claude-session-budget 1.2.0

Track Claude Code's 5-hour session usage from local JSONL and pause tool calls before hitting the limit. No API calls, no network.

claude-session-budget

Track Claude Code's 5-hour session usage locally — and automatically pause task queues before hitting the limit.

Discovered by reverse-engineering ~/.claude/projects/**/*.jsonl
No API calls. No web scraping. Pure local file parsing.

The Problem

Claude Code enforces a rolling 5-hour session limit. When running automated task queues or background agents, the session can hit its limit mid-task with no warning.

How It Works

%%{init: {'themeVariables': {'fontSize': '13px'}}}%%
flowchart TD
    A([You run a task in Claude Code]) --> B[Claude Code logs API response<br/>to local JSONL]
    B --> C[budget_check.py hook<br/>fires before next tool call]
    C --> D[scan_window:<br/>cutoff = now − 5h]
    D --> F[Sum TTL-aware weighted tokens<br/>over all in-window jsonl entries]
    F --> G{Usage % vs<br/>calibrated limit}
    G -->|&lt; 80%| H([✓ Proceed<br/>logs % to stderr])
    G -->|80–93%| I([⟳ Proceed + log sync notice])
    G -->|≥ 93%| J[⏸ Block dispatch until<br/>5-hour session resets]
    J -.->|wait for reset| A

Claude Code writes every API response to local JSONL files: ~/.claude/projects//.jsonl

Each assistant message contains token counts in a usage field. By summing these with pricing-ratio weights and calibrating against one /usage observation, we estimate session usage in real time.

Why no bridge_status anchor?

Earlier versions used type=system, subtype=bridge_status lines as a "5h session start" anchor. That turned out to be wrong: Claude Code emits a bridge_status whenever /remote-control attaches to a new CLI session, not when the underlying 5h Max window begins. Real users open claude many times within one 5h window, so anchoring on the most recent bridge_status silently leapt the cutoff forward and reset the budget count to ~0% mid-window. We now use a plain rolling-5h cutoff and ignore bridge_status entirely.

This means we cannot pinpoint the exact moment Anthropic's server-side 5h window started — the rolling-5h estimate is off by however much idle time preceded the earliest in-window jsonl message. In practice that's usually a few minutes; the fallback oldest = earliest in-window usage ts keeps the reset-time estimate within the same ballpark as /usage.

Token Weighting (cost-equivalent, input = 1.0)

Weights mirror Anthropic's published list-price ratios so the weighted total approximates dollar cost — the dimension the 5h Max cap actually tracks.

Token Type	Weight	Notes
input_tokens	1.00×	base
output_tokens	5.00×
cache_read_input_tokens	0.10×
cache_creation.ephemeral_5m_input_tokens	1.25×	default cache TTL
cache_creation.ephemeral_1h_input_tokens	2.00×	extended cache TTL
cache_creation_input_tokens (legacy)	1.25×	fallback when no TTL breakdown

The per-TTL breakdown landed in newer Claude Code builds; older jsonl entries that only carry the flat cache_creation_input_tokens field are weighted at 1.25× (the historical default and a safe lower bound). When both fields are present in the same entry, the breakdown supersedes the legacy field to avoid double-counting.

Calibration

The calibrated limit is auto-learned from real Anthropic API errors:

1. Every time budget_check.py runs, it inspects each in-window jsonl entry for the structural API-error signature: type=system, subtype=api_error with HTTP status=429, or any nested error.type containing rate_limit / usage_limit.
2. When it finds a new event, it takes the weighted token total at that moment as a real-world 100% reading.
3. The stored limit is EWMA-merged with the observation (default α=0.3) and written to ~/.claude/.budget_calibration.json.

Why structural matching, not text? An earlier version regex-matched "rate limit" / "limit reached" in the raw jsonl line. That picked up any user/assistant message body that mentioned the topic — including conversations debugging this very tool — and produced a self-poisoning EWMA loop that drove the calibrated limit from 63M down to 16M, causing false 100% BLOCKING. Structural signature matching eliminates that class of false positive.

You can also seed/refine the limit manually with one /usage reading:

# Mode A — explicit percentage
python3 scripts/calibrate.py --observed-pct 67

# Mode B — paste the full /usage panel; we extract the % for you
pbpaste | python3 scripts/calibrate.py --from-stdin

Known baselines (used until auto-learning kicks in):

• Claude Max (5x): ~63,226,913 weighted tokens = 100% (measured 2026-05-09)
• Claude Pro: unknown — contributions welcome

Background auto-calibration (zero user input)

Platform	Backend	Out-of-the-box
macOS / Linux	stdlib pty + subprocess.Popen	✅ works as-is
Windows	pywinpty (ConPTY wrapper)	⚙️ pip install pywinpty once

On Windows without pywinpty, auto_calibrate_supported() returns False and the hook silently skips dispatch — the base jsonl-scan estimation still works, you just lose the periodic server-truth refinement. Either install pywinpty (pip install "claude-session-budget[windows]") or use the manual --from-stdin path above.

The hook runs auto_calibrate.py in the background when pct first crosses each milestone in BUDGET_AUTO_CAL_MILESTONES (default 80,90,95). The worker:

1. Spawns claude under a pseudo-terminal so slash commands work.
2. Sends /usage once the input prompt renders.
3. Captures the panel, ANSI-strips, parses the Current session NN% row.
4. EWMA-merges into the calibration file and exits.

Designed to need zero user interaction — no copy-paste, no extra terminal. Cooldown of BUDGET_AUTO_CAL_COOLDOWN_SECS (default 300s) prevents bursts; each milestone fires AT MOST once per 5h window. Worst case: 3 spawns per window, each adding one claude startup's worth of cache-write tokens.

Disable entirely with BUDGET_AUTO_CAL_ENABLED=0. Inspect activity:

tail -20 ~/.claude/.budget_auto_calibrate.log

Manual one-shot run (useful for smoke-testing the spawn pipeline under limit):

python3 scripts/auto_calibrate.py

Installation

Option A — Claude Code Plugin Marketplace (Recommended)

This repo is itself a Claude Code marketplace. Inside Claude Code:

/plugin marketplace add Star001-KR/claude-session-budget
/plugin install session-budget

The PreToolUse hook is wired automatically via hooks/hooks.json, the skill becomes available as /session-budget:budget-check, and all scripts run from ${CLAUDE_PLUGIN_ROOT}/scripts/.

Option B — Homebrew

brew tap Star001-KR/claude-session-budget https://github.com/Star001-KR/claude-session-budget
brew install Star001-KR/claude-session-budget/claude-session-budget

This installs the budget-check and budget-calibrate commands into $(brew --prefix)/bin. To wire budget-check as a Claude Code PreToolUse hook, add to ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "*",
        "hooks": [{"type": "command", "command": "/opt/homebrew/bin/budget-check"}]
      }
    ]
  }
}

Option C — PyPI

pip install claude-session-budget

This installs the budget-check and budget-calibrate console scripts and makes the package importable as claude_session_budget. Wire budget-check as a Claude Code PreToolUse hook the same way as the brew option (the binary will be on your $PATH):

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "*",
        "hooks": [{"type": "command", "command": "budget-check"}]
      }
    ]
  }
}

For PM-layer / orchestrator integration:

from claude_session_budget.session_budget_manager import SessionBudgetManager

budget = SessionBudgetManager()
status = budget.get_status()

Option D — Manual Hook (no plugin, no brew, no pip)

curl -fsSL https://raw.githubusercontent.com/Star001-KR/claude-session-budget/main/install.sh | bash

The installer pins to the tagged release and verifies downloaded scripts against SHA256SUMS from that release before writing to ~/.claude/hooks.

Or manually add to ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "*",
        "hooks": [{"type": "command", "command": "python3 ~/.claude/hooks/budget_check.py"}]
      }
    ]
  }
}

Option E — Claude Code Skill (manual copy)

mkdir -p .claude/skills/session-budget
cp skills/budget-check/SKILL.md .claude/skills/session-budget/SKILL.md
cp scripts/budget_check.py .claude/skills/session-budget/check.py

Option F — PM Layer / Orchestrator (manual)

import sys; sys.path.insert(0, "scripts")  # or install as a package
from session_budget_manager import SessionBudgetManager

budget = SessionBudgetManager()

async def dispatch_task(task):
    wait_secs = await budget.check_before_dispatch()
    if wait_secs:
        await asyncio.sleep(wait_secs)
    # Optional: log current state for dashboards / observability
    s = budget.get_status()
    log.info(f"{s['pct']}% — resets in {s['remaining_str']} (epoch={s['reset_at']})")

get_status() returns a dict with both raw numbers and a human-friendly remaining string:

{
  "pct": 13.2,
  "weighted_tokens": 2_128_235,
  "calibrated_limit": 63_226_913,
  "reset_at": 1778355198.018,        # epoch seconds (anchor + 5h, or oldest msg + 5h)
  "remaining_secs": 16_755,
  "remaining_str": "4h 39m",         # or "already reset" when remaining == 0
}

Thresholds

Threshold	Default	Behavior
Sync	80%	Re-reads JSONL and logs updated estimate
Pause	93%	Blocks by default; optional hook sleep mode can wait and re-check

Set thresholds via env vars or a .env file (loaded automatically):

BUDGET_SYNC_PCT=80 BUDGET_PAUSE_PCT=93 python3 scripts/budget_check.py

.env lookup order — process env always overrides:

1. ~/.claude/.env (global default, always loaded)
2. ./.env (current working directory) — opt-in: set BUDGET_LOAD_PROJECT_ENV=1 to enable the per-project override
3. Built-in defaults

Migration from <1.1.4: ./.env is no longer auto-loaded by default. To restore per-project override behavior, add BUDGET_LOAD_PROJECT_ENV=1 to ~/.claude/.env (or your shell env).

Copy .env.example to get started:

cp .env.example ~/.claude/.env

Pause Modes

By default, budget_check.py blocks immediately when usage reaches the pause threshold. Leave BUDGET_PAUSE_MODE unset, empty, or set to block:

BUDGET_PAUSE_MODE=block

You can opt into sleep mode:

BUDGET_PAUSE_MODE=sleep
BUDGET_RECHECK_SECS=60
BUDGET_RESET_GRACE_SECS=60
BUDGET_MAX_SLEEP_SECS=14400

In sleep mode, the PreToolUse hook process stays alive, periodically re-checks local JSONL usage, and exits 0 once usage falls below the pause threshold. This lets the original tool call continue after the 5-hour window has rolled forward enough. Before resuming, it sleeps BUDGET_RESET_GRACE_SECS and checks one more time.

Important Risks

Sleep mode is experimental and disabled by default.

• The hook process may remain alive for minutes or hours.
• Claude Code, your shell, terminal, OS, or task runner may impose timeouts.
• The UI can appear stuck while the hook is sleeping.
• If the reset estimate is wrong, the hook may still block after waiting.
• Sleep mode is best for supervised local use, not unattended automation.

For reliable queue pause/resume behavior, prefer SessionBudgetManager in an orchestrator or PM layer.

Environment Variables

All variables can be set in process env or ~/.claude/.env. Process env always overrides. ./.env (cwd) is opt-in via BUDGET_LOAD_PROJECT_ENV.

Variable	Default	Description
BUDGET_SYNC_PCT	80	Sync threshold (% of limit). At/above this, hook logs an estimate update
BUDGET_PAUSE_PCT	93	Pause threshold (% of limit). At/above this, hook blocks (or sleeps)
BUDGET_PAUSE_MODE	block	block → exit 2 immediately. sleep → keep hook alive, re-check periodically
BUDGET_RECHECK_SECS	60	sleep mode: jsonl re-scan interval
BUDGET_RESET_GRACE_SECS	60	sleep mode: extra wait after threshold drop, before resume
BUDGET_MAX_SLEEP_SECS	14400	sleep mode cap (4h). After this, hook gives up and exits 2
BUDGET_EWMA_ALPHA	0.3	EWMA smoothing factor for auto-learned limit
BUDGET_CALIBRATED_LIMIT	(unset)	Hard override of stored calibrated limit (weighted tokens)
BUDGET_PROJECTS_DIR	~/.claude/projects	jsonl scan root
BUDGET_CALIBRATION_FILE	~/.claude/.budget_calibration.json	Persistence path for auto-calibration
BUDGET_LOAD_PROJECT_ENV	(unset)	Set to 1 to also load ./.env from cwd at module import. Disabled by default to avoid an untrusted-cwd attack surface and import-time side effects

Limitations

• Token weights are a proxy — Anthropic's internal formula is not public
• Peak hours (weekday 5–11am PT) consume limits faster
• Cross-device usage is not tracked (JSONL files are local only)
• The bridge_status anchor is intermittent: it appears when /remote-control activates, not on every tool call. When stale (long idle gaps) the tool falls back to the plain 5h rolling window
• The rate-limit api_error signature is conservative — accepts both status=429 and any inner error.type containing rate_limit/usage_limit. We haven't directly observed a real 429 jsonl line yet, so the exact inner type string can be tightened once one shows up
• Recalibrate after plan changes

Files

Path	Description
.claude-plugin/plugin.json	Plugin manifest (name, version, author, license)
.claude-plugin/marketplace.json	Marketplace manifest — lets /plugin marketplace add resolve this repo
hooks/hooks.json	PreToolUse hook declaration using ${CLAUDE_PLUGIN_ROOT}
skills/budget-check/SKILL.md	Claude Code skill definition (auto-discovered as /session-budget:budget-check)
scripts/budget_check.py	Lightweight hook script (no deps); also runs auto-calibration
scripts/session_budget_manager.py	Full async class for PM/orchestrator integration
scripts/calibrate.py	Manual calibration entry from a /usage reading
scripts/_budget_core.py	Shared core: .env loader, JSONL scan, anchor detection, signature matcher, EWMA learner
tests/test_budget_core.py	Unit tests (59) — env loading, jsonl scan, anchor, signature matcher, EWMA
.env.example	Copy to ./.env or ~/.claude/.env
install.sh	One-line installer for the manual (non-plugin) hook setup
Formula/claude-session-budget.rb	Homebrew formula (used when this repo is added as a brew tap)
pyproject.toml	PyPI packaging metadata (PEP 517/621) — pip install claude-session-budget
scripts/__init__.py	Marks scripts/ as the claude_session_budget package via package-dir mapping
docs/internals.md	Architecture deep-dive (anchor + 5h fallback + signature matcher + EWMA)
LICENSE	MIT

Contributing

PRs welcome — especially calibration values for Pro and Max 20x plans.

License

MIT