refactor(src): split the remaining five oversized modules behind facades

devlog/_plan/260915_workflow_budget_window/000_unit.md

@@ -0,0 +1,97 @@
+# 260915 — the root workflow budget outlives the task it was sized for
+
+## What happened
+
+A Codex session spent several hours dispatching subagents. Every dispatch failed,
+across three unrelated providers, with a 429 that reads as a provider rate limit.
+The obvious readings were all wrong: not the model, not the account, not the
+upstream. The refusal came from this proxy.
+
+The reproduction is one line. With the same request body, a probe carrying the
+long-running session id in `x-codex-parent-thread-id` was refused, and a probe
+carrying a freshly invented root id was served. Restarting the proxy served both.
+That is the whole diagnosis: the ceiling is per root, held in process memory, and
+the session had reached it.
+
+## Why the ceiling fired
+
+`DEFAULT_WORKFLOW_BUDGET_POLICY` (`src/lib/workflow-budget.ts`) caps a root at 256
+physical sends and 64 distinct children. `workflowSendCeilingReached` compares
+`state.sends >= policy.maxPhysicalSends`, and `state.sends` is **cumulative for the
+life of the process**. The root id is `x-codex-parent-thread-id`, which for Codex is
+the session. So the cap is not a fan-out guard on a long session; it is an expiry.
+
+The comment that justifies it says a per-request cap "cannot bound a fan-out that
+sends once per child seven hundred times". That is a **burst** concern, and a burst
+is bounded by a rate. A lifetime total cannot tell seven hundred sends in a minute
+from two hundred and fifty-six sends spread over four hours, and it refuses both.
+The second one is ordinary work.
+
+Two things made it expensive to diagnose rather than merely annoying. The refusal
+is a 429 that an operator reads as an upstream rate limit, so the first hours went
+to providers and accounts. And there is no way out except restarting the proxy:
+`resetWorkflowBudgetsForTest` exists, the name says who it is for, and
+`workflowBudgetSnapshot` is never exposed, so the state that decided the refusal is
+invisible from outside the process.
+
+## The rule
+
+> A root budget bounds a **rate**, and says so. A ceiling that fires is a local
+> decision an operator can see, name and clear without restarting the proxy.
+
+## Roadmap
+
+| Doc | Work phase | Outcome |
+| --- | --- | --- |
+| `010_windowed_ceilings.md` | wfb | Sends and distinct children are counted over a bounded window, so a long session is never refused for work it did hours ago while a burst inside one window still is |
+| `020_legible_refusal.md` | wfc | The refusal names the ceiling that fired, is marked as a proxy decision rather than an upstream one, and the root budget can be read and cleared through the management API |
+
+## Write scope
+
+Permitted: `src/lib/workflow-budget.ts`, the workflow call sites in
+`src/server/responses/core.ts` and `src/server/index.ts`, the management read and
+mutation surface under `src/server/management/`, `src/server/request-log.ts` for the
+refusal provenance, their tests, and this unit.
+
+## Verification posture
+
+Local suite, typecheck, install and GUI build are **not run** for this unit by
+explicit instruction. Proof is hosted CI at the exact final head SHA and nothing
+else. Pushes use `--no-verify`.
+
+## What would make this fail
+
+Raising the numbers instead of fixing the shape. A bigger lifetime total is the
+same defect further away: it still refuses a session for work it finished hours
+ago, and it still cannot be seen or cleared. The window is the change; the numbers
+are a consequence of it.
+
+
+## Reproducing it
+
+Both probes carry the same body and differ only in the root id. Against a proxy
+whose process has been up long enough for a session to reach the ceiling:
+
+```bash
+BODY='{"model":"gpt-5.6-terra","input":[{"role":"user","content":[{"type":"input_text","text":"ok"}]}],"max_output_tokens":16,"stream":true}'
+
+# the long-running session's own root: refused
+curl -s -o /dev/null -w '%{http_code}\n' -N -X POST http://127.0.0.1:10100/v1/responses \
+  -H 'Content-Type: application/json' -H 'Accept: text/event-stream' \
+  -H "x-codex-parent-thread-id: <that session id>" -d "$BODY"
+
+# any root the process has not seen: served
+curl -s -o /dev/null -w '%{http_code}\n' -N -X POST http://127.0.0.1:10100/v1/responses \
+  -H 'Content-Type: application/json' -H 'Accept: text/event-stream' \
+  -H "x-codex-parent-thread-id: probe-$(date +%s)" -d "$BODY"
+```
+
+Two answers from one proxy, one body and one upstream, separated only by which
+root the request claims. That is what rules out the provider, the account and the
+model in a single step, and it is the check to run first the next time a fan-out
+starts failing for no visible reason.
+
+After a restart both return 200, which is the other half of the diagnosis: the
+ceiling is process-memory only, so the evidence disappears the moment anyone tries
+the obvious remedy.
+

devlog/_plan/260915_workflow_budget_window/010_windowed_ceilings.md

@@ -0,0 +1,52 @@
+# 010 — wfb: count sends and children over a window, not over a lifetime
+
+## Today
+
+```ts
+export function workflowSendCeilingReached(rootId, policy) {
+  const state = roots.get(rootId);
+  return state !== undefined && state.sends >= policy.maxPhysicalSends;
+}
+```
+
+`state.sends` only ever grows. `state.children` is a Set that only ever gains
+members. Neither has a clock. A root that made 256 sends in its first hour is
+refused for the rest of the process even if it sends nothing for a day.
+
+## The change
+
+Keep the counters, add a window. A root records its sends as timestamped buckets
+and the ceiling compares the count **inside the window** against
+`maxPhysicalSends`. Distinct children get the same treatment: a child seen once,
+hours ago, and never again should not hold a slot forever.
+
+The default window has to be argued for rather than picked. 256 sends is the
+number already in the tree and it was chosen against a fan-out, so the window is
+the interval over which that fan-out would be abusive. A ten-minute window keeps
+the original intent — seven hundred sends in a minute is still refused several
+times over — while an ordinary session that averages well under a send every two
+seconds never approaches it.
+
+`maxConcurrentChildren` stays as it is. Concurrency is already instantaneous; it
+has no lifetime problem to fix.
+
+## What must not change
+
+An unconfigured install must not see a refusal it would not have seen before.
+Windowing only ever admits more, never less, for the same traffic — the count
+inside a window is bounded by the lifetime count — so this direction is safe by
+construction. Say so in a test rather than trusting the argument.
+
+The eviction rules from #4546 stay: a root is evicted only when it is both
+inactive and not exhausted, and a full table refuses rather than laundering a
+fan-out into a fresh allowance. A windowed root that has aged out of its window is
+no longer exhausted, which is exactly the state that makes it evictable again.
+
+## Acceptance
+
+1. A root at the ceiling is admitted once its window rolls, without a restart.
+2. A burst inside one window is still refused at the same count as before.
+3. Distinct children age out of the window the same way sends do.
+4. Bucket storage per root is bounded; a root that sends forever does not grow
+   forever.
+

devlog/_plan/260915_workflow_budget_window/020_legible_refusal.md

@@ -0,0 +1,50 @@
+# 020 — wfc: a refusal an operator can read, name and clear
+
+## Today
+
+The refusal is a 429 with `workflow_budget_exhausted` and a sentence about the
+task's send budget. Two problems, and the first one cost hours.
+
+A 429 from a proxy that also forwards provider 429s is ambiguous. There is nothing
+on the record that says which one this was, so the first move is always to go look
+at the provider. This is the same defect #4639 fixed for the synthetic 503: a
+locally generated refusal presented under a field an operator reads as upstream.
+The fix there was provenance on the record, and it applies unchanged here.
+
+The second is that there is no way out. `resetWorkflowBudgetsForTest` is named for
+its audience and `workflowBudgetSnapshot` has no caller outside the module, so the
+state that decided the refusal cannot be read and cannot be cleared except by
+restarting the proxy — which drops every other root's accounting with it.
+
+## The change
+
+Name the ceiling. The denial type already distinguishes
+`workflow-sends-exhausted` from `workflow-children-exhausted` and the rest; carry
+that through to the error body and onto the request log instead of collapsing it
+into one sentence.
+
+Mark it local. The request log gains the same origin treatment #4639 introduced, so
+a proxy refusal and an upstream 429 are distinguishable on the record and on the
+management read surface.
+
+Expose and allow clearing. `GET` the root's budget through the management API so an
+operator can see a ceiling approaching rather than discovering it, and allow a
+bounded, recorded clear of one root. Clearing one root is not the same as
+restarting: it is scoped, it is logged, and it leaves every other root's accounting
+intact.
+
+## What must not change
+
+The clear is an operator action on the operator's own proxy, not a path a request
+can take. It goes through the management surface, which already requires a
+dashboard session or the admin token, and it must not be reachable from the data
+plane. A fan-out cannot be allowed to clear its own ceiling — that would make the
+budget a suggestion, which is the failure #4546 spent a release removing.
+
+## Acceptance
+
+1. The refusal body and the request log name which ceiling fired.
+2. The record marks the refusal as proxy-origin, distinguishable from an upstream 429.
+3. An operator can read one root's budget and clear it through the management API.
+4. The clear is scoped to one root, is recorded, and is not reachable from the data plane.
+

docs-site/src/content/docs/contributing.md

@@ -172,7 +172,7 @@ does not change `main`/`preview` review rules or allow direct pushes, force-push
 
 ## Adding a provider to the catalog
 
-All provider pickers and seeds derive from the canonical registry (`src/providers/registry.ts`):
+All provider pickers and seeds derive from the canonical registry (`src/providers/registry/entries-extended.ts`):
 
 ```ts
 {

docs-site/src/content/docs/fr/contributing.md

@@ -160,7 +160,7 @@ du dépôt et des chemins sensibles du point de vue de la sécurité est déclar
 
 ## Ajout d'un fournisseur au catalogue
 
-Tous les sélecteurs de fournisseurs et les graines proviennent du registre canonique (`src/providers/registry.ts`) :
+Tous les sélecteurs de fournisseurs et les graines proviennent du registre canonique (`src/providers/registry/entries-extended.ts`) :
 
 ```ts
 {

docs-site/src/content/docs/ja/contributing.md

@@ -123,7 +123,7 @@ Go ネイティブポートを担っていた `dev2-go` は廃止し、2 本の
 
 ## カタログにプロバイダーを追加
 
-すべてのプロバイダー選択肢と seed は canonical レジストリ(`src/providers/registry.ts`)から派生します。
+すべてのプロバイダー選択肢と seed は canonical レジストリ(`src/providers/registry/entries-extended.ts`)から派生します。
 
 ```ts
 {

docs-site/src/content/docs/ko/contributing.md

@@ -122,7 +122,7 @@ Go 네이티브 포트를 담당했던 `dev2-go`는 정리했고, 두 라인을
 
 ## 카탈로그에 프로바이더 추가하기
 
-모든 프로바이더 선택기와 seed는 canonical registry(`src/providers/registry.ts`)에서 파생됩니다.
+모든 프로바이더 선택기와 seed는 canonical registry(`src/providers/registry/entries-extended.ts`)에서 파생됩니다.
 
 ```ts
 {

docs-site/src/content/docs/ru/contributing.md

@@ -124,7 +124,7 @@ Pull request'ы с ребейзом приветствуются: ребейз 
 ## Добавление провайдера в каталог
 
 Все селекторы провайдеров и seed-данные выводятся из канонического реестра
-(`src/providers/registry.ts`):
+(`src/providers/registry/entries-extended.ts`):
 
 ```ts
 {

docs-site/src/content/docs/tr/contributing.md

@@ -186,7 +186,7 @@ sahipliği `.github/CODEOWNERS` dosyasında bildirilmiştir.
 ## Kataloğa sağlayıcı ekleme
 
 Tüm sağlayıcı seçicileri ve tohumları kurallı kayıt defterinden
-(`src/providers/registry.ts`) türetilir:
+(`src/providers/registry/entries-extended.ts`) türetilir:
 
 ```ts
 {

docs-site/src/content/docs/zh-cn/contributing.md

@@ -113,7 +113,7 @@ bun run release:watch               # 观察最新的 Release workflow run
 
 ## 向目录中添加 provider
 
-所有 provider picker 与 seed 都来自 canonical registry（`src/providers/registry.ts`）：
+所有 provider picker 与 seed 都来自 canonical registry（`src/providers/registry/entries-extended.ts`）：
 
 ```ts
 {

docs-site/src/content/docs/zh-tw/contributing.md

@@ -131,7 +131,7 @@ bun run release:watch               # 觀察最新的 Release workflow run
 
 ## 向目錄中新增 provider
 
-所有 provider picker 與 seed 都來自 canonical registry（`src/providers/registry.ts`）：
+所有 provider picker 與 seed 都來自 canonical registry（`src/providers/registry/entries-extended.ts`）：
 
 ```ts
 {