Savings

PRECC tracks estimated token savings from every interception. Use precc savings to see how much waste PRECC has prevented.

Quick Summary

$ precc savings
Session Token Savings
=====================
Total estimated savings: <span data-stat="session_tokens_saved">8,741</span> tokens

Breakdown:
  Pillar 1 (cd prepends):         <span data-stat="session_p1_tokens">3,204</span> tokens  (<span data-stat="session_p1_count">6</span> corrections)
  Pillar 4 (skill activations):   <span data-stat="session_p4_tokens">1,560</span> tokens  (<span data-stat="session_p4_count">4</span> activations)
  RTK rewrites:                   <span data-stat="session_rtk_tokens">2,749</span> tokens  (<span data-stat="session_rtk_count">11</span> rewrites)
  Lean-ctx wraps:                 <span data-stat="session_lean_tokens">1,228</span> tokens  (<span data-stat="session_lean_count">2</span> wraps)

Detailed Breakdown (Pro)

$ precc savings --all
Session Token Savings (Detailed)
================================
Total estimated savings: <span data-stat="session_tokens_saved">8,741</span> tokens

Command-by-command:
  #  Time   Command                          Saving   Source
  1  09:12  cargo build                      534 tk   cd prepend (cargo-wrong-dir)
  2  09:14  cargo test                       534 tk   cd prepend (cargo-wrong-dir)
  3  09:15  git status                       412 tk   cd prepend (git-wrong-dir)
  4  09:18  npm install                      824 tk   cd prepend (npm-wrong-dir)
  5  09:22  find . -name "*.rs"              387 tk   RTK rewrite (output truncation)
  6  09:25  cat src/main.rs                  249 tk   RTK rewrite (lean-ctx wrap)
  7  09:31  cargo clippy                     534 tk   cd prepend (cargo-wrong-dir)
  ...

Pillar Breakdown:
  Pillar 1 (context resolution):   <span data-stat="session_p1_tokens">3,204</span> tokens  <span data-stat="session_p1_pct">36.6</span>%
  Pillar 2 (GDB debugging):            0 tokens   0.0%
  Pillar 3 (mined preventions):        0 tokens   0.0%
  Pillar 4 (automation skills):    <span data-stat="session_p4_tokens">1,560</span> tokens  <span data-stat="session_p4_pct">17.8</span>%
  RTK rewrites:                    <span data-stat="session_rtk_tokens">2,749</span> tokens  <span data-stat="session_rtk_pct">31.5</span>%
  Lean-ctx wraps:                  <span data-stat="session_lean_tokens">1,228</span> tokens  <span data-stat="session_lean_pct">14.1</span>%

How Savings Are Estimated

Each correction type has an estimated token cost based on what would have happened without PRECC:

These are conservative estimates. Actual savings are often higher because Claude’s reasoning about errors can be verbose.

Cumulative Savings

Savings persist across sessions in the PRECC database. Over time, you can track the total impact:

$ precc savings
Session Token Savings
=====================
Total estimated savings: <span data-stat="session_tokens_saved">8,741</span> tokens

Lifetime savings: <span data-stat="total_tokens_saved">142,389</span> tokens across <span data-stat="total_sessions">47</span> sessions

Status Bar

After installation, PRECC wires a statusLine entry into ~/.claude/settings.json so the Claude Code status bar shows live session metrics:

$0.42 spent | 1.2M in/out | 📊 last cmd: −1.2K | PRECC: 7 fixes | 5.8ms avg | this session: 320 saved over 7 cmds (~$0.05) | lifetime: 8.9K saved over 217 cmds (~$2.85)

Set PRECC_LANG to render the labels in your language — see the Localization chapter.

Each segment:

The lifetime: segment is placed last so it’s the first to be truncated if Claude Code’s UI clips the bar at the right edge.

Why cost and token count don’t divide

The displayed 1.2M in/out is not the denominator that produced $0.42 spent. Claude Code’s cost.total_cost_usd is computed from the API’s full token breakdown — base input, output, plus cache reads and cache creations. The session-wide cumulative cache token counts are not exposed in the statusline schema, so PRECC can only show the visible (non-cache) portion.

On long sessions with heavy file rereads, cache reads can be 10× the visible token count. That’s why pairing the two as a ratio would mislead — PRECC shows them as independent segments instead.

Why PRECC doesn’t compute the cost

The cost number is authoritative. PRECC reads cost.total_cost_usd verbatim from the JSON Claude Code pipes into the status command on stdin. That’s the same number Claude Code charges against your subscription/usage budget. You can verify it any time with the built-in /cost slash command — both should agree.

What drives the cost

For Claude Opus 4.6:

The biggest drivers on long sessions are usually output tokens (the most expensive per-token type, especially on the 1M context tier), repeated cache reads (cheap individually but accumulating fast across many turns), and cache creations (written once per file read at ~1.25× the base input rate). PRECC reduces the visible-token cost by compressing Bash output (the 📊 last cmd: segment shows the per-command saving), but it cannot reduce cache reads of files Claude has already loaded.

Stable session counts

The “PRECC: N fixes” segment counts events since the persisted session start, written to ~/.local/share/precc/sessions/<session_id>.start on the first statusline refresh of each session. This makes the count monotonic — it cannot drop mid-session even if cost.total_duration_ms is missing on a particular refresh.

Auto-refreshed lifetime snapshot

The lifetime: segment reads ~/.local/share/precc/.lifetime_summary.json, which is rewritten on every PostToolUse measurement and on every precc savings invocation. The this session: segment reads the same lifetime file but subtracts a per-session baseline persisted on the first refresh of each session. No manual refresh needed — the files update themselves.

Suppressing the status bar

If you’d rather keep your existing status bar, set your own statusLine command in ~/.claude/settings.json. PRECC’s installer will detect the custom value and leave it alone on subsequent updates.

To suppress only the per-interaction 📊 PRECC line (in additionalContext), set PRECC_QUIET=1 in your shell environment.

Related research

PRECC’s three savings mechanisms each have a counterpart in the recent literature. These are related work — the ideas PRECC’s design draws on. Their reported figures are their measurements, not PRECC’s: PRECC only ever quotes numbers measured on your own machine (see “measured, not estimated”, above).

• Output/trajectory trimming (PRECC’s diet + bash-output compression) — Reducing Cost of LLM Agents with Trajectory Reduction (AgentDiet), FSE 2026, arXiv:2509.23586. Removes redundant/expired trajectory content post-hoc; reports −39.9–59.7% input tokens. PRECC applies the same idea pre-execution and deterministically (no extra LLM call).
• Skills as programs (PRECC’s mined + builtin rewrite skills) — Harnessing LLM Agents with Skill Programs, arXiv:2605.17734. Frames reusable agent skills as executable program functions — the same analogy behind PRECC’s command-rewrite skills (a pattern → a deterministic rewrite).
• Context compression (PRECC’s compress + lean-ctx wrapping) — Compress the Context, Keep the Commitments: A Formal Framework for Verifiable LLM Context Compression, arXiv:2605.17304. Recent work on compressing context without losing required information — the property PRECC’s deterministic, cache-stable rewrites aim to preserve.