Find the token
that kills cache.

╭ Cachecatch Local ────────────────────────────────────────────────────────────────────────────────────────╮ │ │ │ Cachecatch Local Agent Report │ │ Generated 2026-07-01T12:00:00Z · time window 7d │ │ │ │ Main finding: Cachecatch analyzed 89 local agent sessions across Claude Code, OpenCode, Codex. Token │ │ activity is mixed; cost telemetry is unavailable and pricing coverage is unavailable. The observed │ │ Claude Code, OpenCode cache profile is 9% (LOW). Missing cache fields are not counted as 0%. The main │ │ fixable signal is context hygiene in the transcript-visible sessions: git diffs, terminal output, and │ │ timestamps appear before repo instructions in 68% of sessions │ │ Cachecatch found 342 stored sessions total and 89 sessions in this time window. "Analyzed" means │ │ Cachecatch could parse the session format safely. │ │ │ │ ╭ Builder Scorecard ───────────────────────────────────────────────────────────────────────────────────╮ │ │ │ AGENTIC SESSIONS 89 parsed in this time window │ │ │ │ TOKEN ACTIVITY 2,840,000 mixed observed/estimated │ │ │ │ TOOL CALLS 1,247 observed where local storage exposes tool records │ │ │ │ SUBAGENT RUNS 38 explore (20), general (10), writer (6), editor (2) │ │ │ │ IDE AGENTS USED Claude Code, OpenCode, Codex │ │ │ │ MODELS DETECTED 4 │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ Cache Profile ───────────────────────────────────────────────────────────────────────────────────────╮ │ │ │ CACHE READ PROFILE 9% LOW │ │ │ │ CACHE READ TOKENS 186,000 observed │ │ │ │ CACHE WRITE TOKENS 412,000 observed │ │ │ │ CACHE-TOKEN TELEMETRY unavailable │ │ │ │ COST TELEMETRY unavailable │ │ │ │ PRICING COVERAGE unavailable │ │ │ │ TRANSCRIPT HYGIENE SIGNALS available │ │ │ │ │ │ │ │ Scope observed on Claude Code, OpenCode only; transcript-only agents are excluded. │ │ │ │ Trust CacheCatch does not treat missing cache telemetry as zero. Missing fields are reported as not │ │ │ │ reported. │ │ │ │ Limit Claude/Codex local transcripts are not counted as 0%. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ Context Hygiene ─────────────────────────────────────────────────────────────────────────────────────╮ │ │ │ CONTEXT HYGIENE SCORE 52 / 100 WATCH ████████████░░░░░░░░░░░░ │ │ │ │ EFFICIENCY UPSIDE SIGNAL HIGH based on evidence-backed hygiene findings │ │ │ │ RESEARCH BENCHMARK RANGE 41-80% cost reduction in agentic workloads; not this audit's claim │ │ │ │ KNOWN MODEL COST $12.40 equivalent observed where local telemetry exposes cost │ │ │ │ RECOVERABLE TOKEN GAP 12-18% signal dollar estimate unavailable because cost/pricing coverage is │ │ │ │ partial │ │ │ │ BILLING CONFIDENCE PARTIAL subscriptions/promos/enterprise pricing not visible │ │ │ │ │ │ │ │ Caveat Dollar values are model-cost equivalents only. │ │ │ │ Basis Unpriced or partially priced token opportunity is not converted to dollars. │ │ │ │ Pricing Conservative cache-discount range used where exact cached-input price was unavailable. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ Report Confidence ───────────────────────────────────────────────────────────────────────────────────╮ │ │ │ Confidence MEDIUM │ │ │ │ Medium confidence: token/cache direction is useful, but some agents or models are estimated or not │ │ │ │ priced. │ │ │ │ Use sessions, observed token/cache rows, model mix, and visible context-hygiene findings as the │ │ │ │ reliable parts. Unsafe dollar values are suppressed. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ■ IDE AGENT BREAKDOWN │ │ Stored = all local sessions found. In time window = sessions whose modified/update time is inside the │ │ requested time window. Analyzed = parsed safely. │ │ │ │ ╭ Claude Code (detected) ──────────────────────────────────────────────────────────────────────────────╮ │ │ │ Visibility telemetry-rich cache telemetry not reported │ │ │ │ Sessions 42 analyzed 42 in time window · 156 stored │ │ │ │ Tokens 1,420,000 observed from Claude local token events │ │ │ │ Cache telemetry not reported no explicit cache fields observed │ │ │ │ Cache read not reported missing cache field │ │ │ │ Input tokens 1,060,000 observed │ │ │ │ Cached input not reported missing cache field │ │ │ │ Output tokens 360,000 observed │ │ │ │ Cost $6.80 observed/estimated from telemetry │ │ │ │ Source OTel logs │ │ │ │ Context hygiene 48 / 100 WATCH │ │ │ │ Tool calls 624 │ │ │ │ Subagents 18 explore (12), general (6) │ │ │ │ Top models claude-sonnet-4-20250514, claude-haiku-3.5 ranked by parsed sessions │ │ │ │ Finding Git diffs and terminal output load before AGENTS.md rules in 71% of sessions, breaking │ │ │ │ prefix stability. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ OpenCode (detected) ─────────────────────────────────────────────────────────────────────────────────╮ │ │ │ Visibility telemetry-rich cache telemetry not reported │ │ │ │ Sessions 38 analyzed 38 in time window · 128 stored │ │ │ │ Tokens 980,000 observed from OpenCode local database │ │ │ │ Cache telemetry not reported no explicit cache fields observed │ │ │ │ Cache read not reported missing cache field │ │ │ │ Input tokens 740,000 observed │ │ │ │ Cached input not reported missing cache field │ │ │ │ Output tokens 240,000 observed │ │ │ │ Cost $4.10 observed/estimated from telemetry │ │ │ │ Source local database │ │ │ │ Context hygiene 58 / 100 WATCH │ │ │ │ Tool calls 486 │ │ │ │ Subagents 14 explore (8), writer (6) │ │ │ │ Top models gpt-4.1, gpt-4.1-mini ranked by parsed sessions │ │ │ │ Finding Timestamps and session metadata inject before stable project instructions, reducing prefix │ │ │ │ reuse. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ Codex (detected) ────────────────────────────────────────────────────────────────────────────────────╮ │ │ │ Visibility transcript-rich / cache-telemetry limited cache telemetry not reported │ │ │ │ Sessions 9 analyzed 9 in time window · 58 stored │ │ │ │ Tokens 440,000 estimated from transcript text length │ │ │ │ Cache telemetry not reported no explicit cache fields observed │ │ │ │ Cache read not reported missing cache field │ │ │ │ Input tokens 320,000 estimated │ │ │ │ Cached input not reported missing cache field │ │ │ │ Output tokens 120,000 estimated │ │ │ │ Cost $1.50 observed/estimated from telemetry │ │ │ │ Source transcript │ │ │ │ Context hygiene 64 / 100 WATCH │ │ │ │ Tool calls 137 │ │ │ │ Subagents 6 general (4) │ │ │ │ Top models codex-mini-latest ranked by parsed sessions │ │ │ │ Finding Cache telemetry not exposed in local Codex session files; hygiene score based on transcript │ │ │ │ analysis. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ■ TELEMETRY SETUP │ │ ╭ Enable Future Exact Telemetry ───────────────────────────────────────────────────────────────────────╮ │ │ │ Run `npx --yes cachecatch init claude`, then start Claude Code with the generated env file to enable │ │ │ │ future cache/token telemetry. │ │ │ │ Run `npx --yes cachecatch init codex` to enable future Codex OTel telemetry. │ │ │ │ Then run `npx --yes cachecatch daemon` while using the agent. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ■ TOP PROJECTS │ │ Project advice is based on the local project folders found in agent storage, plus AGENTS.md / CLAUDE.md │ │ presence. │ │ │ │ ╭ ~/code/support-copilot ──────────────────────────────────────────────────────────────────────────────╮ │ │ │ Sessions 38 │ │ │ │ Token activity 1,180,000 │ │ │ │ Cache read 9% observed │ │ │ │ AGENTS.md present keep stable │ │ │ │ CLAUDE.md present Claude-specific rules available │ │ │ │ → AGENTS.md is present and looks stable, but observed cache-read is only 9% — the leakage is below │ │ │ │ the instruction files, in task context. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ ~/code/docs-rag ─────────────────────────────────────────────────────────────────────────────────────╮ │ │ │ Sessions 28 │ │ │ │ Token activity 820,000 │ │ │ │ Cache read 7% observed │ │ │ │ AGENTS.md missing add stable repo rules │ │ │ │ CLAUDE.md missing optional unless Claude Code is used here │ │ │ │ → 28 sessions in this project started from a clean room; they re-built the same project context │ │ │ │ every run. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ ~/code/refund-service ───────────────────────────────────────────────────────────────────────────────╮ │ │ │ Sessions 23 │ │ │ │ Token activity 840,000 │ │ │ │ Cache read not reported not reported │ │ │ │ AGENTS.md present keep stable │ │ │ │ CLAUDE.md missing optional unless Claude Code is used here │ │ │ │ → Keep AGENTS.md byte-stable across sessions. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ■ TOP FIXABLE ISSUES │ │ │ │ ╭ Dynamic context loads before stable rules ───────────────────────────────────────────────────────────╮ │ │ │ What we found: git diffs, terminal output, and timestamps appear before repo instructions in 68% of │ │ │ │ sessions │ │ │ │ Why it matters: Prompt caching is prefix-sensitive. If diffs, logs, timestamps, and tool output │ │ │ │ appear before stable repo rules, the reusable prefix changes sooner. │ │ │ │ Where: Claude Code + OpenCode + Codex · 89 sessions · top projects: ~/code/support-copilot, │ │ │ │ ~/code/docs-rag │ │ │ │ What to do next: Put AGENTS.md/CLAUDE.md first, push diffs and logs to the tail. │ │ │ │ Validation: Rerun the same project/window and compare cache-read %, token activity basis, and │ │ │ │ whether this finding still appears. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ AGENTS.md missing or minimal in key projects ────────────────────────────────────────────────────────╮ │ │ │ What we found: docs-rag has no AGENTS.md; refund-service has no CLAUDE.md │ │ │ │ Why it matters: Agent instruction files are the best place for stable context: repo layout, │ │ │ │ commands, constraints, style, and testing expectations. │ │ │ │ Where: Claude Code + OpenCode + Codex · 89 sessions · top projects: ~/code/support-copilot, │ │ │ │ ~/code/docs-rag │ │ │ │ What to do next: Add stable repo rules to both projects. │ │ │ │ Validation: Rerun the same project/window and compare cache-read %, token activity basis, and │ │ │ │ whether this finding still appears. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ Repeated boilerplate across sessions ────────────────────────────────────────────────────────────────╮ │ │ │ What we found: Repo layout and command hints repeated in 19 of 89 sessions │ │ │ │ Why it matters: Repeating repo instructions inside ad hoc prompts wastes context and makes every │ │ │ │ agent rebuild the same prefix instead of reusing one stable project file. │ │ │ │ Where: Claude Code + OpenCode + Codex · 89 sessions · top projects: ~/code/support-copilot, │ │ │ │ ~/code/docs-rag │ │ │ │ What to do next: Use a shared instruction file instead of repeating boilerplate. │ │ │ │ Validation: Rerun the same project/window and compare cache-read %, token activity basis, and │ │ │ │ whether this finding still appears. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ Cache telemetry unavailable for Codex ───────────────────────────────────────────────────────────────╮ │ │ │ What we found: Codex local sessions do not expose cache-read/cache-write token fields │ │ │ │ Why it matters: This is a measurement limitation, not proof of bad Codex/Claude behavior. The honest │ │ │ │ output is 'not reported' until local files expose cache-read/cache-write tokens. │ │ │ │ Where: Claude Code + OpenCode + Codex · 89 sessions · top projects: ~/code/support-copilot, │ │ │ │ ~/code/docs-rag │ │ │ │ What to do next: Use transcript-based hygiene findings as the reliable signal. │ │ │ │ Validation: Rerun the same project/window and compare cache-read %, token activity basis, and │ │ │ │ whether this finding still appears. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ■ BEFORE / FIX / AFTER │ │ ╭ 1 BEFORE - Local Context Today ──────────────────────────────────────────────────────────────────────╮ │ │ │ × git diffs, terminal output, and timestamps appear before repo instructions in 68% of sessions │ │ │ │ × Prompt caching is prefix-sensitive. If diffs, logs, timestamps, and tool output appear before │ │ │ │ stable repo rules, the reusable prefix changes sooner. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ 2 CACHECATCH FIX ────────────────────────────────────────────────────────────────────────────────────╮ │ │ │ → Put stable repo identity, rules, tools, and output conventions in AGENTS.md / CLAUDE.md. │ │ │ │ → Keep that stable block byte-stable across sessions. │ │ │ │ → Put AGENTS.md/CLAUDE.md first, push diffs and logs to the tail. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ╭ 3 AFTER - Cache-Ready Local Agent Workflow ──────────────────────────────────────────────────────────╮ │ │ │ ✔ Stable prefix first: role, repo rules, architecture constraints, command policy. │ │ │ │ ✔ Dynamic tail last: task notes, terminal output, diffs, errors, current state. │ │ │ │ ✔ Validation: rerun audit and compare observed cache-read %, token basis, and whether the same │ │ │ │ finding remains. │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ■ PUBLIC SHARE SUMMARY │ │ ╭ Public Share Summary ────────────────────────────────────────────────────────────────────────────────╮ │ │ │ 89 agentic sessions in 7d │ │ │ │ 2.84M token activity analyzed · mixed observed/estimated │ │ │ │ 1,247 tool calls │ │ │ │ 38 subagent runs │ │ │ │ 9% observed cache-read profile │ │ │ │ 4 models detected │ │ │ │ │ │ │ │ npx --yes cachecatch share ./reports/<report>.json │ │ │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ │ │ │ │ ■ AGENT REPAIR PROMPT │ │ You are improving local agent prompt-cache hygiene. │ │ │ │ 1. Dynamic context loads before stable rules. │ │ Evidence: git diffs, terminal output, and timestamps appear before repo instructions in 68% of sessions │ │ Fix: Put AGENTS.md/CLAUDE.md first, push diffs and logs to the tail. │ │ │ │ 2. AGENTS.md missing or minimal in key projects. │ │ Evidence: docs-rag has no AGENTS.md; refund-service has no CLAUDE.md │ │ Fix: Add stable repo rules to both projects. │ │ │ │ Validate by rerunning CacheCatch for the same window and checking whether the same findings remain. │ │ │ │ ■ DISCLAIMER │ │ Dollar values are model-cost equivalents only. Subscriptions, promotions, and enterprise pricing are not │ │ visible. Cache-discount ranges are conservative where exact cached-input pricing was unavailable. │ │ │ ╰────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

╭ ♥ Support + Share ────────────────────────────────────────────────────────────────────────────────────────╮ │ Support Cachecatch: copy and run npx --yes cachecatch share to generate your share banner. │ │ It will ask for your X handle, make the PNG, and print ready-to-use X copy/link. │ ╰────────────────────────────────────────────────────────────────────────────────────────────────────────────╯