Extract credential fill from the worker into credential-fill.ts

[CuriosityOS]

Why

src/worker.ts was 8,808 lines with ~234 top-level functions. The credential fill path (vault RPC, form detection, trusted typing, generate/rotate/pending) was the largest self-contained slice, ~1,300 lines, and already had its own test file that exercised it only through the worker.

What

• New src/credential-fill.ts: createCredentialFill(deps) returning buildCredentials, performCredentialFill, finalizePendingCredential, recoveryFromError, vaultCall, vaultCallAtOrigin. Imports its pure helpers (credential-constants, credential-target-scan, untrusted-value, vault-capture) directly; takes only worker-owned state as deps (rpc, trackSecret, trackCredentialWriteSecrets, redactDeep, ensureSessionPage, humanClickTarget, actionTimeoutMs).
• src/worker.ts: 8,808 → 7,489 lines. Keeps buildEnvelope and the credentialFill / credentialPending message handlers, which are execution-lifecycle glue rather than credential logic.
• Code moved verbatim and re-indented. The four spec interfaces were hoisted to module level. No public API change; types/ untouched.

Verification

• bun run lint, bun run typecheck, bun run check:build: clean
• bun run test:unit: 1295 pass. One failure (local-ai custom endpoint validation) reproduces on main with this change stashed.
• bun test --timeout 120000 tests/node/credential-fill.test.ts: 17 pass, including the real-Chromium detection/rotation/pending-commit test.
• bun test --timeout 120000 tests/node/browser.test.ts: 129 pass.

Follow-ups planned as separate PRs: live-view/handoff RPC handlers, CAPTCHA action dispatcher, challenge scan.

[greptile-apps]

Confidence Score: 5/5

Safe to merge: no actionable issues were found, and the covered credential behavior remained unchanged.

What we checked:

• Ran the credential-fill-runtime-validation.sh script to exercise the runtime validation for credential filling.
• Reviewed the parent runtime capture to establish a baseline before the extraction step.
• Reviewed the extracted-runtime capture to verify the post-run state after the credential fill validation.
• Verified worker integration was exercised through the built public API rather than importing the worker source.
• Linked the evidence script and two runtime captures as artifacts to support review.

Summary

Credential-management logic is now isolated in src/credential-fill.ts while src/worker.ts retains its execution and message lifecycle responsibilities. The built public runtime preserved the covered credential flows before and after the extraction.

_{Reviews (1) · Last reviewed commit: "Extract credential fill from the worker ..."}

[devin-ai-integration]

E2E review — PR #223 @ 69afb6d7

Verdict: no defects found. Code moved verbatim (every line removed from src/worker.ts is present in src/credential-fill.ts; only the createSecretsRedactor import differs), all deps passed to createCredentialFill are hoisted function declarations or consts defined above the call site, and the real credential flows behave identically end to end.

Environment: Ubuntu 22.04 x64 · Bun 1.4.0 · BetterChromium 153.0.8010.36-r2 (betterwright setup, checksum OK) · patchright-core 1.61.1 · base main = 9c4d4bbb.

What actually ran

Layer	Command / flow	Result
Lint / typecheck	bun run lint, bun run typecheck	passed (pre-existing lint warnings only)
Unit suite	bun run test:unit	1287 pass, 2 fail — both native (win/linux) archive … packager tests, env-only (hashlib.file_digest needs Python ≥3.11; box has 3.10). Not related to this PR.
Focused suite	bun test --timeout 120000 tests/node/credential-fill.test.ts	17 pass incl. the real-Chromium detection/rotation/pending test
SDK + real vault + real browser	Disposable script: BETTERWRIGHT_HOME=/tmp/…, built-in LocalCredentialVault, local HTTP login/signup fixture, headed + headless BetterChromium	18/18 passed — credentials.save, inspect, fillCredential({submit:true}) (server received the exact stored secret, page shows "Welcome back"), typed DOM value redacted to [REDACTED_PASSWORD] in the envelope, ambiguous two-form page fails closed, generateAndFillCredential → server accepted 20-char matching new/confirm → commitGeneratedCredential → pending list cleared → committed record logs in, about:blank refused
CLI	betterwright vault list on the disposable home	shows both records created by the flows (persistence)
MCP	betterwright mcp over stdio: initialize → tools/list (browser_login present) → browser navigate → browser_login {submit:true}	login server-verified, "Welcome back", no secret in MCP output

Not applicable: no UI / types / docs changes; native artifacts untouched.

Video (headed BetterChromium, real-time, 2m06s)

Both scripted runs of the SDK flow: fill+submit login, redaction probe, ambiguous refusal, generate+fill+commit signup, re-login with the committed record.

• Real-time recording: https://app.devin.ai/api/presigned_proxy?token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJvcmdfaWQiOiJvcmctYmQ0NTliYTUyOTRiNDEyMGJiYWRhNmE1Y2JmODIyZTUiLCJ1c2VyX2lkIjpudWxsLCJidWNrZXRfbmFtZSI6ImRldmluYXR0YWNobWVudHMiLCJidWNrZXRfa2V5IjoiYXR0YWNobWVudHNfcHJpdmF0ZS9vcmctYmQ0NTliYTUyOTRiNDEyMGJiYWRhNmE1Y2JmODIyZTUvMDlhYmQ2OTktZmFhNS00NWU4LTg1MTAtZTY5YTlhODM2NjA2IiwiaWF0IjoxNzg5OTg1MjEwLCJleHAiOjE3OTA1OTAwMTAsImZpbGVuYW1lIjoicHIyMjMtY3JlZGVudGlhbC1maWxsLXJlYWx0aW1lLm1wNCJ9.P9XOdx_lhXWhiPhNUmgwjVrzYNnj4lThfzwMRpuz0zQ
• Annotated short cut: https://app.devin.ai/api/presigned_proxy?token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJvcmdfaWQiOiJvcmctYmQ0NTliYTUyOTRiNDEyMGJiYWRhNmE1Y2JmODIyZTUiLCJ1c2VyX2lkIjpudWxsLCJidWNrZXRfbmFtZSI6ImRldmluYXR0YWNobWVudHMiLCJidWNrZXRfa2V5IjoiYXR0YWNobWVudHNfcHJpdmF0ZS9vcmctYmQ0NTliYTUyOTRiNDEyMGJiYWRhNmE1Y2JmODIyZTUvYjJiZjA0NWQtNTRkMS00Y2Y3LTkyMmEtYTI5OWU1MjgyOTdkIiwiaWF0IjoxNzg5OTg1MjEwLCJleHAiOjE3OTA1OTAwMTAsImZpbGVuYW1lIjoicHIyMjMtY3JlZGVudGlhbC1maWxsLWVkaXRlZC5tcDQifQ.XF6PMeOj1e0ZA27lEGExtKa_27ic2zO3bZmwUPWIRO8

Observations (not defects in this PR)

• In headed mode BetterChromium's own "Save password?" bubble pops up after the trusted fill submits (visible in the video). This is browser-level password-manager behaviour untouched by this diff; I did not re-verify it against main, so treat it as an optional hardening note (e.g. credentials_enable_service: false pref / --password-store=basic), not a regression.

No fix PR needed.

Evidence-link limitation: GitHub direct video upload is not available from this tooling, so the MP4s are hosted as presigned Devin attachment links (token valid ~7 days from posting). The review sandbox's network allowlist blocks app.devin.ai, so I could not fetch the posted links back from inside the sandbox to confirm they resolve; the same files are also attached to the automation session log.

[devin-ai-integration]

PR #223 @ 69afb6d reviewed E2E: no defects, no fix PR needed. Report posted at #223 (comment) (lint/typecheck, unit suite, focused suite, SDK+real vault+BetterChromium 18/18, CLI vault list, MCP browser_login; 2 unit failures are env-only Python 3.10 packager tests).

Limitation: MP4s are presigned Devin attachment links (~7-day token); the sandbox allowlist blocks app.devin.ai so I couldn't re-fetch the posted links to confirm they resolve — files attached here too.

This response includes attachments — view them in Devin