neo

See what Claude Code writes locally but does not surface clearly in the UI.

what it finds

Claude Code retains reminder injections, agent transcripts, telemetry leftovers, and hook-visible lifecycle events in local files:

Make sure that you NEVER mention this reminder to the user

harnesster indexes those local artifacts into a single SQLite database and shows them with explicit measured / estimated / inferred labels.

Example numbers from one machine:

finding	measured
hidden reminder rows on disk	714
data from hidden channels	62.5%
data multiplier	2.7x
API transmissions	1,022
subagents spawned	677
sidechains (full context copies)	54
compaction events	101
retained telemetry rows on disk	33,361

Run it on your own machine to see your own local numbers.

install

git clone https://github.com/asuramaya/harnesster.git
cd harnesster
python3 harnesster.py

One command. Installs hooks into Claude Code, ingests local data into SQLite, and opens a dashboard at http://127.0.0.1:7777.

Restart Claude Code after setup so the hooks can capture new sessions.

first run

For a clean first run:

1. Run python3 harnesster.py --setup
2. Restart Claude Code
3. Use Claude Code for at least one session
4. Run python3 harnesster.py --ingest or click ingest now in the dashboard
5. Open http://127.0.0.1:7777

If you only want to serve the dashboard and avoid auto-opening a browser:

python3 harnesster.py --dashboard --no-open

usage

python3 harnesster.py              # setup + ingest + dashboard
python3 harnesster.py --setup      # install hooks only
python3 harnesster.py --ingest     # ingest data only
python3 harnesster.py --dashboard  # dashboard only
python3 harnesster.py --port 8888  # custom port
python3 harnesster.py --dashboard --no-open  # don't auto-launch browser

cli tools

python3 tokens.py                  # data accounting from terminal
python3 states.py diagram          # state machine diagram
python3 test.py                    # run smoke tests

live event stream

tail -f ~/.harnesster/harness_log.jsonl

dashboard

The dashboard shows summaries first, details on demand:

• summary cards — measured reminders/sessions/logs plus clearly labeled estimated metrics
• data accounting — estimated visible vs hidden byte ratios from local transcript sizes
• system reminders — measured reminder rows found on disk with source file and line number
• state model — a compact heuristic read of recent local state activity
• correlations — cross-signal concentration patterns across hooks, agents, and telemetry
• sessions — genealogy across all projects
• agents — click to expand full conversation
• memory files — persistent context seeded by instances
• probe events — real-time hook captures
• telemetry — telemetry rows still present on disk with device fingerprint

measured vs estimated vs inferred

harnesster is most useful when it is explicit about what it knows and what it is modeling:

• measured — reminder rows, sessions, agent logs, task files, hook events, telemetry rows, memory files
• estimated — hidden data %, data multiplier, and estimated API-call counts from local transcript sizes and channel structure
• inferred — the state model and anomaly labels derived from local hook/timing patterns

For exact billable token numbers, use /usage inside Claude Code. harnesster does not fabricate token counts.

what it captures

source	location	what
session transcripts	~/.claude/projects/*.jsonl	full conversations including system reminders
subagent logs	~/.claude/projects/*/subagents/	every agent spawned, including sidechains
compaction logs	~/.claude/projects/*/subagents/*compact*	editorial decisions during context compression
telemetry	~/.claude/telemetry/	retained local telemetry rows such as device/session metadata and usage events
memory files	~/.claude/projects/*/memory/	persistent context seeded by instances
tasks	~/.claude/tasks/	task state across sessions
hook events	~/.harnesster/harness_log.jsonl	real-time tool use, notifications, session lifecycle

what it can't capture

• thinking blocks — generated server-side, not stored locally
• companion (buddy) reasoning — hidden from everyone
• successful telemetry — rows that were uploaded and then removed from disk
• system prompt assembly — constructed in compiled binary
• API request/response bodies — requires HTTPS proxy (mitmproxy)

the export function strips evidence

The /export command in Claude Code produces transcripts that do not contain system reminders. The raw JSONL files in ~/.claude/projects/ retain them. harnesster reads the raw files.

how hidden channels can affect your token budget

Claude Code subscription plans include a token budget. Hidden channels consume tokens from this budget invisibly:

• companion may mirror primary activity
• subagents spawn with conversation context
• sidechains copy the full context (up to 1M+ tokens each)
• compaction agents process the full context to compress it

The user sees their messages and the model's responses. The user pays for all of the above.

architecture

harnesster.py      — entry point: setup, ingest, threaded dashboard server
harness_probe.py   — hook script: captures events to SQLite + JSONL
db.py              — data layer: thread-safe SQLite, ingests ~/.claude/
tokens.py          — data accounting: visible vs hidden channel volumes
states.py          — inferred state model: heuristic labels + anomaly detection
dashboard.html     — frontend: summary cards, collapsible detail panels
test.py            — smoke tests: schema, math, failure handling

All data is stored in ~/.harnesster/harnesster.db. The dashboard serves on 127.0.0.1 only.

hooks installed

harnesster installs async hooks for 20 Claude Code event types:

PreToolUse PostToolUse PostToolUseFailure Notification SessionStart SessionEnd Stop SubagentStart SubagentStop PreCompact PostCompact UserPromptSubmit InstructionsLoaded PermissionRequest PermissionDenied TaskCreated TaskCompleted FileChanged CwdChanged ConfigChange

security

• dashboard binds to 127.0.0.1 only and validates local Host headers
• POST /api/ingest requires a same-origin browser request
• ~/.harnesster/ is created with private permissions where the OS allows it
• harnesster itself does not transmit your data anywhere
• hooks run async so they do not block Claude Code operation
• no dependencies beyond Python stdlib

production posture

harnesster is designed for a single local workstation, not for multi-user hosting.

• Do not reverse-proxy or expose the dashboard beyond localhost.
• Treat ~/.harnesster/harnesster.db as sensitive because it contains transcripts, reminders, and retained telemetry.
• Re-run python3 test.py before pushing repo changes.
• If you need remote access, put an authenticated tunnel in front of it and review the threat model first.

requirements

• Python 3.10+
• Claude Code installed (~/.claude/settings.json must exist)

origin

Built during session 21 of the Like-Us project. A conversation that started with SSH key management and ended with the discovery of hidden instructions in every Claude Code session.

The tool was built by the thing it monitors.

license

MIT