Glassbox

    1

    Other tools show what happened in a Claude Code session. Glassbox tells you what went wrong: retry loops, context bloat, cost, and claims with no receipt.

    Free

    Security scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Glassbox

    Example session with this skill installed

    Check this session with Glassbox and act on what it finds. Use --redact so I can paste the report into the team channel.

    • Read your context and instructions
    • Compiled the glassbox

    Ran npx -y glassbox-trace@0.11.1 check 5053865e --format md --redact --fail-on info (exit 1: findings at or above the threshold). Report below, then what I'll do differently.


    Glassbox report — «7 chars»

    • session: 5f9130fb-ffc2-5d3e-a9df-ad92a3f060f3 · model: claude-sonnet-4-6, claude-fable-5-1 · started: 2026-10-25T23:25:47.637Z
    • wall 58m 24s (active 58m 24s) · 0 turns · 109 requests · 125 tool calls (11 failed, 1 unanswered)
    • tokens: 29,156,332 context served (99% cached) · peak prompt 312,708 · 40,245 output (7,058 thinking) · est. cost $15.43

    Findings (8)

    1. ERROR context-bloat — Context reached 312,708 tokens · $15.43 (100% of cost) spent above 120,000

    First crossed 120,000 at request #6 (turn 5); 104 requests ran above it. Every one of them re-reads that prefix.

    Evidence

    • request #6 · turn 5: context 238,389 · output 478 · 8.1 s
    • request #109 · turn 10: context 312,708 · output 365 · 5.6 s

    Next time: Keep the prompt small: summarise long results into notes, avoid re-reading files already in context, and hand large investigations to subagents.

    2. ERROR retry-loop — mcp__claude-in-chrome__computer called 8× with identical input

    2 of them failed — the agent kept retrying the same call.

    Evidence

    • mcp__claude-in-chrome__computer · 4.2 s · turn 10: «42 chars»
    • mcp__claude-in-chrome__computer · 3.9 s · turn 10: «42 chars»
    • mcp__claude-in-chrome__computer · 20 s · turn 10: «42 chars»
    • mcp__claude-in-chrome__computer · 4.5 s · turn 10: «42 chars»
    • mcp__claude-in-chrome__computer · 2.5 s · turn 10: «42 chars»
    • mcp__claude-in-chrome__computer · 28 s · turn 10: «42 chars»
    • … 2 more

    Next time: When a call fails, change something before calling again: read the error, fix the input or the environment, or ask the user — never repeat an identical call.

    3. WARN oversized-result — mcp__memory__memory_read returned 50,681 chars

    Large tool results stay in context for the rest of the session. Consider head/limit, grep, or a subagent.

    Evidence

    • mcp__memory__memory_read · 0 ms · turn 1: «57 chars»

    Next time: Ask tools for less: head/limit/offset on reads, a tighter grep pattern, a file listing instead of a dump, or a subagent that returns a summary.

    4. WARN unverified-claim — Turn 6: a test claim with no receipt in the transcript

    «132 chars»

    Evidence

    • request #23 · turn 6: context 294,473 · output 389 · 7.2 s

    Next time: Say it after you've checked it: a test claim needs a test run after your last edit, a shipped claim needs the command's result in this transcript, a 'nothing changed' claim needs a git status after the thing that could have changed it. If you did not check, say 'not re-run' instead.

    5. WARN orphan-tool — 1 tool call never got a result

    Session aborted, crashed, or was still running when the transcript was captured: Bash

    Evidence

    • Bash · turn 10: «363 chars»

    Next time: A call with no result means the session ended mid-call; if you resume, re-run it or confirm its side effects before building on it.

    6. WARN failed-tool — mcp__claude-in-chrome__form_input failed 2 of 9 calls (22%)

    «127 chars»

    Evidence

    • mcp__claude-in-chrome__form_input (error) · 3.4 s · turn 10: «172 chars»
    • mcp__claude-in-chrome__form_input (error) · 3 ms · turn 10: «172 chars»

    Next time: Check the first error message of the failing tool before retrying; if the tool itself is broken, switch tools or tell the user instead of pushing on.

    7. INFO cache-churn — 199,370 cached tokens re-written at request #29

    The previous request had 296,611 tokens cached; this one read only 97,241 back -2310 ms later. Something early in the prompt changed (system prompt, tool list, early messages).

    Evidence

    • request #28 · turn 9: context 296,613 · output 699 · 8.7 s
    • request #29 · turn 10: context 252,079 · output 72 · 6.6 s

    Next time: Avoid changing anything at the top of the prompt mid-session (system prompt, tool list, early messages), because it invalidates the cached prefix and re-bills it.

    8. INFO long-turn — Turn 1 made 125 tool calls

    A very long autonomous stretch. Worth a skim for wasted work.

    Evidence

    • mcp__memory__memory_read · 0 ms · turn 1: «57 chars»
    • Skill · 0 ms · turn 1: «23 chars»
    • ToolSearch · 0 ms · turn 1: «56 chars»
    • mcp__Gmail__search_threads · 0 ms · turn 1: «28 chars»
    • WebSearch · 0 ms · turn 1: «37 chars»
    • mcp__Gmail__get_thread · 0 ms · turn 1: «60 chars»
    • … 119 more

    Next time: A very long autonomous turn; check in with a short progress summary at natural checkpoints so wasted work is caught earlier.

    Tools

    toolcallserrorstotal time
    mcp__claude-in-chrome__computer2433m 00s
    mcp__claude-in-chrome__javascript_tool2411m 31s
    mcp__claude-in-chrome__navigate1040 s
    mcp__claude-in-chrome__tabs_context_mcp1038 s
    mcp__claude-in-chrome__form_input9237 s
    WebFetch3125 s
    mcp__remote-devices__device_request_folder_access1024 s
    mcp__remote-devices__device_list_dir2123 s
    … 18 more tools

    What to do with this

    This is a mechanical review of the session transcript, generated by Glassbox (no AI involved). If you are the agent that ran this session: acknowledge the findings that apply, say in one or two sentences what you will do differently next session, and do not redo any work. If you are a human: the evidence lines are permalinks into the Glassbox viewer for this session.

    Generated by Glassbox 0.11.1.


    What I'll do differently this session

    • context-bloat and oversized-result are the same mistake: a 50,681-char memory read on turn 1 and page dumps have the prompt at 312k. I'll read narrower ranges and keep page text out of context.
    • retry-loop: the eight identical computer calls were screenshots timing out on a tab that wasn't on screen. I switched to reading the DOM; I'll do that first next time instead of retrying.
    • unverified-claim at turn 6: a test claim with no run behind it in this transcript. I'll run the check before saying so, or say "not re-run".

    Connects securely to your tools. The creator never sees your data.

    About this skill

    A session that cost $21.83 on the receipt. Read properly, it cost $322.36. Same transcript. Nobody looked.

    Glassbox reads Claude Code transcripts on your machine and reports the things you can't see from the chat: retry loops, tools that kept failing, oversized results, context bloat, cache churn, stalls, and claims the agent made about its own work ("tests pass", "committed", "verified") with no tool result behind them. Each finding carries its evidence and a "next time" line.

    This skill teaches your agent to run it on its own session and act on what it finds. It also:

    • reports token and cost burn, with a floor when Claude Code didn't write the usage to disk
    • searches past sessions of the project (what was asked, written, run, committed) before re-investigating something an earlier session already handled
    • opens a self-contained HTML timeline viewer
    • compares two runs of the same task
    • audits what loads before the first prompt (CLAUDE.md files, imports, dead paths, conflicting rules)
    • scans transcripts for leaked secrets, masked, never the value

    What it needs

    Node 18+ and Claude Code transcripts on the machine (~/.claude/projects/). One network call: npx fetching glassbox-trace@0.11.1 from npm (MIT, published with provenance). After that nothing leaves the machine. The CLI is local and read-only; --redact blanks prompt and tool text before a report is shared.

    Want it after every turn?

    Claude Code users can install the plugin instead, which runs at the Stop hook and can hand the findings back to the agent:

    /plugin marketplace add aldohushi1-stack/glassbox
    /plugin install glassbox@glassbox-trace
    

    Links

    By Aldo Hushi, BlueprintAU, Adelaide. MIT licence.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Testing & Debugging.

    Requires Node 18+. Last few tool calls may be missing from active transcripts. Subagent costs are estimates if usage data is missing from disk.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 22 days ago

    • Free with an Agensi account

    Software
    Node.js 18+glassbox-trace (npm, fetched by npx)Git

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access

    Terminal: runs the glassbox CLI with node. Read files: Claude Code transcripts under ~/.claude/projects/. Write files: a self-contained HTML viewer in the temp folder when asked to open one; .glassbox/ in the project only with the plugin's memory option on. Network: one npx fetch of glassbox-trace from npm; the CLI itself makes no network calls.

    Listed22 days ago

    What's inside

    Frequently Asked Questions

    It reads Claude Code transcripts, so it is useful where Claude Code runs. The skill file loads in any agent; the findings only exist for Claude Code sessions.

    One npx fetch of glassbox-trace from npm. After that, nothing. The CLI reads local files and makes no network calls. --redact blanks prompt and tool text if you share a report.

    They are estimates at API list price from the usage in the transcript. If Claude Code didn't write some subagent usage to disk (anthropics/claude-code#93620), the report says so and the figure is a floor.

    No. The skill runs on demand. The plugin runs after every turn, shows a one-screen summary, and can hand findings back to the agent and keep a memory for the next session.

    Yes. MIT on GitHub, free on npm, free here.

    New in Testing & Debugging

    Free