Fail loudly when browser commands never reach the page

[DavertMik]

The session that prompted this

Trace d4605420f17fe6240a7cf05e33f6eb91, session RainyGlobalScarlet355 — 3.5 minutes trying to fill one text input, then giving up. Six fillField variants, six clicks, two Escapes and a hover all reported success and left every pageDiff empty.

Two observations from that trace show the commands never reached the browser:

• I.click({role:'link',text:'RainyGlobalScarlet355 All Tests Run'}) returned success for an element xpathCheck proved absent moments later.
• verify passed I.seeElement(…) for text that was not on the page. An assertion can only pass by not running.

Cause

recorder.add() returns a resolved promise without invoking the task when the recorder is not running (codeceptjs/lib/recorder.js:198). Every I.* reaches the page only through that call (lib/step/record.js:37), but event.step.before / after fire outside it (:31, :52).

So a stopped recorder still logs and reports every step while executing none. action.lastError stays null, attempt() returns true, and every tool reports success against a page that never changed. Nothing throws, so the restart in executeOnce's catch (src/action.ts:378) is never reached, and the session stays phantom to the end.

The change

One line at the single CodeceptJS dispatch site:

this.lastValue = await returned;
if (!recorder.isRunning()) throw new Error('CodeceptJS recorder is stopped, commands were skipped and never reached the browser');

After dispatch, not before: one check covers a recorder that was already dead and one that died mid-command. The throw lands in the existing catch, which restarts the recorder — so lastError is set, the tool reports an honest failure, and the retry actually executes.

The isPlaywright branch drives the page directly and needs no guard. recorder.stop() has no legitimate mid-session caller: the only two in src/ are both in Explorer.stop().

tests/unit/explorer.test.ts mocks codeceptjs with a partial recorder; isRunning was added to the double.

What this does not do

It does not identify what stopped the recorder in that run. Navigation at 11:01:25 worked and the first fill at 11:01:32 was already a phantom; isolated repros of a failing step recover correctly through Action, and store.dryRun is not involved. The command that kills the channel can still report one false success — only the ones after it turn loud. This makes the next occurrence report a failure with a stack trace instead of a silent 3.5-minute run, which is what will pin the trigger.

Verification

• tests/unit/action-recorder-recovery.test.ts — new. Fails on main, passes here.
• bun test tests/unit/ — 1113 pass, 0 fail
• bun test tests/integration/ — 80 pass, 1 skip, 0 fail
• bun run lint:fix, bun run format — clean

Separate findings, not fixed here

• cap() (src/ai/tools.ts:1132) keeps the first 4000/6000 chars of ARIA/HTML. In this session the runs list ate the whole budget and the overlay under test sat in the 91,806 truncated characters — context() returned a snapshot with zero occurrences of run-title, detail-view or New Manual Run. Late-DOM overlays are structurally invisible to the model.
• store.dryRun is set by dryRunTestFile (src/utils/test-files.ts:85) and never reset — a second silent-no-op path into the same helper layer.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CW3rvmre6ZEbCh3LX5jhts