Skip to main content
These examples assume an authenticated active saved profile with the intended organization and project. Confirm it with tp context show before running commands. Add root --json before the command when machine parsing needs JSON output; use root --profile <name> only to override the active profile. Run commands using the active saved profile:
When tp is not on PATH, run from thinkplatform-lib with .venv/bin/tp or pass --tp-bin /absolute/path/to/.venv/bin/tp to the collector. The collector auto-detects the repository .venv/bin/tp. All commands in this reference are read-only.

​
Context and profile

Confirm that org_id and project_id match the requested scope. Do not silently switch the active profile.

​
Cost and attribution

Use explicit dates:
Useful options:
  • --provider openai|azure_openai|azure|all
  • --metric ActualCost|AmortizedCost
  • --force-refresh when stale cached provider data would materially affect the result
  • --exclude-unallocated only for a focused view; retain unallocated cost in the reconciliation

​
Fallback for older CLI installations

The server must also support the requested grouping. If it does not, record the missing dimension as a limitation and use tasks, Agents, and provider breakdowns as secondary evidence.

​
Agent and task configuration

Only retrieve raw task messages after selecting an anomalous task:
Do not paste raw message content into the report. Report sizes, repetition, tool names, and short paraphrases.

​
Schedules and triggers

Inspect:
  • interval/trigger and enabled status;
  • target Agent and schedule prompt;
  • run count, status mix, attempts, repeated task IDs, and errors;
  • overlapping schedules, Hooks, Webhooks, and task follow-ups that can trigger the same work.

​
Data-context investigation

After message summaries indicate file-heavy tool use, inspect only the implicated paths:
Look for repeated reads of large or numerous files, raw PDFs/images requiring conversion, and missing curated indexes or summaries. Do not enumerate the entire Data Lake without evidence that it is required.

​
Snapshot collector

The default top-task and schedule summaries are bounded. Increase limits deliberately when the initial ranking shows a long tail that matters. Read collection_warnings, derived.dimension_coverage, and derived.collection_coverage before conclusions. Scheduler status_counts describe enqueue records, not downstream task outcomes; correlate them with task status.