Reduce repetition, not verification. Context size, cached input, output and reasoning usage are different measurements; fewer bytes are not proof of a lower bill or unchanged answer quality.
Where usage actually goes
Measured from one heavy user's local logs (2026-10-10, ~1,100 sessions across
Codex, Claude Code, Hermes and Grok). Run python TOOLS/usage_report.py for
your own numbers before changing anything.
- Requests can include substantial prior context. In one historical Codex log sample, cumulative input divided by tool-call count was ~124k; this is not a measured median request size or an uncached charge. Stable instructions, skill indexes, MCP schemas and large outputs can affect subsequent requests; caching, deferred loading and compaction determine their actual cost. Keep those small; end a phase by updating STATE.md and continue in a fresh session instead of one marathon thread.
- Codex approval reviews were ~23% of Codex usage: each escalated command
started a reviewer that re-read the transcript. Fewer escalations mean fewer
reviews: work inside the writable roots, batch commands, and let the user
decide whether to widen
sandbox_workspace_write.writable_rootsfor folders they always work in. Never change sandbox or approval settings yourself. - Forked subagents were ~12%: a fork re-reads the parent history. Spawn only for independent work and give each a narrow written brief.
- 71 sessions were resumptions after a usage limit. A current STATE.md makes a
resume a short read instead of a re-investigation (
state-resumption).
Wait once, report meaningful changes
Use the host's background/notification mechanism for long builds and CI. Polling the model repeatedly costs inference even when nothing changes. Time spent waiting in a process does not itself generate model tokens, though tool notifications, hosted execution and later model calls may have costs.
For GitHub releases, record the exact pushed SHA, discover its required runs,
and wait with TOOLS/Wait-External.ps1 -GitHubRun <run-id>. It runs
gh run watch <run-id> --compact --exit-status and, on failure, fetches
gh run view <run-id> --log-failed once. Verify every required check
succeeded on that SHA before publishing.
No runs, failed/cancelled runs, or a different SHA are not success.
Respect the host's wait limits and keep the user informed without busy polling.
Cache evidence, not assumptions
Cache rules and prices depend on the provider, model and billing plan. Do not apply one API's rates to another API or a subscription quota. For example, Anthropic's standard API cache writes are 1.25x base input for five minutes or 2x for one hour, and reads 0.1x; verify current applicable pricing first. Output is not made free by caching an input prefix.
Keep durable instructions/tool definitions stable and put changing task state
later where the host permits. Prefix changes can reduce cache reuse, but the
exact matching, minimum lengths and invalidation rules are host-dependent.
Use actual usage fields (for example Anthropic cache_read_input_tokens) or
provider telemetry before claiming a cache hit or saving. Never schedule
requests just to keep an otherwise unused cache warm.
Load relevant capabilities
- Keep skill names/descriptions compact; load bodies only when relevant.
- Prefer a sufficient native tool or CLI already available. A dormant CLI adds no MCP schema; its instructions, commands and output still consume context.
- Scope optional MCP servers. Claude Code supports deferred Tool Search; Codex and Hermes support native tool filters. Respect user-owned servers.
- Advertised schema bytes/4 is only a rough size estimate, not loaded, cached or billed tokens. Measure the current provider rather than quoting old totals.
- Do not change models, auxiliary routing or providers merely to save tokens without the user's requested quality and configuration constraints.
- Do not route the main model through
openrouter/free. That router picks a free model at random, so model/cache reuse is not predictable. A $0 OpenRouter account is 50 requests/day and 20/minute. Cloudflare's free pool is 10,000 neurons/day and is not a token allowance. Confirmed daily quota exhaustion is not a retry fan-out. A 429 can also mean temporary rate limits or capacity: inspect the error/reset metadata and honorRetry-Afterwith bounded backoff. Do not silently switch models or paid routes. Checked 2026-10-07.
Free-route sources: OpenRouter limits, free router, Cloudflare pricing, Cloudflare error codes.
Batch and bound output
Batch independent reads, reuse evidence already in context, and request only
the fields/lines needed. A workspace map (WORKSPACE.md, see
workspace-organization) replaces re-exploration: read the map, then update
the row you changed. Check exit codes and truncation. Do not rerun a passing
gate unless its inputs or environment changed. Prefer targeted searches to
dumping entire trees, histories or configuration files.
RTK is lossy and command/version dependent, not a universal 97% saving. This pack's narrow safe hook covers standalone human-facing status/tests; exact diffs, search, reads, structured output and pipelines stay raw. Read the measured policy before changing rewrites. Preserve failures, exit status and access to full output.
Sources checked 2026-09-13: Anthropic cache rules, Claude Code caching.