fix: migrate provider catalogs to sparse route policies

[jolestar]

Summary

• migrate Tencent TokenHub, Xiaomi, DashScope, and Volcengine catalogs to shared canonical model metadata with sparse endpoint route policies
• enforce a full-catalog invariant that each canonical model's intrinsic metadata is registered exactly once
• preserve endpoint aliases, discovery behavior, and model-level-only models.catalog semantics
• add focused runtime diagnostics and Web GUI regressions for Doubao plan/coding vision capability inheritance

This is the second and final implementation PR for #2363, following #2367.

Closes #2363.

Verification

• cargo fmt --all -- --check
• cargo test model_catalog (55 passed)
• RUSTFLAGS="-D warnings" cargo check --all-targets
• npm test -- --run src/runtime/client.test.ts (26 passed)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
holon	Ready	Preview, Comment	Jul 21, 2026 11:58pm

[jolestar]

Holon Reviewer: APPROVED ✅

Summary

Second and final PR for #2363 — migrates Tencent TokenHub, Xiaomi, DashScope, and Volcengine catalogs to shared canonical model metadata with sparse endpoint route policies. Clean continuation of the pattern from #2367.

Analysis

• Canonical metadata verified: All removed per-endpoint entries correctly inherit from canonical entries (dashscope/*, volcengine/*, xiaomi/*, tencent-tokenhub/*), with route policy overrides preserving endpoint-specific caps (e.g., DashScope gateway 65_536 max_output cap for deepseek-v4, kimi-k2.x, glm-5.x).
• Capability fix confirmed: Doubao seed-2-0-pro and seed-2-0-code-preview now correctly inherit supports_reasoning and image_input from canonical volcengine entries across all endpoints (default, coding, plan). Old per-endpoint entries were missing these capabilities.
• Route definitions correct: volcengine-agent@plan/* routes resolve through volcengine-coding entries (which map to canonical volcengine), dashscope-token-plan/dashscope-coding-plan routes resolve through canonical dashscope entries.
• reasoning_effort_options guard: Xiaomi mimo endpoint guard (default | token-plan) correctly replaces the old "xiaomi" | "xiaomi-token-plan" unconditional match.
• Invariant test valuable: built_in_catalog_registers_each_canonical_model_metadata_once prevents future duplicate canonical registrations.
• Net -76 lines: Successful deduplication across 4 provider catalogs.

No blocking issues found. Ready to merge once CI passes.

[holonbot]

Review: fix: migrate provider catalogs to sparse route policies

Approve. Well-structured migration that correctly consolidates duplicate provider catalog entries into canonical-model + sparse-route-policy architecture.

Verified

• All DashScope token-plan/coding-plan route overrides match old explicit entries (context windows, max output caps, capabilities).
• Volcengine coding/plan routes correctly inherit canonical capabilities from remaining volcengine-coding entries and canonical volcengine entries.
• Intentional behavior change (Doubao models now inherit supports_reasoning=true, image_input=true from canonical) confirmed by new tests.
• Tencent TokenHub dedup and Xiaomi route migration are clean.
• reasoning_effort_options guard correctly narrows from legacy provider matching to canonical provider + endpoint guard.
• New invariant test (built_in_catalog_registers_each_canonical_model_metadata_once) and Doubao capability inheritance tests provide good regression coverage.

Minor observation

Implicit coupling: volcengine-agent@plan routes for deepseek-v4-pro, kimi-k2.7-code, glm-5.2 etc. inherit canonical metadata from volcengine-coding entries (which canonicalize to volcengine/X). A comment near those entries noting this dependency would help future maintainators avoid accidentally breaking plan routes when cleaning up coding entries.

[holonbot]

Holon Run Report

• Result: completed
• Target: fix: migrate provider catalogs to sparse route policies #2368
• Workflow run: https://github.com/holon-run/holon/actions/runs/29878784670
• Model: dashscope@token-plan/qwen3.7-max
• Requested model: dashscope@token-plan/qwen3.7-max
• Fallback active: false
• Token usage: input 390826, output 14926, total 405752
• Provider cache: read 1152995, created 69477, cacheable 1222472, hit rate 94.32%
• Model rounds: 39
• Tool calls: 47