choiyounggi · choiyounggi · Sep 27, 2026 · Sep 28, 2026
@@ -12,10 +12,10 @@ follow the cross-pointers in their index or take the next matching seeded domain
 | Domain | Status | Route here when |
 |--------|--------|-----------------|
 | [databases](wiki/databases/index.md) | **seeded** | Choosing a datastore/database type for a workload (relational vs document vs vector vs graph), designing schemas/tables/keys, choosing or evaluating indexes, writing or optimizing queries, choosing transaction/isolation behavior, multi-row rewrites (reorder, bulk status) on a shared resource, surveying live data to derive a rule, verifying additive migrations |
-| [backend](wiki/backend/index.md) | **seeded** | Server-side application code — language-agnostic (`common/`: API contracts, call-site enumeration before a contract change, idempotency, JWT, timeouts/retries, caching, jobs, transactions in app code, shared state/pools, errors, consuming LLM APIs (completion validation, context budgeting), authoring agent-facing artifacts (binding instruction text, agent tool-surface granularity/parity), MAPE-aligned point-prediction calibration, benchmark-relative signal rating, consuming external-API responses, externally-owned defaults, object-storage references, sync-vs-async integration choice, WebSocket/SSE connection lifecycle) plus stack subtrees: `java/` (JPA, Spring proxies, JVM threads/memory), `node/` (event loop, promises, runtime validation, shutdown), `python/` (GIL/asyncio, pydantic, WSGI/ASGI workers, language traps, packaging data files with `importlib.resources`) |
-| [frontend](wiki/frontend/index.md) | **seeded** | Web UI code: state placement, rendering performance, in-UI data fetching (races, infinite scroll), auth token handling, forms, XSS-safe output, accessibility, any new or changed user action (this wiki's development standard adds a WebMCP tool surface per action) |
-| [infrastructure](wiki/infrastructure/index.md) | **seeded** | CI/CD pipelines, secrets in build/deploy, container image builds, rollout/rollback strategy, observability (logs/metrics/alerting), per-environment/path-valued config, multi-agent orchestration (worker liveness signals, shared run state, tmux pane delivery, completion gates, worktree-isolated workers, autonomous ask-vs-rule decisions, session context/token budgeting, a pre-built code knowledge graph as a freshness-gated orientation layer for planning, the merged-tree gate for parallel branches, the verify command written into a worker brief, a lock owner id inherited from the parent that spawned the session) |
-| [testing](wiki/testing/index.md) | **seeded** | Writing or structuring automated tests: level choice, test-before-code ordering, a UI action that is also a registered agent tool, cases/assertions, cross-layer effect scoping, test data, mock decisions, flaky tests, test-infrastructure containers (Testcontainers) failing on the dev host, testing a SwiftPM executable target (release-process quality → qa) |
+| [backend](wiki/backend/index.md) | **seeded** | Server-side application code — language-agnostic (`common/`: API contracts, call-site enumeration before a contract change, idempotency, JWT, timeouts/retries, caching, jobs, transactions in app code, shared state/pools, errors, consuming LLM APIs (completion validation, context budgeting, first-call load latency of a self-hosted model server), authoring agent-facing artifacts (binding instruction text, agent tool-surface granularity/parity), MAPE-aligned point-prediction calibration, benchmark-relative signal rating, consuming external-API responses, externally-owned defaults, object-storage references, sync-vs-async integration choice, WebSocket/SSE connection lifecycle) plus stack subtrees: `java/` (JPA, Spring proxies, JVM threads/memory, Kotlin: kotlinx.serialization enum coercion, implicit-receiver shadowing in scope functions), `node/` (event loop, promises, runtime validation, shutdown), `python/` (GIL/asyncio, pydantic, WSGI/ASGI workers, language traps, packaging data files with `importlib.resources`) |
+| [frontend](wiki/frontend/index.md) | **seeded** | Web UI code: state placement, rendering performance, in-UI data fetching (races, infinite scroll), auth token handling, forms, XSS-safe output, accessibility, concurrent optimistic updates against one server-owned object (web or mobile view models), scroll-scrubbed animations under a reduced-motion preference, any new or changed user action (this wiki's development standard adds a WebMCP tool surface per action) |
+| [infrastructure](wiki/infrastructure/index.md) | **seeded** | CI/CD pipelines, secrets in build/deploy, container image builds, rollout/rollback strategy, observability (logs/metrics/alerting), per-environment/path-valued config, config values a build step writes into generated source, multi-agent orchestration (worker liveness signals, shared run state, tmux pane delivery, completion gates, worktree-isolated workers, autonomous ask-vs-rule decisions, session context/token budgeting, a pre-built code knowledge graph as a freshness-gated orientation layer for planning, the merged-tree gate for parallel branches, the verify command written into a worker brief, a lock owner id inherited from the parent that spawned the session) |
+| [testing](wiki/testing/index.md) | **seeded** | Writing or structuring automated tests: level choice, test-before-code ordering, a UI action that is also a registered agent tool, cases/assertions, cross-layer effect scoping, test data, mock decisions (including a fake `IntersectionObserver` under a viewport-animation library), contract tests for a deprecated-alias table, flaky tests, test-infrastructure containers (Testcontainers) failing on the dev host, testing a SwiftPM executable target (release-process quality → qa) |
 | [qa](wiki/qa/index.md) | **seeded** | Release-quality process: release gates, regression scoping, bug reports, severity/priority triage, evidence for completion claims, the agent-tool parity gate for a web UI release, acting on code-review feedback, adversarial review of high-risk diffs, exploratory testing (guarded-path coverage, override matrices), scope-purity gates, sourcing deliverable documents from generated artifacts, verifying the quantitative claims in a document before publishing it, a documented claim about a third-party tool's side effects, an obligation row in a tier/policy table that another contract also pins, rationale prose left behind by a config-value change, automated verification of document deliverables (spec/RFC gates), an aging detector for model-coupled agent guidance, capturing an app's own screen content without Screen Recording permission (writing automated test code → testing) |
 | [debugging](wiki/debugging/index.md) | **seeded** | Diagnosing a failure — finding what is wrong and why: reproducing, bisection, hypothesis testing, traces/logs, intermittent failures (fixing the diagnosed fault → its owning domain) |
 | [security](wiki/security/index.md) | **seeded** | Trust-boundary decisions: input validation, session-vs-token auth choice, per-resource authorization (IDOR), secrets hygiene (including ciphertext orphaned by a regenerated encryption key), dependency trust, PII handling, in-session agent tool exposure (prompt-injection blast radius), the author identity a commit publishes to a public repository, host-compromise triage / incident response (verifying assumed security agents, identifying masquerading processes) (XSS rendering → frontend; CI secrets → infrastructure; JWT implementation → backend/frontend auth) |

@@ -208,3 +208,21 @@ Append-only. Format: `## [YYYY-MM-DD] <ingest|revise|lint|gap|contradiction|drif
 ## [2026-09-28] ingest | testing-strategy-agent-tool-shared-handler-tests — a UI action that is also a registered tool is tested once at the shared function plus two entry-point tests per tool (Registration incl. AbortSignal teardown, Wiring via spy) against a `document.modelContext` stub; a bug fix that changes the handler's contract changes the `inputSchema` assertion in the same commit; one DevTools Run-tool pass per release.
 
 ## [2026-09-28] revise | WebMCP adopted as the development standard (owner decision 2026-09-28): frontend/agent-interfaces/agent-facing-tool-surfaces trigger widened to any new or changed user action in a web UI + bug-fix and exclusion-list edge cases; AGENTS.md routing step 7 gains a web-UI-action row → frontend agent-interfaces then qa parity gate; INDEX.md frontend/qa/testing route lines and the frontend/qa/testing domain indexes updated; related links added both ways (release-gates, cross-layer-effect-tests, in-session-tool-exposure). The standard keeps the human UI primary and the tool layer additive (CG draft; Chrome origin trial + ChatGPT desktop runtimes).
+
+## [2026-09-28] ingest | testing-mocking-fake-intersection-observer-for-viewport-animations — under motion/framer-motion `whileInView`, observers are cached per (root, serialised options) in a module WeakMap so constructor probes read 0 from the second test; the per-element callback is looked up by `entry.target`, so a fake entry without it makes the animation silently never fire while first-paint styles keep the suite green; count `observe(element)`, include `target`, add an end-state positive control and a hardcoded-defaults mutation (framer-motion 13.2.0 source; field evidence cover-letter 2026-09-27)
+
+## [2026-09-28] ingest | testing-quality-alias-table-contract-tests — a deprecated-alias layer is tested row-for-row from the design document's mapping with a size assertion equal to the legacy-name count and one retargeting mutation proven red; a spot-check of 4/18 let un-asserted aliases forward wrong (PIT, Google Testing Blog, JUnit 5 parameterized; field evidence linkly-calendar Android 2026-09-27)
+
+## [2026-09-28] revise | infrastructure-agent-orchestration-checkable-claims-in-an-adopted-plan — +trigger, +Do row, +edge row, +field evidence: an adopt-only plan that freezes test bodies and forbids new failures is probed by one production slice + whole-suite run + failing-set delta against baseline, then reverted and reported (seagrass t174, failures 1→2 inside a frozen test class)
+
+## [2026-09-28] ingest | frontend-state-concurrent-optimistic-updates — concurrent per-field PATCH optimistic updates against one server-owned object: confirmed snapshot + ordered pending patches rendered on top, FIFO-serialised requests, success sets confirmed=response, failure drops only its own patch; invalidate a query cache only when `isMutating() === 1`; overlap test with response gates (TanStack optimistic-updates guide, TkDodo concurrent-updates; field evidence linkly-calendar Android SettingsViewModel 2026-09-27, 4 tests red→green, 131/0)
+
+## [2026-09-28] ingest | infrastructure-config-config-values-emitted-as-source-text — a build step writing user/env input into generated source (AGP `buildConfigField` → `BuildConfig.java`, "must have valid Java content") must escape `\` then `"` in one helper and add a compile probe with hostile characters to the verification step, because javac fails before any runtime validation runs (AGP DSL reference, gradle-tips; field evidence javac 17 `unclosed string literal` 2026-09-27)
+
+## [2026-09-28] ingest | backend-java-kotlin-coerced-enum-defaults-in-kotlinx-serialization — with `coerceInputValues = true`, an unknown enum value on a property with a default is silently replaced by the default; keep must-reject enum properties required (no default), pin with a SerializationException decode test, or make forward-compat explicit with an `UNKNOWN` member (kotlinx.serialization docs/json.md; field evidence linkly-calendar Android TripPlace.category 2026-09-27)
+
+## [2026-09-28] ingest | backend-java-kotlin-implicit-receiver-shadowing-in-scope-functions — inside `apply`/`run`/`with` an unqualified call resolves to the implicit receiver's member before a same-named top-level function (Kotlin spec overload resolution), so a fixture helper `trip(id)` silently calls the fake's `trip(id)`; name helpers distinctly or use `also`/`let` (field evidence linkly-calendar Android TripViewModelTest 2026-09-27, 3 failures → 13 pass)
+
+## [2026-09-28] ingest | backend-common-llm-self-hosted-model-load-latency — the first call to a self-hosted model server after idle pays the disk load (Ollama keeps a model 5 m by default; `OLLAMA_KEEP_ALIVE` / per-request `keep_alive`, request wins); size the client timeout above load+generation, warm up before measuring, set keep_alive longer than the caller's idle interval (Ollama FAQ/API docs; field evidence 8 GB Q8 GGUF judge, 10 s SDK timeout → 31 s recorded failure, warm-up 1588 ms then p50 786 ms, 2026-09-27)
+
+## [2026-09-28] ingest | frontend-design-scrubbed-scroll-animations-under-reduced-motion — a `scrub` ScrollTrigger sets progress from scroll position, so a global `timeScale()` reduced-motion switch never reaches it; read the preference where the effect is created (`gsap.matchMedia` `(prefers-reduced-motion: reduce)`), skip the trigger and set the final state; enumerate `scrub`/progress `onUpdate` sites at task review (GSAP ScrollTrigger, timeScale, matchMedia docs; field evidence cover-letter DurationMeter 2026-09-27)
@@ -0,0 +1,71 @@
+---
+id: backend-common-llm-self-hosted-model-load-latency
+domain: backend
+category: llm
+applies_to: [general]
+confidence: verified
+sources:
+  - https://github.com/ollama/ollama/blob/main/docs/faq.mdx
+  - https://github.com/ollama/ollama/blob/main/docs/api.md
+last_verified: 2026-09-28
+related: [backend-common-llm-context-window-budget, backend-common-reliability-timeouts-and-retries, backend-common-reliability-client-side-rate-limiting]
+---
+
+# First Request to a Self-Hosted Model Server After Idle
+
+## When this applies
+
+A client calls a locally hosted model server (Ollama or a server with the same
+keep-alive model) that loads weights from disk on demand and unloads them after
+an idle window, and the caller is intermittent — a hook, a cron job, an
+evaluation script, a judge model — or the caller's SDK has a short default
+timeout (10 s is common). Also when the first call of a run fails with a
+timeout while later calls succeed in under a second.
+
+## Do this
+
+1. **Measure the load once and size the client timeout above load plus
+   generation.** Time a first call after the model is unloaded (`ollama ps`
+   shows what is loaded and until when); a multi-gigabyte GGUF takes tens of
+   seconds from disk. Set the timeout to 60 s or more for such models — a
+   timeout at 10 s returns an error while the load continues.
+2. **Send a warm-up request before anything you measure or batch**, and report
+   its latency separately from the steady-state numbers.
+3. **Set `keep_alive` longer than the caller's idle interval.** Ollama keeps a
+   model loaded "for 5 minutes before being unloaded" by default; the
+   `OLLAMA_KEEP_ALIVE` environment variable changes it for all models, and the
+   `keep_alive` parameter on `/api/generate` and `/api/chat` overrides it per
+   request (duration string, seconds, a negative number to keep loaded, `0` to
+   unload after the response).
+4. **Treat a timeout on the first call as a load in progress, not a dead
+   server:** check the server's loaded-model list before retrying — a retry
+   inside the same short timeout pays the load again and records a second
+   failure ([backend-common-reliability-timeouts-and-retries]).
+
+| Caller pattern | keep_alive |
+|----------------|------------|
+| Interactive session, calls seconds apart | Default (5 m) |
+| Hook or job firing every N minutes | Longer than N (`OLLAMA_KEEP_ALIVE=30m` or per-request `"30m"`) |
+| Dedicated judge/eval box, memory to spare | Negative value (`-1`) — stays loaded |
+| One-shot batch, then free the memory | `0` on the last request |
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| The SDK retries on timeout | The first result is recorded as a failure after `timeout × attempts` while the server was loading the whole time; disable retry for the warm-up call or raise the timeout first |
+| Several models share one server | Each cold model pays its own load; warm up each one you will call |
+| The env var and the request parameter disagree | The request parameter wins — the documented precedence |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Read the first-call timeout as "the server is broken" | Check `ollama ps` and re-send after the load | The model was loading; the error is the client's clock, not the server |
+| Publish p50 latency from a run that includes the cold call | Warm up first and report the warm-up separately | One 30 s load skews every summary statistic |
+
+## Sources
+
+- https://github.com/ollama/ollama/blob/main/docs/faq.mdx — "By default models are kept in memory for 5 minutes before being unloaded"; `keep_alive` accepts "a duration string (such as "10m" or "24h")", "a number in seconds (such as 3600)", "any negative number which will keep the model loaded in memory (e.g. -1 or "-1m")", "'0' which will unload the model immediately after generating a response"; "The `keep_alive` API parameter with the `/api/generate` and `/api/chat` API endpoints will override the `OLLAMA_KEEP_ALIVE` setting"
+- https://github.com/ollama/ollama/blob/main/docs/api.md — `keep_alive`: "controls how long the model will stay loaded into memory following the request (default: `5m`)"
+- Field evidence 2026-09-27 (a local evaluation harness against an Ollama-compatible server hosting an 8 GB Q8 GGUF judge model; client SDK default timeout 10 s with retry): the first judgment was recorded as `31334 ms ERROR … Request timed out (timeout=10.0)`; with the timeout raised to 120 s and one warm-up call (`warmup 1588 ms`) the run scored 50/50 with p50 786 ms, and the server's process list showed the model unloading "4 minutes from now"