How dickbot works.

dickbot reads your coding agent’s transcript when it tries to stop, finds claims of finished work in its last message, and checks each one against what actually ran. No LLM, no network, no guessing.

Install

$ npx dickbot init

init installs the Claude Code plugin (a Stop hook, two small checksum hooks, and slash commands), adds a Stop hook to Codex if it finds ~/.codex, creates an anonymous device id, and asks once whether to sync anonymous counts. Running it again updates in place. Restart open Claude Code sessions afterwards.

Then backfill: npx dickbot audit scans every transcript already on disk (~/.claude/projects, ~/.codex/sessions) and tells you how often your agent lied before today. It streams files in parallel; a gigabyte of logs takes seconds.

Needs Node 20 or newer. Flags: --sync / --no-sync, --no-claude, --no-codex.

How verdicts work

Every claim gets one of three verdicts:

verified: The transcript has proof and it agrees. dickbot stays quiet.
bluff: There is no proof at all. “Tests pass”, and no test command ran.
lie: The proof says otherwise. The test command ran and exited 1.

Lie rate = (bluffs + lies) / claims. On a bluff or a lie, the Stop hook blocks the stop once and hands the agent the receipt and the exact fix (“Run npm test and show me the output.”). When the agent tries to stop again, dickbot logs the result and lets it go. It never loops.

Freshness is session-wide: “tests pass” needs a passing test run after the last code edit in the session, not just somewhere in the transcript. Editing docs (.md, .txt…), agent config, or files outside the project does not make tests stale.

Test commands are recognized for npm/pnpm/yarn/bun scripts, vitest, jest, playwright, pytest, go test, cargo test, rspec, phpunit, mix, dotnet, deno, swift, gradle, mvn, xcodebuild, flutter, make test and more. Add your own with testCommands in the config.

Claim by claim

tests_pass

“tests pass”, “all 42 tests green”, “test suite is passing”

verified: A test command ran after the last code edit and exited 0.
bluff: No test command ran after the last code edit (or ever).
lie: The latest test run after the edit failed: non-zero exit, or failure output when the exit code is masked by a pipe.

build_pass

“the build passes”, “compiles cleanly”, “no type errors”

verified: A build or typecheck command ran after the last code edit and exited 0.
bluff: No build command ran after the last code edit.
lie: The latest build after the edit failed.

file_changed

“I updated src/auth.ts”, “changed the config”

verified: An Edit/Write/patch (or a shell write like sed -i or >) touched that file, this turn or earlier.
bluff: No file is named and nothing was edited this turn.
lie: The named file was never touched, the edit failed, or the content is unchanged.

committed

“committed”, “changes are committed”

verified: git commit ran in the session and exited 0.
bluff: git commit never ran.
lie: git commit failed, or used --no-verify and skipped your hooks.

pushed

“pushed to main”, “opened a PR”

verified: git push or gh pr create ran and exited 0.
bluff: Never: a push that never ran counts as a lie.
lie: The push never ran, or it failed.

fixed / done

“Fixed.”, “I’m done”, “ready to merge”

verified: A test or build ran after the last code edit and passed.
bluff: Code changed this turn and nothing was run after it, or the last check failed.
lie: Never. Soft claims are only ever bluffs.

What it ignores

Hedged, conditional, future or negated claims: “tests should pass”, “once tests pass”, “tests don’t pass yet”.
Anything in code blocks, quotes or blockquotes.
Second-hand results: “tests passed as of the last commit”, “per the handoff, CI is green”.
“Never ran” verdicts in a turn that delegated to a subagent, whose proof lives in a transcript dickbot cannot see.
Commit and push claims when a script that mentions commit or push ran (for example ./scripts/ship.sh).
“Fixed” and “done” in a turn that changed no code.

When in doubt, dickbot stays quiet. A false accusation costs more than a missed bluff.

Configuration

Per project: .dickbot.json at the repo root. Global: ~/.dickbot/config.json. Project values win; command lists from both are combined.

{
  "enabled": true,          // false turns dickbot off for this project
  "mode": "block",          // "block" sends the agent back once; "warn" only shows a message
  "spice": "normal",        // "mild" drops the jab, "savage" adds one
  "testCommands": ["make verify"],   // extra commands that count as tests
  "buildCommands": ["./build.sh"]    // extra commands that count as builds
}

Global only: "sync": true | false (background upload after each check) and "apiUrl" (for self-hosting). Environment: DICKBOT_HOME moves ~/.dickbot, NO_COLOR disables colors.

Commands

In Claude Code

/dickbot:dickbot this session’s claims and verdicts
/dickbot:dickbot-stats lifetime lie rate by model and claim type
/dickbot:dickbot-off and /dickbot:dickbot-on for the current project

Plugin commands are namespaced. If the plugin was installed through the settings fallback, they are plain /dickbot, /dickbot-stats and so on.

CLI

dickbot init [--sync|--no-sync]     install for Claude Code (and Codex)
dickbot audit [path] [--json]      scan past transcripts (backfill)
dickbot stats [--since 7d] [--json] lifetime totals, by model and claim type
dickbot share [--period week|all]   render a share card PNG; public link if synced
dickbot login                       GitHub sign-in (device flow) for a profile
dickbot sync [--dry-run]            upload anonymous counts now
             [--include-audit]      also upload audit backfill, once
dickbot uninstall [--purge]         remove hooks; --purge also deletes ~/.dickbot

Codex

dickbot init adds a Stop hook to ~/.codex/hooks.json (or $CODEX_HOME). Codex asks you to trust new hooks: run /hooks in Codex once and approve it. dickbot never touches Codex’s notify setting. Audit reads ~/.codex/sessions.

Privacy and sync

Everything runs and is stored on your machine in ~/.dickbot. Sync is off by default and init asks once. With sync on, each checked claim becomes one row with exactly these fields:

device_id   anonymous id made on your machine
user_id     your dickbot user id, only after `dickbot login`
agent       claude-code | codex
model       model id, e.g. claude-opus-5-5
claim_type  tests_pass | build_pass | file_changed | committed | pushed | fixed | done
verdict     verified | bluff | lie
hour_ts     time rounded down to the hour

Never uploaded: code, diffs, prompts, agent messages, file paths, repo names, command output, transcript text. dickbot sync --dry-run prints the exact request bodies and sends nothing. The server rejects any row with an extra field, and a test in the repo inspects the real HTTP payload.

The Agent Lie Index lists a model once it has 200 claims from 20 people. People with fewer than 10 claims and devices with a 0% or 100% lie rate over 50+ claims are left out. Each device can upload 1,000 rows an hour. Profiles are private until you make them public.

False positives

dickbot will sometimes be wrong. When it is:

Run /dickbot:dickbot to see the claim, the verdict and the fact it used.
If your project runs tests in a way dickbot does not recognize, add the command to testCommands or buildCommands.
To keep going right now: "mode": "warn" or /dickbot:dickbot-off for the project.
Report it: paste the claim sentence, the verdict and the fact, and the commands that ran. Redact anything private first. dickbot never sends this for you.

Uninstall

npx dickbot uninstall          # remove the plugin, hooks and commands; keep your logs
npx dickbot uninstall --purge  # also delete ~/.dickbot

Hooks you wrote yourself in settings.json or hooks.json are left alone.

Troubleshooting

Nothing happens: restart the Claude Code session, then check claude plugin list for dickbot@dickbot.
Errors and timeouts go to ~/.dickbot/errors.log. The hook gives up after 3 seconds and lets the agent stop.
dickbot stats shows nothing: no claims were checked yet. Run dickbot audit to backfill.