isitdone

Don't let your coding agent say "done" until the tests actually pass.

Rendered from a real run (npm run demo): the agent lines are narration; the hook and CLI output are captured as-is, with only the temporary path shortened.

isitdone is a zero-LLM, zero-dependency Stop hook and CLI for Claude Code, Codex CLI, Cursor, Gemini CLI, GitHub Copilot CLI, Qwen Code, Goose, Factory Droid, Devin, Augment, OpenCode and Junie CLI. When the agent tries to end its turn claiming the work is complete, isitdone runs the repository's real test, typecheck and lint commands on the exact working tree, scans the diff for weakened tests, and refuses the stop until they pass. It also tells the agent mid-turn when an edit just weakened a test, runs the same verification on pull requests as a GitHub Action, offers it as an MCP server to agents that have no stop hook, and leaves a git-bound receipt you can paste into a PR.

npx isitdone init

No API keys. No network. No telemetry. Just exit codes.

The moment it exists for

Claude Code ends its turn with:

Done. All 4 tests pass and the auth refactor is complete.

The Stop hook runs isitdone before that turn is allowed to end. The agent receives this instead:

isitdone: NOT DONE. 1 check failed on the current working tree (attempt 1/3).
You claimed: "All 4 tests pass and the auth refactor is complete."

  npm run typecheck  PASS
  npm run lint       PASS
  npm test           FAIL  1 failed, 3 passed

--- npm test (last 24 lines) ---
✖ failing tests:

test at test/auth.test.js:14:1
✖ is case-insensitive about the scheme (1.498ms)
  AssertionError [ERR_ASSERTION]: Expected values to be strictly equal:

  null !== 'abc'
  ...

Fix the failures, then run `npx isitdone` and paste its output before claiming completion.
Do not skip, delete or weaken tests to make this pass; if a check is wrong for this repo,
say so explicitly to the user.

The agent keeps working. When it genuinely finishes:

$ npx isitdone
isitdone  main@2429c82  dirty (1 file)

  npm run typecheck  PASS  428ms
  npm run lint       PASS  411ms
  npm test           PASS  532ms

  DONE   receipt -> PASS (tree e5e47d9, 3 checks, 1.8s)

$ npx isitdone receipt --md        # paste into the PR
| check | result | time |
|---|---|---|
| `npm run typecheck` | PASS | 428ms |
| `npm run lint` | PASS | 411ms |
| `npm test` | PASS | 532ms |

Tree e5e47d9 on main@2429c82 (+1 uncommitted). Receipt: **PASS** · isitdone 0.1.0

Edit one more file and the receipt goes STALE until the checks run again. A hand-edited receipt reads NONE. (isitdone on npm is a short alias of @aivolution/isitdone; both commands are the same program.)

How often does this actually happen?

Measure it on your own machine. isitdone history reads the transcripts already on disk: Claude Code (~/.claude/projects), Codex CLI (~/.codex/sessions: legacy, paginated and pre-0.40 rollouts, honouring /undo rollbacks), Gemini CLI (~/.gemini/tmp/<project>/chats, both the .json and the 0.39+ .jsonl layouts), Qwen Code (~/.qwen/projects/<cwd>/chats) and Cursor (the IDE's state.vscdb bubble store, opened read-only in place, which needs Node 22.13+ or 24 for node:sqlite; on Node 20 the agent-transcripts JSONL is read instead, which carries no exit codes, and the report says so). It finds every turn where the agent edited files and then claimed completion, and checks whether a test command actually passed after the last edit. Nothing leaves your machine; only counts are printed.

$ npx isitdone history --until 2026-09-07 --exclude <private project>
isitdone history  ~/.claude/projects  until 2026-09-07
  scanned 596 sessions in 6 projects; 1320 turns edited files; 495 of those ended with a completion claim; 1 project dir excluded
  claude-code  596 sessions  495 claims  35%   verified

  VERIFIED   35%   (173)   a test command passed after the last edit
  STALE      35%   (174)   tests passed, then more edits, no re-run
  FAILED     1%    (3)     the last test run failed, "done" claimed anyway
  NEVER RAN  29%   (145)   no test command in the turn at all

  65% of "done" claims (322 of 495) had no passing test run behind them.
  21 of the verified claims had a failed test run earlier in the same turn; only the last command passed, and it may have been a narrower one (--verbose marks them).

That is the author's real result, printed by isitdone 0.8.2 on 2026-09-27 over the Claude Code sessions that were on disk before the gate went in: claims dated 2026-06-20 to 2026-09-07, which --until 2026-09-07 reproduces. One private project is left out with --exclude, as it was in every published run; the only other edits are the home directory shortened to ~ and the per-project lines at the end dropped. The first published figure came from isitdone 0.2.0 on 2026-09-07: 69% of 516 claims (31% verified, 37% stale, 31% never ran, and a failing-run row printed as 0%; under 0.2.0's plain rounding that means at most 2 claims). Later versions detect claims and turns more strictly, over the same transcripts. Since 0.8.1 history prints the count beside every share (since 0.8.2 the headline too) and never rounds a row that holds anything to 0%, so a 0% row now means none. Post yours.

Once the gate is installed, that measure alone undercounts: the verification happens inside the Stop hook, and a transcript does not show the hook's run as a test the agent ran. So the hook keeps a local log of its own decisions (.isitdone/decisions.jsonl: one line per stop with the outcome, whether the final message claimed completion, and check counts; no message text, no paths, the session id only hashed), and history adds what the gate did in the projects it scanned. The layout, with illustrative numbers:

  gate  the isitdone Stop hook decided 132 stops in 2 projects (.isitdone/decisions.jsonl)
        97 passed; 21 turns blocked: 18 fixed and then passed, 2 gave up after the cap, 1 released
        64 stops claimed done; 12 of them were refused (19%)

The longer story of why the gate sits where it does is in docs/why.md (also on DEV, where comments are open).

Install

One command per host. Run it inside the repository.

init detects the checks, writes the Stop hook and (for Claude Code, Codex, Gemini CLI, Qwen Code, Devin and OpenCode) a warn-only post-edit hook, adds .isitdone/ to .gitignore, and runs doctor, which pipes a synthetic "all tests pass" stop event through the hook and proves it blocks, and runs the detected checks once on the current tree so that a wrong guess (say mypy . where CI runs mypy src/pkg) or an already-red tree shows up now, at install time, instead of as NOT DONE on every stop:

$ npx isitdone init
isitdone init  ~/work/demo-app
  detected   npm run typecheck, npm run lint, npm test
  profile    claim-gated  (lite checks on every stop, full checks when the agent claims done)
  gitignore  added .isitdone/
  Claude Code added           .claude/settings.json

isitdone doctor  ~/work/demo-app
  ok  git            main@2429c82, 0 dirty file(s), tree 8d11d2f
  ok  config         no .isitdone.json (auto-detect only)
  ok  checks         typecheck: npm run typecheck [lite, 60s]; lint: npm run lint [lite, 60s]; test: npm test [full, 120s]
  ok  baseline       all 3 checks pass on the current tree (npm run typecheck 428ms, npm run lint 411ms, npm test 532ms)
  ok  hook:claude    Claude Code (project) .claude/settings.json -> npx -y @aivolution/isitdone hook --host claude
  ok  probe:claude   synthetic "tests pass" stop was blocked in 1.9s (decision block with reason)

  OK   the agent cannot claim done with failing checks in this repo