[diegosouzapw]

OpenCode Go models without synced metadata can inherit the provider's 200,000-token default or omit output limits in imported catalog rows. Add the exact provider-scoped limits for 12 verified models: update three existing rows and register nine missing rows. Catalog enrichment, capability resolution, and context management then use the documented windows while retaining operator and synced-data precedence.

The values come from the official models.dev Go catalog pinned at 26d4dc7, resolving its base_model inheritance. Request formats follow the official Go endpoint table. GPT-6 Luna retains its narrower 922,000-token input ceiling; Space Bunny uses a 524,288-token input ceiling. The provider URL, default model order, executor, global model specs, and existing effort aliases are unchanged.

This addresses the verified subset of #16579. deepseek-flash, minimax-m2.5, omen-alpha, and the Qwen high/max alias IDs still need Go-specific limit evidence. The inspected sources did not establish Go-specific limits for those IDs, and the executor forwards the Qwen aliases as distinct IDs. Their existing rows remain available; limits from another provider are not assumed to apply. The issue should remain open for that evidence.

Validation uses frozen release base 6a1437e6e262db43b50d4bc8cb7b1d658b2e6308 and the audited exact lockfile dependency tree (2,524 installed packages checked, zero version mismatches). This does not claim validation of a later merge result.

• TDD RED on the unchanged base: 12 expected failures and five passing controls (exit 1).
• GREEN after implementation: 17/17 tests pass (exit 0), covering public catalog enrichment, capability/context resolution, separate input ceilings, operator/synced precedence, request formats, unknown limits, and environment overrides.
• Public capability comparison on unchanged base versus fix: all non-limit capability fields remain identical for all nine newly registered rows, including reasoning, tools, vision, audio/video, and supported efforts.
• Prettier, changed-file ESLint, file-size, changelog integrity and test discovery pass.
• Core TypeScript, dashboard typecheck, cyclomatic complexity and cognitive complexity all pass on commit 114e112f1c310c5884f5bb66c24d797aefa27499. Dashboard findings remain within the frozen baseline.
• Final regression selection: 256/256 tests pass across 26 native test files, with zero failures, cancellations or skips.

Refs #16579

⚠️ base-red inherited: #15306 — the release-green tracker remains open for release-wide docs, full-suite, packaging/build and agent-skills failures. Its signatures must be compared on the exact base before attributing a failure. No failing local gate is waived here: the scoped checks above passed on the stated commit, including dashboard typecheck. This does not establish full release-green status.

[github-actions]

CI Coverage Report

• Coverage job: skipped
• PR test policy: success

Coverage artifact was not available for this run.