silently
- A secure context is required. On a plain-HTTP origin
document.modelContextis absent even with the flag enabled, and nothing reports why.mkcertcovers local development. - The dashboard embeds
ui/distinto its binary. Rebuilding the distribution is not deploying it: a running process can serve a bundle without tools while the built one has them, or serve a bundle older than its own binary.