fix(executors): repair DuckDuckGo AI Chat challenge solver (418 ERR_CHALLENGE)

[Mynacol]

Summary

• Fixes the duckduckgo-web provider, which returned HTTP 418 ERR_CHALLENGE on every chat request while duck.ai worked normally in a browser from the same IP. The provider was fully unusable.
• Ground truth was established by driving a real headful Chromium at duck.ai from the same IP, which returned 200. That ruled out IP blocking / rate limiting and proved the fault was in our anti-abuse challenge solver.
• Six independent defects were found and fixed. The first alone disabled the solver outright; the rest were exposed only once it started running again.
• Also removes a redundant "seed" chat POST that doubled request volume against an IP-rate-limited endpoint and surfaced as spurious 429 ERR_RATE_LIMIT.

Root causes

#	Defect	Effect
1	export keywords inside CHALLENGE_STUBS. That string is run by vm.runInContext, which compiles in script mode, so module syntax is a hard SyntaxError. A refactor mass-added export to its 5 function declarations.	Solver threw on every call → raw unsolved challenge sent → 418
2	\\s inside a String.raw template reached the sandbox as a literal backslash	getComputedStyle display probe silently read empty
3	buildHtmlLookup subtracted 1 from the descendant count	Chromium reports 3 for <li><div></li><li></div; we sent 2
4	Sandbox failed 9 of 13 browser-fidelity probes (prototype chains, NodeList identity, live body.children, native-code toString, sloppy-mode this)	Challenge classified us as a bot
5	Solved payload omitted meta.origin / meta.stack / meta.duration, which the real duck.ai bundle always sends	418 even when every client_hash was correct
6	reasoningEffort is now mandatory on duckchat/v1/chat	400 ERR_BAD_REQUEST

Note on #4: Math must not be sealed — real Chromium reports Object.isSealed(Math) === false, and sealing it made our vector differ by one.

Related Issues

• Closes #
• Related to feat(resilience): Circuit breaker for no-auth providers — stop advertising dead models #6999 (DDG circuit breaker), fix(providers): DuckDuckGo AI Chat executor broken — VQD token acquisition fails (503) + rate limit (429) #6996 (VQD 429 misclassification), fix(providers): duckduckgo-web consistently broken (400 ERR_BAD_REQUEST) #8000 (DDG model catalog churn)

The regression was introduced by the challenge-solver extraction in #5925 (Release v3.8.44), which added export to the functions inside CHALLENGE_STUBS while moving them.

Validation

• Change type: provider
• Focused tests and category gates from the golden path
• npm run lint
• Reconciled with the current active release base; focused checks rerun afterward
• Production-code changes include a new or updated automated test in this PR
• SonarQube PR analysis is green or any remaining issues are explicitly documented below — not runnable locally; deferred to CI

Commands run (Node 22.23.2 provisioned via nix, since the sandbox had no system Node):

# Provider golden-path gates
npm run check:provider-consistency   # OK — 225 REGISTRY entries, 304 canonical providers
npm run check:provider-assets        # Provider asset budget passed

# Focused DuckDuckGo + adjacent suites — 177/177 pass
node --import tsx/esm --test \
  tests/unit/duckduckgo-*.test.ts \
  tests/unit/ddg-circuit-breaker-null-content-6999-7000.test.ts \
  tests/unit/executor-web-cookie-sweep.test.ts \
  tests/unit/search-handler-duckduckgo.test.ts \
  tests/unit/free-web-search.test.ts \
  tests/unit/session-pool-modular.test.ts \
  tests/unit/noauth-provider-validation.test.ts

# Pre-commit gates (hooks were skipped locally: npx not on PATH)
npm run lint
npm run check:any-budget:t11         # PASS
npm run check:tracked-artifacts      # OK
npm run check:docs-sync              # PASS
npx prettier --check <changed files> # All matched files use Prettier code style

Live end-to-end validation — the executor now returns 200 against real DuckDuckGo, 4/4 cases:

Case	Result
non-streaming, gpt-4o-mini (alias path)	200 → "OK"
streaming, gpt-5.4-mini	200 → "OK"
claude-haiku-4-5 (reasoningEffort: low)	200 → "OK"
math prompt, gpt-5.4-nano	200 → "42"

This satisfies the AGENTS.md hard-rule #18 bug-fix gate via both paths: automated regression tests (TDD) and a documented live real-environment test.

Tests Added Or Updated

• tests/unit/duckduckgo-challenge-solver-regression.test.ts — new, 32 tests
• tests/unit/duckduckgo-reasoning-effort-required.test.ts — new, 5 tests
• tests/unit/duckduckgo-challenge-split.test.ts — updated, +4 tests guarding the script-mode/export invariant
• tests/fixtures/duckduckgo/challenge-variants.json — new fixture: 8 real challenge programs captured from duckduckgo.com, each paired with the probe vectors a real headful Chromium produced for that exact program

The suite asserts against recorded real-browser behaviour, not against our own output — matching Chromium bit-for-bit is the actual correctness criterion here. Every fix was individually reverted to confirm its test fails:

Reverted fix	Failing tests
re-add export to stubs	30
re-break \\s escaping	3
restore count - 1	4
drop meta augmentation	2
seal Math	4
reasoningEffort: null	2

Coverage Notes

This PR changes open-sse/executors/duckduckgo-web.ts and open-sse/executors/duckduckgo-web/challenge.ts.

• solveDuckDuckGoChallenge, CHALLENGE_STUBS, and buildHtmlLookup are covered by the 8 real-variant fixtures plus 16 individually-named browser-fidelity probe tests, so a future stub regression names itself instead of failing as one opaque hash mismatch.
• The reasoningEffort change and the seed-POST removal are covered by duckduckgo-reasoning-effort-required.test.ts, which stubs fetch and asserts on the actual outgoing wire payload (including a test that exactly one POST /duckchat/v1/chat is issued per user request).
• Coverage moves up on both touched files: previously the solver had no test that executed the sandbox at all — which is precisely why a SyntaxError inside CHALLENGE_STUBS shipped unnoticed.

Reviewer Notes

Base-red inherited — not caused by this PR. Three failures pre-exist on origin/release/v3.8.50 and reproduce identically on a pristine checkout of that tip:

1. tests/unit/models-catalog-route.test.ts — ERR_MODULE_NOT_FOUND: open-sse/services/antigravityProjectPersistence.ts. The file is imported by open-sse/services/combo/quotaStrategies.ts but does not exist in git at that ref.
2. tests/unit/provider-translate-path-golden.test.ts — golden drift from devin-cli-agentic and raycast missing from the snapshot. Verified the duckduckgo-web entry is byte-identical and untouched by this PR.
3. npm run lint — 3 no-explicit-any errors in tests/unit/vertex-functioncall-id-3440.test.ts, a file this PR does not touch.

Per CONTRIBUTING.md these are not fixed here; a base-red fix belongs in its own fix/release-v3.8.50-basereds-* PR.

Risk areas

• vm sandbox is a supply-chain surface. solveDuckDuckGoChallenge executes upstream-supplied JS. The security posture is unchanged — still vm.runInContext with the 5s timeout — and the SECURITY NOTE remains. This PR only makes the emulated browser environment more faithful; it does not widen what the sandbox can reach.
• The stubs are anti-bot-sensitive. DuckDuckGo rotates challenge variants (8 distinct programs observed in one session). If upstream adds a new probe this can regress again. The fixtures make that failure mode legible: a new variant will show as a specific probe mismatch against recorded browser behaviour.
• Two comment-formatting traps inside CHALLENGE_STUBS, both hit during development and now documented inline: (a) a backtick in a comment terminates the String.raw template; (b) \\x in a comment is still a double escape. A WARNING block at the top of the literal spells out the script-mode constraint.
• Behaviour change: every request now sends reasoningEffort ("none" by default, "low" for claude-haiku-4-5 / tinfoil/gpt-oss-120b). Verified live as strictly required — the previous "let the server default it" path returns 400.

No migrations, no feature flags, no DB changes.

Manual validation reviewers may want to repeat: the live 4/4 table above. Note DuckDuckGo aggressively IP-rate-limits (429 ERR_RATE_LIMIT) under repeated testing; that is environmental and distinct from the 418 this PR fixes.

[Mynacol]

I manually tested the DuckDuckGo provider on a fully running omniroute server.

[diegosouzapw]

Thanks for the PR. Please address the mandatory items (tests and/or merge blockers) in this branch, then rerun checks before /merge-prs.

[diegosouzapw]

Thanks for the PR. Please address the mandatory items (tests and/or merge blockers) in this branch, then rerun checks before /merge-prs.

[diegosouzapw]

Thanks for the PR. Please address the mandatory items (tests and/or merge blockers) in this branch, then rerun checks before /merge-prs.

[diegosouzapw]

Obrigado pelo PR. Mantive a revisão de fix-in-place e não foi possível concluir o ajuste completo aqui:

• Para os PRs em fork: não consigo aplicar push de correção diretamente na sua branch.
  Por favor, faça um rebase/sync com release/v3.8.50, resolva conflitos se houver, e rode os checks dessa branch.
  Se preferir, posso aplicar a correção na próxima rodada assim que você mandar o branch atualizado ou confirmar que o PR está limpo pra esse merge.