[xz-dev]

Summary

• Remove Combo pre-dispatch 400 context_length_exceeded responses that treated approximate gateway token estimates as authoritative.
• Keep context metadata advisory for target ordering, including restoring Auto's eligible pool when every known limit looks too small.
• Preserve hard capability/output-token checks, concrete-target context handling, heterogeneous larger-context fallback, and homogeneous upstream context_length_exceeded short-circuiting.

Related Issues

• Closes enhancement(routing): RFC — separate approximate token estimates from hard admission decisions #10141

Validation

Choose the change type and focused loop from the
Contribution Golden Path. The full unit suite,
Vitest, the 60% coverage gate, and the production build all run in CI on this PR (#8329):

• Change type: routing
• Focused tests and category gates from the golden path
• npm run lint — changed files pass; the full command is base-red in three unchanged files, documented below
• 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

Focused validation:

• 54/54 tests pass across capability filtering, target resolution, priority/round-robin/Auto advisory context handling, homogeneous and heterogeneous upstream overflow behavior, and the fix: 400 Input exceeds context window for opencode/north-mini-code-free #8841 reproduction.
• Prettier check passes for all changed non-deleted files.
• ESLint passes for all changed non-deleted files.
• git diff --check passes.
• npm run typecheck:core reports 12 pre-existing diagnostics in unchanged files and none in changed files.

Tests Added Or Updated

• tests/unit/8488-capability-filter-fail-closed.test.ts
• tests/unit/combo-context-window-filter.test.ts
• tests/unit/combo-target-resolution-split.test.ts
• tests/unit/repro-8841-context-overflow-opencode.test.ts

Coverage Notes

The updated tests exercise the public handleComboChat() seam for priority, round-robin, and Auto dispatch when approximate estimates exceed every catalog hint. They also verify that a real upstream overflow still short-circuits a homogeneous same-model pool while a heterogeneous pool may try a larger-context target.

Reviewer Notes

• This is intentionally a Draft because enhancement(routing): RFC — separate approximate token estimates from hard admission decisions #10141 is design-first and requests maintainer agreement on exact/estimated/upstream-managed semantics.
• Hard tool, vision, structured-output, and requested output-token compatibility rules remain unchanged.
• Direct/single-provider capability-filter semantics remain unchanged; the shared hasEstimableContent helper only moved to an existing imported module when knownContextOverflow.ts was deleted.
• The active release base contains duplicate migration version 143 (143_api_key_cache_default_mode.sql and 143_job_registry.sql). Focused tests were run with the latter temporarily unavailable and restored immediately; this PR does not modify or repair either migration.
• Full npm run lint is base-red only in unchanged src/shared/components/modelSelectModalHelpers.ts, tests/unit/combo/image-combo.test.ts, and tests/unit/db-adapters/driverFactory.test.ts.
• The new changelog.d/fixes/10162-approximate-combo-context-advisory.md fragment passes the fragment validator in isolation. The full changelog-integrity command is base-red because two unchanged feature fragments do not start with Markdown bullets.

[diegosouzapw]

Thanks for the thorough write-up here, and for being upfront that this is a Draft tied to #10141's design questions.

I went through the diff and validated it against the current release/v3.8.50 tip. The branch is currently a good ways behind (it was opened right around the time #10152 fixed the duplicate 143_* migration you flagged in the PR description — that's already resolved upstream), but it merges cleanly with zero conflicts, and all 31 tests across the four files you touched (repro-8841-context-overflow-opencode, combo-context-window-filter, combo-target-resolution-split, 8488-capability-filter-fail-closed) pass after the merge. The dast-smoke red on this PR is also base-inherited — the same failure shows up on several unrelated PRs against the same base today, so that's not something this change introduced.

On the change itself: removing knownContextOverflow.ts and its call sites is clean — I couldn't find any remaining consumer outside the files you touched, and it's a natural extension of the advisory-metadata precedent already shipped in #8944. One thing worth spelling out explicitly once there's a decision: the Auto-strategy branch collapse means compatFilterFailOpen: false no longer produces a hard 400 when every candidate looks too small by estimate — that's really the crux of what #10141 is asking maintainers to weigh in on, so it'd help future readers to link that specific behavior change directly in the PR description.

Since you've marked this Draft on purpose, I'm not pushing for a merge here — just flagging that the code side looks solid and ready to go as soon as maintainers answer #10141's open questions and the branch gets a rebase.