cargo-diagnostics

Explain what a Cargo run or batch actually did, after the fact — trace one run node by node, draw the graph it executed with the failing step marked, sweep a batch or play for errors grouped by root cause, and attribute credit spend down to the node and the provider. Triggers: "why did this fail", "it succeeded but the output is wrong", "half my rows are empty", "why is this column blank", "what broke in this batch", "why did that cost so much", "which node is burning credits", "it worked yesterday", "these results look wrong", "it went down the wrong path", "this step never ran", "show me what the run did". Skip when: setting up an alert for next time — use cargo-observability; just downloading the data — use cargo-analytics.

Cargo CLI — Diagnostics

Forensic runbooks for workflow behavior: trace one run, sweep a batch for errors, profile a play's credit spend. This skill is the interpretation layer — the raw surfaces (run get, orchestration SQL, billing metrics) are documented in cargo-orchestration and cargo-billing; each runbook here tells you which of them to pull, in what order, and what each output shape means.

Bootstrap

Already signed in (cargo-ai whoami returns a workspace)? Skip to the next section.

bash

npm install -g @cargo-ai/cli            # no global install? prefix every command with `npx @cargo-ai/cli`
cargo-ai login --email [e-mail]  # emailed code, no browser; creates the account on first use
                                        # alternatives: --oauth (browser) · --token <api-token> (CI)
cargo-ai whoami                         # confirm the active workspace before any write

Every command prints JSON to stdout; failures exit non-zero with {"errorMessage": "..."}. Anything that creates a run or a batch is async — pass --wait-until-finished or poll the matching get. Credit attribution steps (billing usage get-metrics, billing subscription get) need a token with admin access; everything else works with a standard token. When the full skill bundle is installed, ../cargo/references/prerequisites.md adds the CLI version pin, token scopes, and the admin-only surface.

Which runbook?

text

What are you diagnosing?
│
├── One run / one record ("why did this record fail?",
│   "run succeeded but the output is wrong/empty")
│   └── references/run-trace.md
│
├── Many runs ("the batch has errors", "error rate spiked",
│   "which node keeps failing?")
│   └── references/batch-error-sweep.md
│
└── Cost ("this play is expensive", "where do the credits go?",
    "make this cheaper")
    └── references/play-optimize-credits.md

Rule of thumb: start with the sweep when you don't yet know which run to look at — it ends by handing you exemplar run UUIDs to feed into the trace.

No run UUID at all ("look at the last run", "the run for acme.com", "what did my play do in the editor")? That's references/run-trace.md § 0, which resolves a symptom to a UUID. Note that orchestration run list requires --workflow-uuid and cannot answer it — orchestration SQL over runs takes no filter and can. Never conclude that run inputs and outputs are inaccessible because run list refused.

Boundary with cargo-analytics: analytics measures and exports ("what's the error rate?", "download the batch results", "export this segment"); this skill explains ("why is the error rate up?", "why is this record's output empty?"). A diagnosis often starts from an analytics signal (error count spiked, batch reports failedRunsCount > 0) and ends back in analytics — once the cause is fixed and runs re-executed, bulk retrieval goes through run download-outputs / batch download / segment download, all documented in ../cargo-analytics/SKILL.md. This skill's evidence surfaces (run get, orchestration SQL, billing metrics) are for diagnosis, not bulk export.

References

Doc	What it covers
references/run-trace.md	Find a run from a symptom when you have no UUID (§ 0), then walk it end-to-end: per-node executions, runContext outputs, branch routing, per-node credits and timing.
references/batch-error-sweep.md	Find errored runs across a batch/play/workspace, group failures by root cause, pick exemplars, decide fix vs report.
references/play-optimize-credits.md	Attribute credit spend to workflows and nodes, then apply the cost levers in priority order.

The surfaces every runbook draws on

Surface	Command	Gives you
Run detail	cargo-ai orchestration run get <run-uuid>	run.executions[] (node-by-node trace), runContext (per-node output keyed by nodeSlug), runComputedConfigs (what each node was actually called with)
Orchestration SQL	cargo-ai orchestration query execute "<sql>"	Aggregates over runs, batches, spans, records (ClickHouse; no schema prefix; workspace-scoped)
Billing metrics	cargo-ai billing usage get-metrics --from <date> --to <date>	Credit totals, filterable and groupable by workflow_uuid, connector_uuid, agent_uuid, integration_slug, model_uuid
Graph picture	cargo-ai orchestration node diagram --run-uuid <uuid> --highlight <slug> --format ascii --raw	The graph the run executed, with the failing node marked. Free, runs nothing

Draw the graph before explaining a routing bug. For "it took the wrong branch" or "this step never ran", the picture is the evidence, and it shows one thing run get does not make obvious: the on failure edges. A step that looks skipped is often one the run reached via a fallbackChildUuid edge, which means the provider errored rather than returning nothing — a different diagnosis with a different fix. Flags and the ASCII legend: ../cargo-orchestration/references/node-diagram.md.

Full query syntax, table columns, and caps: ../cargo-orchestration/references/examples/queries.md. Debugging field semantics: ../cargo-orchestration/references/troubleshooting.md.

Presenting findings

Follow ../cargo/references/interaction.md: lead with the conclusion ("18 of 20 failures are one cause: the connector's token expired"), summarize evidence in a short table, never dump raw run get JSON or full query results into the conversation. Any fix that re-runs paid nodes goes through the pilot gate: re-run 10–20 records first, report the observed cost and hit-rate, then ask the user to approve the rest quoting the record count and credit estimate — a diagnosis is not approval to re-bill the batch that produced it. Full spend rules in ../cargo-gtm/references/cost-discipline.md.

When diagnosis dead-ends

If the evidence contradicts documented behavior (a field missing from run get, a query cap that doesn't match the docs, an error that makes no sense), file a report — that's the official channel and the team reads every one:

bash

cargo-ai workspaceManagement report create \
  --title "<one-line summary>" \
  --description "<commands run, errorMessage verbatim, expected vs actual, UUIDs>"