The scan terminal service had accumulated cache, payload, filtering, AI prompt, AI merge, METAR gate, ranking, and city-row construction details in one file. This splits those stable responsibilities into focused modules while preserving the endpoint payload shape and existing behavior.
Constraint: User-visible behavior and release version must remain unchanged for this internal refactor
Rejected: Rewrite the terminal scan flow around a new abstraction | too risky while production behavior is being stabilized
Confidence: high
Scope-risk: moderate
Directive: Keep scan_terminal_service.py as orchestration; add detailed rule changes to the focused modules instead of re-growing the service file
Tested: py_compile for extracted modules; ruff check .; pytest tests/test_scan_terminal_modules.py tests/test_web_observability.py; full pytest; npm run test:business; npm run build; git diff --cached --check
The scan terminal had grown into overlapping CSS, request-state, AI-provider, and city-card data responsibilities. This refactor separates those boundaries without changing product behavior: CSS modules are split by surface, city AI prompt/provider/fallback logic is isolated, and scan terminal request state now has reusable RemoteData adapters plus business-state tests.
Constraint: Preserve existing global scan-terminal class names and API responses during the refactor
Constraint: No new dependencies; keep this as a file-boundary cleanup
Rejected: Introduce React Query now | higher migration risk than the requested lightweight query-client path
Rejected: Rewrite AI stream behavior | progressive/fallback states are product-sensitive and were only adapter-split
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep AI stream state changes covered by business snapshots before changing fallback/cache wording
Tested: npm run test:business; npx tsc --noEmit; npm run build; python pytest -q; ruff check; py_compile targeted city AI modules
Not-tested: Live DeepSeek provider network replay and browser visual QA
The AI helper extraction left a stale private helper import in the scan terminal service. Removing it keeps CI aligned with the current call graph without changing runtime behavior.
Constraint: CI runs ruff F401 as a blocking check.
Confidence: high
Scope-risk: narrow
Reversibility: clean
Tested: .\.codex-tmp\pydeps\bin\ruff.exe check .
Not-tested: Full backend test suite; change is import-only.
The scan terminal service had grown into a 3.5k-line file that mixed endpoint orchestration, AI provider calls, fallback copy, JSON repair, and deterministic evidence guards. This extracts the city-AI helper layer into a focused module while preserving the old private names through imports for existing tests and callers.
Constraint: Keep behavior unchanged after the previous evidence-guard fixes and avoid a broad service rewrite.
Rejected: Split every scan-terminal concern at once | too much regression risk for this maintenance pass.
Confidence: high
Scope-risk: narrow
Tested: pytest tests/test_web_observability.py -q
Tested: npm run build
DeepSeek can return a polished city forecast that conflicts with already-computed stale-observation, observed-break, or peak-window evidence. The completion path now carries the deterministic fallback guard state and overwrites only the critical fields when provider text or numbers contradict those local facts.
Constraint: Airport-read latency optimization keeps provider output narrow, so backend completion remains the authority for final highs and evidence conflicts.
Rejected: Trust provider final wording when present | it can reintroduce stale METAR anchors or miss observed high breaks.
Confidence: high
Scope-risk: narrow
Tested: pytest tests/test_web_observability.py -q
Tested: npm run build
City-card fallback reads now stop using stale METAR or official observations as strong live anchors. A stale observation no longer forces high/low revisions, and both backend and browser AI cache keys include the observation fingerprint so updated report times, receipt times, temperatures, or stale status invalidate old AI text.
Constraint: Cached city AI reads must not survive a material observation update
Rejected: Let stale METAR trigger observed-break revisions | stale reports can be older than the active temperature path
Confidence: high
Scope-risk: moderate
Tested: pytest tests/test_web_observability.py -q
Tested: npm run build
Fallback city-card reads now distinguish three cases that previously collapsed into the generic fast-evidence copy: observed highs above the model path, observed highs still lagging after the peak window, and low observations before the peak window that should wait for confirmation rather than down-revise immediately. The same pass removes three unused private helpers from the scan terminal service.
Constraint: Fallback output must be useful before the full AI airport-bulletin read returns
Rejected: Treat any low latest METAR as a down-revision | early-day observations can be below the forecast before the peak window
Rejected: Keep unused helper wrappers | they were unreferenced and added noise to an already large module
Confidence: high
Scope-risk: moderate
Tested: pytest tests/test_web_observability.py -q
Tested: npm run build
Fast evidence mode should not say DEB and models support the center when the latest METAR has already exceeded that center or the model upper edge. The fallback now treats the live observation as a lower bound for the daily high and explains the upward revision pressure.
Constraint: Fallback output must remain useful before the full AI bulletin read returns
Rejected: Keep the original DEB center until AI completes | it can be lower than an already-observed temperature
Confidence: high
Scope-risk: narrow
Tested: pytest tests/test_web_observability.py::test_city_ai_fallback_revises_up_when_latest_metar_breaks_above_models tests/test_web_observability.py::test_city_ai_fallback_reasoning_identifies_fast_evidence_mode tests/test_web_observability.py::test_city_ai_stream_request_only_asks_provider_for_observation_read -q
Tested: npm run build
The city card only needs the provider to interpret the latest METAR or official observation. Deterministic fields such as the high-temperature center, model cluster note and fallback risks are already available server-side, so the stream request now asks DeepSeek for only the observation read and concise reasoning.
Constraint: City cards still need a complete payload for both Chinese and English UI modes
Rejected: Keep generating the full decision schema in the stream | too much model output for every card
Rejected: Retry failed streams by default | it can double latency and the fallback can use partial streamed text
Confidence: high
Scope-risk: moderate
Tested: pytest tests/test_web_observability.py::test_city_ai_stream_request_only_asks_provider_for_observation_read tests/test_web_observability.py::test_city_ai_fallback_reasoning_identifies_fast_evidence_mode tests/test_web_observability.py::test_city_ai_partial_json_trims_dangling_taf_clause tests/test_web_observability.py::test_city_ai_schema_completion_trims_dangling_taf_clause -q
Tested: npm run build
The city decision fallback path is generated when the full DeepSeek city-airport read has not completed, so the reasoning copy now labels the state as fast evidence mode instead of saying the AI read is normal.
Constraint: Fallback output may use only DEB, model cluster, and latest observation evidence
Rejected: Keep 'AI read normal' wording | it implies a completed AI interpretation when the fallback path is active
Confidence: high
Scope-risk: narrow
Tested: pytest tests/test_web_observability.py::test_city_ai_fallback_reasoning_identifies_fast_evidence_mode tests/test_web_observability.py::test_city_ai_partial_json_trims_dangling_taf_clause tests/test_web_observability.py::test_city_ai_schema_completion_trims_dangling_taf_clause -q
City AI can return a partially streamed JSON string when the provider truncates output. The fallback previously kept an unfinished clause such as '但TAF显示', which made the forecast explanation look broken even though earlier evidence was usable.
Constraint: Provider JSON can be truncated after useful fields have already streamed
Rejected: Drop all partial AI text | would lose valid METAR interpretation already returned before truncation
Confidence: high
Scope-risk: narrow
Tested: pytest city AI truncation regression tests
Tested: npm run build
Not-tested: Live DeepSeek provider response
2569718930@qq.com