RCP-CWK-010 Cowork Token Monitor

Monitors token usage throughout your session using weighted tracking and velocity detection

Logs estimates to a session-scoped file and integrates with the COM Extension (CWK-012) for automatic threshold-based triggers. Extends the Core Token Monitor with persistent logging and Cowork-specific capacity planning.

Cowork Token Monitor with File-Based Logging

Cowork | Bridge (Core Extension) | Difficulty: Medium | Time: Under 1 min (auto) / 1–2 min (on-demand)
Tags: Introduced in Beta, Token, Monitor, Velocity, Capacity, Logging, COM, Threshold, Bridge, Cowork

TL;DR

What It Does
Monitors token usage throughout your session using weighted tracking and velocity detection. Logs estimates to a session-scoped file and integrates with the COM Extension (CWK-012) for automatic threshold-based triggers. Extends the Core Token Monitor with persistent logging and Cowork-specific capacity planning.
How It Works
The AI calculates a weighted token estimate by classifying content types — simple exchanges count at 1x, code blocks at 2x, recipe implementations at 4x, and so on. It tracks consumption velocity (tokens per exchange, acceleration trend, tool-call density) and predicts how many exchanges remain until warning (70%), critical (85%), and auto-handoff (90%) thresholds. At 70%, it activates Late-Session Mode, which triggers a confidence decay penalty in downstream verification recipes. All estimates are logged to a session-scoped file for intra-session diagnostics.
What To Expect
In automatic mode, the monitor runs silently until a threshold is reached — then you see a warning, critical alert, or auto-handoff trigger. On-demand mode gives you a full status report any time. Before-task mode tells you whether an upcoming task fits within your remaining token budget.
Best Results When You
Let automatic mode run in the background. Use on-demand checks before large tasks. Pay attention to the 70% warning — it means Late-Session Mode is active and downstream confidence scoring is now stricter.
Typical Time
Under 1 minute for automatic checks. 1–2 minutes for on-demand or before-task reports including velocity analysis.
Difficulty
Medium — the recipe runs automatically via COM integration. You interact with it when you request on-demand checks or need before-task capacity planning. Understanding the threshold system and Late-Session Mode helps you plan your session effectively.

How To Start

The token monitor operates in three modes. Automatic mode runs via COM integration without you doing anything. On-demand and before-task modes are triggered when you ask for a status check.

STEP 1Choose Your Check Mode

Tell the AI which type of token check you need, or let COM handle it automatically.
Available parameters
  • check_type · string · required
    Type of token check. Options: “automatic,” “on_demand,” “before_task.” Automatic runs silently via COM. On-demand gives a full report. Before-task checks capacity for a specific upcoming task.
  • upcoming_task · string · optional
    Task name for capacity planning. Used with “before_task” mode.
  • force_report · boolean · optional · default: false
    Force a full report regardless of threshold status.
  • log_to_file · boolean · optional · default: true
    Write the token estimate to the session log file.
Invocations
Automatic (COM-Triggered)
#H->AI::Directive: (Monitor token usage — Cowork mode)
On-Demand Status Check
#H->AI::Directive: (Monitor token usage — Cowork mode with check_type=on_demand)
Before-Task Capacity Check
#H->AI::Directive: (Monitor token usage — Cowork mode with check_type=before_task, upcoming_task=develop recipe)

Usage Examples

Three scenarios — automatic threshold detection, an on-demand status check, and a before-task capacity warning.
Example 1 · Automatic mode triggers 70% warning
AI: [running silently in background via COM]
AI: Token usage: approximately 70% of session limit. Late-Session Mode is now active — claims routed through CWK-024 receive an automatic -10 confidence penalty unless grounded by a fresh tool observation. Checkpoint recommended.
Example 2 · On-demand status check
User: How are we on tokens?
AI: Token usage: approximately 55% of limit. Velocity: roughly 800 tokens per exchange, steady. Approximately 12 exchanges until 70% warning, 24 until 85% critical. Tool-call density: medium.
Example 3 · Before-task capacity warning
User: Do we have room for a comprehensive analysis?
AI: Current usage: approximately 62%. A comprehensive analysis typically costs around 10% of the session budget. That would bring us to roughly 72%, crossing the 70% warning threshold. Consider running a checkpoint first.
User: Good call, let’s checkpoint first.

How AI Reads This Recipe

When this recipe is triggered, the AI runs a 7-step token monitoring cycle with weighted estimation, velocity tracking, threshold detection, and file-based logging. The AI should:
  1. CALCULATE weighted token usage by classifying content types: simple exchanges (1x), code blocks (2x), structured data (2.5x), test results and tool outputs (3x), long analyses (3x), recipe implementations (4x), file reads (2x), bash executions (1.5x).
  2. CALCULATE velocity: tokens in the last 5 exchanges, average per exchange, acceleration trend (increasing, steady, or decreasing), and tool-call density. Predict exchanges remaining until 70% and 85% thresholds.
  3. IN AUTOMATIC MODE: trigger warnings at 70% (activate Late-Session Mode, fire COM checkpoint recommendation), alerts at 85% (recommend handoff), and auto-handoff at 90% (execute CWK-002). The LATE_SESSION_MODE flag persists for the remainder of the session and causes CWK-024 to apply a -10 confidence penalty on ungrounded claims.
  4. IN ON-DEMAND MODE: report current usage percentage, velocity, exchanges to warning and critical thresholds, and tool-call density.
  5. IN BEFORE-TASK MODE: estimate the cost of the upcoming task using Cowork-specific estimates (e.g., “develop recipe” ~5%, “comprehensive analysis” ~10%, “sub-agent delegation” ~12%). Warn if the task would cross a threshold.
  6. LOG the estimate to a session-scoped file with timestamp, check type, usage percentage, velocity, acceleration, tool-call density, exchanges to thresholds, upcoming task, and action taken. The log does not persist across sessions.
  7. INTEGRATE with CWK-012 COM Extension: set flags that CWK-012’s auto-trigger rules act on. Do not call CWK-003 directly — COM handles checkpoint triggering.
The Late-Session Mode flag is critical — once set at 70%, it does not reset for the remainder of the session. CWK-024 checks this flag in its scoring logic. If unable to estimate token usage, warn at 60%+ as a safety margin.

When to Use This Recipe

Use this recipe when you:
  • Want to know how much session capacity remains before starting a large task.
  • Need automatic warnings as you approach token limits so you can checkpoint and hand off gracefully.
  • Want velocity-based predictions of how many exchanges you have left at your current pace.
  • Need the Late-Session Mode flag set so downstream verification recipes apply appropriate confidence penalties.
Do not use this recipe when:
You are in the first 10 exchanges of a fresh session and nowhere near any threshold. Running manual token checks this early wastes time — let automatic mode handle it. On-demand and before-task checks become valuable after you have accumulated enough context for meaningful velocity calculation (roughly 10+ exchanges).

Recipe FAQ

Q.What are the weighted content multipliers?

Simple exchanges: 1x. Code blocks: 2x. Structured data: 2.5x. Test results and tool outputs: 3x. Long analyses and reports: 3x. File reads (Read tool): 2x. Bash executions: 1.5x. Recipe implementations: 4x.

Q.What is Late-Session Mode?

A flag activated when token usage reaches 70%. Once active, it persists for the rest of the session. CWK-024 (Verification Gate) uses this flag to apply an automatic -10 confidence penalty on claims that are not grounded by a fresh tool observation from the current step. This guards against context compaction degrading earlier observations.

Q.Does the log file persist across sessions?

No. The session-scoped log file resets when the VM resets between sessions. The handoff captures final token state. The log is for intra-session diagnostics and COM pattern detection only.

Q.What Cowork-specific task estimates are built in?

Develop recipe: approximately 5%. Comprehensive analysis: 10%. Multiple test cases: 15%. Full documentation: 10%. Implement recipe: 8%. Sub-agent delegation (CWK-008): 12%. Roadmap update: 6%. CLAUDE.md generation (CWK-007): 4%. Git checkpoint plus push: 2%. Handoff creation: 5%.

Q.How does this recipe interact with CWK-012 COM Extension?

This recipe sets threshold flags that CWK-012’s auto-trigger rules act on. When automatic mode detects 70%, COM fires CWK-003 (mid-session checkpoint), logs the trigger, and updates handoff metadata. This recipe does not call CWK-003 directly.

Q.What happens at the auto-handoff threshold?

At 90%, the recipe triggers an error-level alert and directs immediate execution of CWK-002 (session handoff). This is the hard ceiling — the session should be wrapped up at this point.

Version History

Changes to this recipe over time. Most recent first.
v1.01a 2026-03-15
Added LATE_SESSION_MODE flag: activates at 70% threshold, persists for remainder of session, triggers -10 confidence decay in CWK-024 for ungrounded claims. Added tool-call density tracking. Added audit trail logging per R-12.
v1.00a 2026-02-28
Initial release. Extends Core Token Monitor with Cowork-specific features: 7-step monitoring cycle, weighted content classification, velocity tracking with acceleration, three check modes (automatic, on-demand, before-task), session-scoped file logging, CWK-012 COM integration for threshold-aware auto-triggers. Cowork-specific task cost estimates for capacity planning.

Get this recipe with CRAFT for Claude Cowork

Cowork recipes ship bundled with CRAFT for Claude Cowork — there’s no separate download. Clone the framework once, and your AI runs every recipe automatically when invoked.

Pull anytime to stay on the latest version — free to clone, no login or email required.

Then start your session

Once CRAFT is in your project folder, open a new Cowork session and ask Claude to initialize. For example:

You

Please initialize my CRAFT session.

Claude

CRAFT session ready. Your project is loaded, your persona is active, and your recipes are available. What would you like to work on?

What is CRAFT for Claude Cowork?

Not familiar with Git? Download as a ZIP

No command line needed. Just download, move, and unzip:

  1. Open the CRAFT framework repo on GitHub.
  2. Click the green Code button, then choose Download ZIP.
  3. Move the downloaded ZIP into your Claude Cowork project folder.
  4. Unzip it: double-click on Mac, or right-click → Extract All on Windows.

Similar Posts