Cloud Intelligence™ CLI v2, built for humans and agents alike

Half the invocations of a modern CLI come from an AI agent, not a keyboard. `dci` v1.x was designed before that was true, and it showed in both directions: humans got output formatted for machines, and agents got output formatted for humans. v2.0, shipping today, rebuilds the output layer, the error handling, and the machine contract. The commands and the API are unchanged.

Run a monthly cost report on CLI v1.x and you got something close to a flat database dump. One row per group per time period, raw floats like 239927.13841529994, and epoch timestamps. A report in euros looked identical to one in dollars, which at current exchange rates is an easy way to misreport spend by 8%. Listing anomalies squeezed sixteen columns into your terminal width, truncated every cell, and hid the one column you needed, the date.

AI agents had it even worse. Typo a command and the CLI printed 258 lines of help, then exited 0, so the script kept running as if nothing happened. Put an unknown field in a query body and the CLI silently ignored it and returned default report data unrelated to your SQL, which an agent would present as the answer with full confidence. A successful delete came back wrapped in a "retryable" error, so agents retried an operation that had already succeeded. And a 30-day query grouped by service could dump 7,500 rows, roughly 120,000 tokens, straight into the agent's context window.

What humans get with CLI v2

Reports pivot by default. Groups by time, currency-signed totals, ranked by spend, with a trend column, no flag needed. The three-month, four-service report that used to be 12 interleaved rows of floats is now a table you can paste into a budget review. Numbers carry their currency symbol, so a EUR report finally looks like one. If you script against flat rows, --flat keeps them, with readable dates and aligned columns, and --raw-numbers restores exact values when you need them.

Listings got the same treatment. Identity and date columns always survive the width fit, hidden columns are named along with the flags to bring them back, and nulls render as empty cells instead of a literal <nil>.

Mistakes fail fast. dci list-bugets returns one line, "did you mean list-budgets?", with exit code 2 your script can act on. An invalid customer context is rejected the moment you set it, not three commands later as a mysterious 403. The same failure maps to the same exit code in every mode.

Two additions connect the terminal to the rest of your work: dci open report <id> deep-links into the console, and dci query --output csv exports straight to a spreadsheet.

What agents get with the upgrade

For AI agents, the changes come down to trust and cost. Unknown request fields now fail with a structured error and a hint pointing to the schema, so a malformed query can no longer return plausible-looking wrong data. Deletes that succeed report success, and client errors are never marked retryable, which ends the retry loops. Running without credentials in CI returns a structured error in about 10 milliseconds instead of hanging on a browser login that will never open.

Token costs dropped across the board. Large results are bounded at 500 rows by default, with the truncation stated explicitly and instructions to narrow the query or raise the limit. That takes the 120,000-token response down to about 9,500, a 92% reduction. --help output shrank 84%, with the full request and response schemas one flag away via --help-full. And --rows keyed returns schema-named objects with an explicit currency field, so nothing has to count positional array cells to find the cost.

Try it

dci upgrade                # get v2
dci skill <agent>          # refresh the embedded agent skill
dci query --help-full      # the schemas are still there when you need them

The full before/after comparison, with real output from both versions, is in the Help Center, alongside the What's new in v2 brief.