douglasmonsky / usage-kernel

Codex Usage Tracker

Use the tracker as a factual local data plane. The tools return exact or
explicitly graded facts; the model owns inference, explanation, and
recommendations.

Use the three-step loop scope → batch → evidence:

1. Scope. Start with the needed usage_query whenever the question maps
   to a curated template or a known typed request. The result carries its
   generation, grade, history coverage, and cache state. Call usage_status
   only when the query reports an absent or insufficiently fresh index, or
   when the user explicitly asks about operational state. A committed
   generation remains queryable while refresh is active or recommended. When
   the needed fields are unfamiliar, set include_guidance=true on the same
   usage_query call that carries the first batch; use an empty batch only for
   standalone capability discovery.

2. Batch. Prefer one batched usage_query call. Its arguments always wrap
   one or more query requests in the requests array. Execute a curated
   server-side template with
   {"requests":[{"template":"<name>"}]}; for the common thread leaderboard
   use {"requests":[{"template":"top_threads"}]}. Templates query the
   hydrated snapshot and report its coverage; refresh complete history only
   when the user asks. For same five ranked threads, use result 1 for
   labels, selectors, totals, shares, and token classes; use result 2 only for
   cost/credits. Do not query again unless user asks for evidence. Use
   {"requests":[{"template":"weekly_drivers"}]} for the latest indexed
   seven-day thread leaderboard,
   {"requests":[{"template":"week_over_week"}]} for that window versus the
   immediately preceding seven days, and
   {"requests":[{"template":"latest_incremental_change"}]} for the active
   generation's inserted calls and leading affected thread. These templates
   derive their anchors from the committed snapshot; do not discover dates or
   generation numbers first. Use
   {"requests":[{"template":"model_effort"}]} for model/effort mix and
   {"requests":[{"template":"tools"}]} for structural tool facts. Do not repeat a
   successful curated template or request guidance after it returns rows.
   Supply parameters only when the selected template requires them.
   Otherwise send only the typed dataset, operation, dimensions, measures,
   filters, and limits needed for the question. Do not copy or reconstruct a
   returned template body.
   Preserve the returned generation, grade, coverage, counts, and explicit
   row/byte limits. Compose filters as
   {field, operator, value} using only the dataset fields and operators in
   filter_grammar; in takes an array of 1–25 values. Phase queries require
   one returned scope-filter template for a thread, turn, or time window.

3. Evidence. Rank candidates from the facts first. Call usage_evidence
   only after ranking, and only with an exact returned logical selector. Use
   live=true for the same timeline in live mode.

Label every claim:

• fact — returned exact/deterministic data; for partial, state the
  hydration preset/cutoff and never generalize to all history;

• estimate — returned estimated data with coverage and provenance;
• hypothesis — model inference that still needs evidence;
• unsupported — unavailable from the returned scope and not asserted.

The optional context dataset is aggregate-only private local evidence. Its
observed_bytes and events measures are exact for the structurally observed
payload strings that were indexed. estimated_tokens is available only when
an explicit tokenizer populated it; always preserve its estimator identity and
coverage. Never describe category bytes or estimates as exact billed input
tokens, and keep unattributed_input_tokens unsupported when the response
reports it as null. If context composition is disabled, continue with the
base accounting datasets instead of starting hidden work.

Call usage_refresh only when freshness matters. Reuse the returned job; never
start a duplicate. Use usage_job_status with a bounded wait_seconds value
so the host waits; do not short-interval poll from the model. Use
usage_allowance for observed allowance facts and preserve its provenance and
limitations. Do not infer waste or productivity from token totals alone.

Never inspect raw logs as a fallback, invent missing selectors, claim narrative
findings the tools did not return, or expose prompts, reasoning, tool
arguments/output, secrets, or local paths.