5 .hint spans (ADMIN TOKEN / PACK PATH / PROVIDER ID / SMOKE TEST MODEL /
COMMIT MESSAGE) were being visually overlapped by their sibling <input>
boxes because:
- .hint was inline, padding 10px 12px with background+border
- <input> was inline-block
- label was display:block, font-size 12px
- the inline .hint box was being squeezed into the same flow line as the
input, so its last 8-10px were covered
Verified via getBoundingClientRect + elementFromPoint probe before fix:
hint.top < input.bottom (5 fields) = hint overlapped by input
Fix: force .hint to display:block with margin-top:6px so it always sits
below its input as a separate block; also force label > input/select/textarea
to display:block width:100% so the form field always stretches and never
inlines with the field label text.
Verified after fix: 0 hints overlapped on providers.html (the worst case
with 8 .hint spans).
CSS-only change; no HTML edits. Affects all 8 portal pages. Smoke +
frontend assets tests still PASS.
7-day window 220 commits across 3 active repos (sub2api-cn-relay-manager 115,
user-system 72, llm-intelligence 33). 4 cross-project patterns extracted into
hermes-global skills:
1. numbered-task-board — M-/H-/B- + T-Data- prefix systems
2. six-stage-pipeline-verify — schema→lib→importer→fixture→real-dry→real-upsert
3. docs-only-commit — code/docs commits never mixed
4. test-coverage-iteration-workflow — per-module 30-100 test funcs
Each skill cites the source commits / files for verifiability.
This sub2api-cn-relay-manager view focuses on what the project contributed.
Docs-only commit (skill #3 in action). No code changes.
§9 of codegraph-integration note now documents the 2026-06-03 15:20 upgrade:
- enrich defaults to callWithFallback() chain (current model → fallback_providers)
- skips codex_responses transport providers automatically
- resolves __default._providerModel to endpoint-served name (not user alias)
- strips <think>...</think> blocks before JSON.parse
- explicit --provider still supported (skip chain, useful for fast mode)
Verified on 5 git repos (contract management / gongju / llm-intelligence /
user-system / supply-intelligence) — 13/13 OK, 0 fallback rotation needed,
all from minimax-m3/MiniMax-M3 (current agent model).
Docs-only commit; sibling subagent's parallel code changes are NOT included.
All three admin pages had two parallel inline <script> blocks (a modern S1
that used adminRuntime + a legacy S2 that was self-contained). Both had
a nested <script> text inside S1 that the browser tolerated only because
the second script re-ran any state-affecting calls. Merge into a single
inline script per page; fix the nested <script> comment.
- providers.html: 100371 -> 62761 chars (-37610, -37%)
- accounts.html: 54878 -> 33098 chars (-21780, -40%)
- batch-import: 43861 -> 26570 chars (-17291, -39%)
Also rename draftProviderIDInput -> providerIDInput in providers.html
(the old draft-provider-id input was removed during the earlier workflow
merge, leaving the script with a null addEventListener on draft id).
All scripts pass node --check. Both test_tksea_portal_assets.sh and
verify_frontend_smoke.sh PASS.
Problem: provider manifest form had all-empty fields with cryptic
placeholders, users had to know what model IDs to type.
Fix on /portal/admin/providers.html (Provider Manifest 草稿):
- DISPLAY NAME: datalist of common vendors (OpenAI / DeepSeek /
硅基流动 / Moonshot / 智谱 / Anthropic / 零一万物 / MiniMax / Qwen / Baichuan / 混元)
- PLATFORM: datalist of common platforms (openai / openai-compatible /
deepseek / anthropic / gemini / zhipu / moonshot / minimax / qwen / ...)
- SMOKE TEST MODEL: datalist of common smoke models + auto-fills with
first model from MODELS field if user leaves it empty
- BASE URL PLACEHOLDER: datalist of common base URLs (12 presets)
- MODELS: chip-row of 11 common models (gpt-5.4, gpt-5.4-mini,
deepseek-chat, MiniMax-M2.7-highspeed, kimi-k2.6, glm-4.6,
claude-sonnet-4-5, gemini-2.5-pro, qwen3-coder-plus, gpt-4o, o4-mini)
+ clear button. Click chip → append to MODELS field (dedup).
- KEYS textarea: 6 rows + example placeholder (sk-example-1/2/3)
Fix on /portal/admin-batch-import.html (发起导入):
- HOST ID: datalist of common host_ids + hint about loading pack first
- ENTRIES textarea: 6 rows + multi-line hint explaining
base_url|api_key|model1,model2 format, optional model, batch import
JS change: syncDraftHelperState() in providers.html now auto-fills
smoke_test_model with first model if user hasn't filled it yet.
Also fixed: 2 duplicate copies of syncDraftHelperState (from
earlier batch script restoration) — both now have the new logic.
Verification:
- bash scripts/test/test_tksea_portal_assets.sh → PASS
- bash scripts/test/verify_frontend_smoke.sh → PASS
- browser_console click test: gpt-5.4 + deepseek-chat + kimi-k2.6 chips
→ models='gpt-5.4,deepseek-chat,kimi-k2.6' + smoke='gpt-5.4' auto-fill ✓
- screenshot: /tmp/portal-screenshots/admin-providers-v5.png
Add a dedicated acceptance script for providers.html, cover it in the local real-host script regression suite, and document the current frontend review baseline, closure audit, providers action matrix, and remediation task board.
This keeps the frontend acceptance boundary explicit: providers.html now has a repeatable verification entry point for its page-level actions, while non-UI provider operations remain documented as backend-only capabilities.
When subscription access is requested with an explicit access_api_key, assign the subscription to the real target user, bind that user's API key to the subscription group, and probe readiness with the same key instead of falling back to a managed synthetic user.
Update the runtime/reconcile flows, adapter tests, and source-of-truth docs so subscription_ready now reflects user-visible host access rather than managed-key-only closure success.
