stage-356: small 2-PR follow-up batch — #2234 aux-model routing + #2265 mixed-case provider canonicalization (closes #2245)

[nesquena-hermes]

stage-356 — small 2-PR follow-up batch (v0.51.63)

Fourth full PR sweep of the day. Two non-hold non-draft PRs cleared fit-screen, both → MERGE QUEUE.

+268 / -7 across 5 files. Two PRs from two contributors.

Composition

PR	Author	LOC	Surface	Closes
#2234	Jordan-SkyLF	+11/-4	api/routes.py + tests	(post-v0.51.62 rebase, 3rd follow-up commit)
#2265	Michaelyklam	+257/-3	api/config.py + new test file	#2245

PR Detail

#2234 — Use documented aux model for update summaries

The THIRD and final follow-up commit (97b68f0) from PR #2234, now rebased onto current master. The category-refinement portion shipped in v0.51.61 (stage-354) and the rebase + scrollable-summary portion shipped in v0.51.62 (stage-355).

Concern: the original implementation looked up an auxiliary.update_summary config key — but that's a WebUI-only magic key, not a documented Hermes Agent auxiliary task. Users would have to guess the config key existed.

Fix: route update summary generation through the documented auxiliary.compression text-model slot. Main-model fallback is preserved if auxiliary resolution or generation fails. Adds a route comment explaining the mapping rationale (summary generation conceptually IS summarization, which is what compression provides).

#2265 — Fix configured provider models after key canonicalization (closes #2245)

Real bug: when users define a custom provider with mixed-case (CLIPpoxy) or snake_case (snake_case_provider) key in config.yaml, canonicalization lower-cases the key. Subsequent model-allowlist lookup uses the canonicalized key but config.yaml still stores the raw key, so the lookup returns empty and the model picker dropdown shows no models for that provider.

Fix: api/config.py maps canonical provider IDs back to their raw config.yaml provider keys before loading provider_cfg. Original config keys are preserved for provider settings rendering. Built-in providers (openai, anthropic, etc.) and fallback behavior unchanged.

242-line regression test covering mixed-case + snake_case + built-in unchanged + _PROVIDER_MODELS-only fallback.

Held PRs re-evaluated this sweep

Per the rapid-re-sweep rule: pulled latest comments on each held PR.

• feat: add opt-in streaming text fade #2099 (opt-in streaming text fade) — contributor says "ready for review" since 2026-05-13. Defer — explicit human-review request + 811 LOC + UX gate needed.
• feat(providers): show pooled Codex quota status #2165 (pooled Codex quota status) — starship-s addressed all 4 prior review concerns in commit ab663cd. Defer — 1823 LOC, too large for sweep, deserves dedicated review session.
• feat: add custom WebUI branding images #2208 (custom WebUI branding) — Aron weighed in: "plugin type territory vs core?" — soft defer until maintainer call.
• feat: add per-chat RTL toggle button with keyboard shortcut #1721, Add external memory approval queue UI #1975, feat: Add toggleable session lineage indicators setting #2072, feat: show ambient provider quota indicator #2082, feat: auto-import CLI sessions for seamless continuation #2174, Fix/chat start preserve sidecar #2183, feat(webui): reconcile external session updates #2194, Add OpenRouter cost history backend #2195, feat(deepseek): add V4 Pro/Flash [1m] variants to model dropdown #1418, [codex] Fix workspace-prefixed user turn leakage #2145, [codex] Track update checks by release tags #2146 — no notable change since last sweep; holds remain valid.

Verification

• Targeted pytest (6 PR surfaces): 96/96 pass in 4.9s
• Full pytest: exit 0
• run-browser-tests.sh: 20/20 QA + 11/11 API checks PASSED in 104s
• Live UI smoke on 8789: zero JS errors
• Phase 5 strict merge-marker check: zero markers
• Python syntax: clean across all 5 modified files
• CI on each constituent PR (latest SHA): 3/3 SUCCESS

Opus Advisor Review

SHIP both. No SHOULD-FIX.

Opus directly verified:

• fix: use documented aux model for update summaries #2234 aux-model resolution + fallback (api/routes.py:5467-5485) — try/except wraps both get_text_auxiliary_client("compression", ...) AND the aux_client.chat.completions.create(...) call. A 401 or timeout from configured aux falls through cleanly to main-model AIAgent block at L5487 ✓
• fix: use documented aux model for update summaries #2234 cache key consistency — api/updates.py:347 keys on (name, behind, current_sha, latest_sha, compare_url) — model-agnostic, no cross-model staleness ✓
• Fix configured provider models after key canonicalization #2265 reverse map iteration (api/config.py:2879-2884) — walks every key in cfg.get("providers", {}), canonicalizes via _canonicalise_provider_id, and setdefaults canonical → raw_key. CLIPpoxy → clippoxy maps back to "CLIPpoxy" correctly ✓
• Fix configured provider models after key canonicalization #2265 built-in provider regression risk — pid in _PROVIDER_MODELS or pid in _canonical_to_raw_provider_key condition; for openai/anthropic not in cfg, first clause matches and behavior is identical to pre-fix ✓

Minor observations (not blockers, Opus noted)

• Fix configured provider models after key canonicalization #2265: If user has both CLIPpoxy and clippoxy in providers (canonical collision), setdefault makes only the first win. Pre-fix that case silently lost ALL models for the second. Worth a logger.debug on collision in a future cleanup.
• Fix configured provider models after key canonicalization #2265: tests/test_issue2245_mixed_case_provider_models.py ends with no trailing newline. Cosmetic.

Closes / refs

• Closes #2245 (via Fix configured provider models after key canonicalization #2265 mixed-case canonical key fix)

Stats

5 files changed, 268 insertions(+), 7 deletions(-)