@@ -8,6 +8,28 @@ Notable changes to Cloudfall. The format follows
88
99### Added
1010
11+ - ` cloudfall agent investigate --alert TEXT --base-url URL --model NAME `
12+ lets a model work an alert through the operations catalog, over any
13+ OpenAI-compatible endpoint (the API key from ` --api-key-from-env ` ,
14+ ` --api-key-from-file ` or ` CLOUDFALL_API_KEY ` ). The model sees the same
15+ operation tools ` mcp serve --repository ` gives any client: a read
16+ operation runs and returns what the hosts reported, any other operation is
17+ proposed in check mode and recorded as a decision, and no tool approves
18+ anything. A call repeated with the same arguments is answered from the
19+ first result instead of running again. The investigation is recorded in
20+ ` investigations/ ` as an ` AgentInvestigation ` : the alert, the model, the
21+ tokens, every call and how it ended, and the finding, with any decision id
22+ the model named but never recorded kept apart under ` unrecorded ` . Exit 92
23+ ` MODEL_UNAVAILABLE ` when the endpoint refuses, exit 93
24+ ` INVESTIGATION_INCOMPLETE ` with the record when the model runs out of
25+ ` --max-turns ` or does not answer in the asked-for shape. Proven on a
26+ disposable host with Nemotron 3 Super on Nebius Token Factory: a full disk
27+ behind a PostgreSQL alert was traced to unrotated application logs in five
28+ turns, and the two proposals it recorded, once approved, brought the
29+ database back
30+ - An operation tool of ` mcp serve --repository ` describes a read operation
31+ as running and returning what the hosts reported, instead of saying it
32+ runs check mode only
1133- ` operations propose ` returns a read operation's play output in
1234 ` data.output ` , so an agent acts on what the operation reported without
1335 opening the decision's ` .diff ` file; the record keeps the file as before.
0 commit comments