⛽ context-gauge

A context-saturation gauge for coding-agent harnesses — Claude Code and Grok Build. One small, dependency-free script that shows you — and tells the model — how full the session's reasoning working set is, so you split, scope down, or hand off before quality quietly degrades.

Formerly claude-context-gauge. The binary is now context-gauge; claude-context-gauge remains a shim.

🟢 GREEN   <40K   full reasoning capacity
🟡 YELLOW  40–90K  prefer delegating searches; scope new inputs deliberately
🟠 ORANGE  90–150K split before reading; no new large reasoning input; checkpoint soon
🔴 RED     150–250K handoff imminent — finish what's in hand; take on no new load
⚫ BLACK    ≥250K   STOP — hand off now; the next big step risks lossy auto-compaction

Bands are absolute working-set tokens, not a % of the model window. The same table applies to Claude (~200K–1M) and Grok (grok-4.5 ≈ 500K, composer-fast ≈ 200K). A bigger window adds overflow room, not a longer effective span.

---

Harness support

Harness	Integration	Fill source	Floor
Claude Code	statusLine + UserPromptSubmit	transcript usage (input + cache_read + cache_creation)	first main assistant turn
Grok Build	UserPromptSubmit (+ SessionStart / PostCompact seed)	signals.json → contextTokensUsed	first observation (re-seed after compact)

Auto-dispatch: stdin with transcript_path → Claude; sessionId / GROK_SESSION_ID → Grok.

---

Why a working set, not raw fill

Every session is born tens of thousands of tokens deep — harness system prompt, tool schemas, always-injected files. That floor is cached, position-privileged, and not what the model actively reasons over. Banding raw fill would flag every fresh session as "degraded" on turn one (cry-wolf).

working = current_fill − session_floor

This self-calibrates per project so the band tracks the thing that actually degrades.

Why absolute tokens, not % of the window

Long-context degradation onset is roughly window-independent:

• NoLiMa (2025): models advertised at 128K–1M drop below 50% of short-context baseline by ~32K task tokens.
• Chroma, "Context Rot" (2025): a 200K-context model measurably degrades around ~50K.
• Lost in the Middle (Liu et al., 2023): accuracy sags mid-context — a U-curve, not a cliff at the window edge.

The window size only feeds a secondary auto-compaction-proximity % (cosmetic).

Thresholds are a considered default, not a law of nature. Tune the constants at the top of the script.

---

Install

Requires Python 3 (stdlib only — no third-party deps).

One-liner (uvx / pip)

# run without installing
uvx --from git+https://github.com/jkubo/context-gauge context-gauge --self-test

# install as a user tool
uv tool install git+https://github.com/jkubo/context-gauge
# after PyPI publish:
# uvx context-gauge --self-test
# uv tool install context-gauge
# pipx install context-gauge

From a clone

git clone https://github.com/jkubo/context-gauge
cd context-gauge
./install.sh          # → ~/.local/bin/context-gauge (+ claude-context-gauge shim)
# or: pip install -e .   /   uv tool install .

Claude Code

Merge into ~/.claude/settings.json (see settings.example.json):

{
  "statusLine": {
    "type": "command",
    "command": "$HOME/.local/bin/context-gauge",
    "padding": 0
  },
  "hooks": {
    "UserPromptSubmit": [
      { "hooks": [ { "type": "command", "command": "$HOME/.local/bin/context-gauge" } ] }
    ]
  }
}

statusLine needs a recent Claude Code (present as of v2.1.212). Use either or both integrations.

Grok Build

Copy hooks.grok.example.json to ~/.grok/hooks/context-gauge.json, then /hooks → r (reload). Floor cache: ~/.context-gauge/floors/<session_id>.json.

{
  "hooks": {
    "UserPromptSubmit": [
      { "hooks": [ { "type": "command", "command": "$HOME/.local/bin/context-gauge", "timeout": 5 } ] }
    ],
    "SessionStart": [
      { "hooks": [ { "type": "command", "command": "$HOME/.local/bin/context-gauge", "timeout": 5 } ] }
    ],
    "PostCompact": [
      { "hooks": [ { "type": "command", "command": "$HOME/.local/bin/context-gauge", "timeout": 5 } ] }
    ]
  }
}

Grok has no statusLine yet — inject-only.

---

Configuration (optional env vars)

Variable	Default	Effect
CONTEXT_GAUGE_DISABLE	(unset)	Any truthy value → no-op (kill switch).
CLAUDE_CONTEXT_GAUGE_DISABLE	(unset)	Legacy alias for the same.
CONTEXT_GAUGE_WINDOW	200000	Claude window for cosmetic % only. Set 1000000 for 1M beta.
CLAUDE_CONTEXT_GAUGE_WINDOW	—	Legacy alias.
CONTEXT_GAUGE_FLOOR_DIR	~/.context-gauge/floors	Grok floor cache directory.
GROK_HOME	~/.grok	Grok sessions root.

---

How it reads the numbers

Claude. Streams the session transcript JSONL for main-window assistant usage lines (skips sidechain/meta). Aborted turns that write all-zero usage are ignored (would otherwise false-GREEN a saturated session). Floor = fill of the first genuine assistant turn.

Grok. Resolves ~/.grok/sessions/<urlencode(cwd)>/<sessionId>/signals.json and reads contextTokensUsed / contextWindowTokens / compactionCount. Floor is written on first observation and re-seeded when compaction runs or fill drops below the stored floor. Mid-session first attach seeds floor = current total (tracks growth from attach, under-alerts once rather than cry-wolf).

---

CLI

context-gauge --self-test                 # print all five bands
context-gauge --transcript FILE.jsonl     # Claude: hook + status lines
context-gauge --session-dir DIR           # Grok: one reading from signals.json
context-gauge --signals FILE.json         # Grok: direct signals file

---

Tests

python3 -m pytest tests/ -v
# or
python3 -m unittest tests.test_gauge -v

Packaging / release

uv build                    # → dist/*.whl + dist/*.tar.gz
uv publish                  # needs UV_PUBLISH_TOKEN / PyPI trusted publisher

Entry points: context-gauge and claude-context-gauge (shim) → context_gauge:main.

License

Apache-2.0. See LICENSE.

---

Rename note

This project was published as claude-context-gauge (2026-07-23). It grew a Grok Build backend and the neutral name context-gauge. The old binary name remains a shim; the GitHub repo redirects after rename.