← Back to News

update

Nov 18, 2026

Token savings now separate "saved" from "potential"

The tokens dashboard and weekly digest now count a saving only if the model actually received fewer tokens. Observe-mode measurements show separately. We also added three new ways to cut tokens.

The tokens page in the enterprise dashboard now shows two numbers instead of one. Saved counts tokens the model never received. Potential counts what a mechanism measured in observe mode but did not apply. The weekly tokens digest uses the saved figure, and you can now send that digest on demand for the current week or any of the previous 12 weeks.

If your org runs some compression mechanisms in observe mode, your reported savings will go down. That is the point of the change: the old single number mixed both kinds.

What changed on the tokens page

  • A totals card shows Saved — reached the model next to Potential — measured in observe mode, not applied.
  • Each mechanism row reads like 9k saved (3 enforced) · 3k potential (2 observe-only).
  • Project badges split saved from potential.
  • The Saved/Session and Saved/Line ratios use saved tokens only.

Every savings event recorded by the hooks now carries an applied flag. Rollups keep the applied total (applied_tokens_saved) next to the measured total (tokens_saved_estimate). For older data recorded before the flag existed, we count a mechanism's savings as applied only if it never ran in observe mode. Otherwise historical applied savings show as 0. This undercounts some history, which we preferred to overcounting it.

Token figures in the dashboard are still estimates (characters ÷ 4), not provider-tokenizer counts.

A fix that made the split necessary

While checking these numbers against real Claude Code sessions, we found that Claude Code silently ignores a replacement tool result passed as a plain string. It has to match the shape of the original result: file content for Read, stdout/stderr for shell commands. Two things went wrong because of this:

  1. Read-cache replacements were logged as applied but never reached the model, so savings were over-reported.
  2. Redacted output from the DLP hook was also dropped for Read and cat, so the unredacted text reached the model in those cases.

Both are fixed in the current rclm release. Replacements now keep the original result shape, and we verified with the real Claude CLI that redacted and cached output is what the model receives. Update with pip install -U rclm.

Three new ways to cut tokens

All three shape a tool result once, as it enters the context. Earlier history is never rewritten, so the provider's prompt cache stays valid.

Outline on cross-session re-read. Session A reads a 300+ line file. Later, session B reads the same file from the top, and the file is unchanged on disk (we check its hash). Session B gets a symbol outline with real line numbers instead of the full content, and the model can then read just the range it needs. In one test with the real Claude CLI, a 677-line file cost about 13k tokens on first read and 689 tokens (outline plus one targeted read) the second time, and the model still answered correctly. If the model re-reads the file in full anyway, we record a negative saving equal to that read, so the dashboard doesn't claim a saving that didn't happen. This one is off by default (observe mode). Org admins can switch it to enforce.

Repeated log-line collapse. Retry loops and progress logs often print the same line hundreds of times with only the numbers changing. We keep the first line and add a note like [rclm: +149 similar lines differing only in numbers]. It needs at least 5 repeats and only applies to lines that look like logs. Tables, CSV, JSON and key=value output are never folded.

Blob placeholders. Long base64 runs (160+ characters) and long hex runs (128+) are replaced with a placeholder that includes the length and a short sha256. Shorter hex values such as commit hashes and sha256 digests are left alone. In one test, a command's output went from about 4.6k tokens to 132.

Both minifiers apply only to successful shell output that no command-specific filter already handles. They never touch file contents (Read, cat, head, sed -n, …), edits, or output from failed commands.

We also fixed two correctness gaps in the read cache: its state now resets when Claude Code compacts the conversation, and each subagent keeps its own state. Before this, a "you already read this" note could point at content the model no longer had.

Weekly digest: fixed and on demand

The scheduled weekly tokens digest wasn't being delivered because of a database constraint. That is fixed, and the next digest goes out Monday at 08:00 recipient-local time.

The digest settings dialog now has a Send Digest button with a week picker: This week (so far), Last week, or up to 12 weeks back. It sends the real digest for that week, built from your sessions with the sections you've saved, to your own address. Weeks with no sessions are rejected rather than sent empty. Manual sends don't affect the scheduled Monday email.

A manual send uses Monday 00:00 boundaries, while the scheduled digest covers Monday 08:00 to Monday 08:00, so the two can differ by a few hours of sessions.

Limitations we're watching

  • Small test sample. The numbers above come from a handful of controlled runs, not a fleet-wide measurement.
  • Outline costs. An outline can cost an extra turn when the model needs most of the file anyway. The negative event records the token cost, but not the extra round trip.
  • Collapsed lines. Log-line collapse can still fold lines where the numbers mattered. If a collapsed result hid something your agent needed, please tell us.
  • Client support. Outline replacement is enforced only for Claude Code for now. Codex, Cursor, Gemini CLI and Copilot record it in observe mode.
  • Other pages. The personal /stats page still shows the combined figure.

If the saved or potential numbers on your dashboard look wrong, or a replacement cost your agent a correct answer, send us the session. Those reports are how we tune these thresholds.