An action judgement holds at propose and wait is no longer a dead end. The
new approvals module owns the request: it binds to the decision and the
action's digest, so an answer covers the action the person was shown and
nothing else, and it settles once. The answer wakes the run, which judges the
action again and runs it only as the approval allows: a refusal still refuses,
an altered input runs nothing, and a judgement that came back worse runs
nothing on that approval.
The composer carries the request as a message with the request id and its
deadline, and its answer is a message too, so the trail shows what was asked
and what was answered. A request nobody answers inside its deadline is settled
as expired and nothing runs.
How long a request waits is an organisation setting, edited in the dashboard
and stored in milliseconds, because a Postgres interval scans as text. A run
pins the deadline it started with, like its model.
An agent that proposes the same action twice is asked about it once, and the
notification says when the investigation proposed nothing at all.
It also said its tools only read, which stopped being true when actions became tools. Judgement decides whether an action runs, so calling one is the proposal.
A new judgement module assesses one concrete action: hard rules refuse before
any guard runs; every guard sets a ceiling and the action runs at the lowest;
a guard that fails, panics or gives no ceiling holds it at propose and wait;
an advisory verdict is recorded and sets no ceiling. Reversibility and
Precedent (the fresh-install class default) are the first guards. Every
decision is recorded with its verdicts and class, and the run's trail lists
them. Agent tools now offer connector actions too: each call is assessed and
executed only as decided and only with the judged input, and a held action
comes back to the agent as a failed call naming its level and findings.
Tools gain Resolve, so an MCP server's declared effect reaches judgement.
Since the MCP migration made accepted.tools NOT NULL, the connector and secret
engine accept path never set it, so re-accepting any connector failed with a
not-null violation.
Killing the binary mid-investigation and restarting it failed the run: the
first model request after the resume didn't match the recorded one. The tool
schemas pinned in a step reached the live attempt as written but the resume
as stored, and both the v2 encoder (compacting raw JSON) and jsonb (reordering
keys) rewrite them. Step now returns the decoded recorded value on the live
path too, and runs.steps.output is json, which keeps the text as written.
An alert about a filesystem carries instance and mountpoint, and those are the
labels the forecast query needs, so the prompt names both and gives the query
that answers what an operator asks: the seconds until the disk is full, from
predict_linear and deriv over the same window. A negative figure is the
filesystem not filling, which the prompt says to report as that rather than as
a time.