## Problem/motivation The pre-beta5 review in #3586404+s found live-path holes plus leftover consumer/runtime bugs. Shipped types do not trigger the fatal cases today, but `AiContextConsumerTypeInterface` is the public extension API. Fix in rc1 before contrib types exist in the wild. ### 1. No error isolation on the live AI request `AiContextPreGenerateResponseSubscriber::onPreGenerateResponse()` has no try/catch and runs inside the provider `PreGenerateResponseEvent`. `AiContextBuildSystemPromptSubscriber` has the same gap on the agent path. `AiContextConsumerRouter::getRoutingMap()` instantiates every registered consumer type on cache miss. `routeRequest()` calls `createInstance()` with no `hasDefinition()` guard, so a stale persistent `ai_context_consumer_routing_map` cache entry can feed a vanished type id into `createInstance()`. A contrib type that throws in `create()` or `getRoutingRequestTags()` 500s the first chat request after a cache clear. Context push should be decorative and must not kill the provider call. ### 2. Nested subcontext chat is not excluded from routing Conditional subcontext resolution issues its own `$provider->chat()` with tags `ai_context` and `ai_context_subcontext` from inside the pre-generate handler (`AiContextSubcontextResolver`). The nested call dispatches a fresh event with a fresh `ChatInput`, so the existing per-input reentrancy marker does not apply. `AiContextConsumerRouter::EXCLUDED_REQUEST_TAGS` excludes `ai_agents` and `ai_assistant_api` but not CCC's own tags. A contrib type that routes on a broad tag can do: select → nested chat → route again → unbounded recursion. ### 3. Leftover live-path consumer and request edge cases - Disabled or unavailable consumer types still create `AMBIGUOUS` fail-closed, so a kill switch suppresses the healthy type instead of removing the dead one from contention. - `push_enabled: null` / omitted must resolve to the type default. #3586444+s froze the contract and landed the code; there is no regression test. - Request-attribute leaks across AI calls in one HTTP request: `applyRequestInfo()` writes `path` and nothing clears it (`clearPath()` is dead). A token-supplied path or entity from one AI call leaks into Site Sections matching and usage routes for later calls, and the stored entity shadows the route entity. - Reentrancy marker caches negative terminal statuses. A second `chat()` on the same `ChatInput` re-publishes a stale `DECLINED` even when the automator ID tag is now resolvable. - Once the subscribers are isolated, a usage-tracking failure after selection succeeded would be swallowed by the outer catch. On the generic path the context block is already on the input but no `pushed` marker is written, so a tool-call inner request appends a second block. On the agent path the prompt is never set. ## Proposed resolution - Wrap the pre-generate subscriber body in `try/catch (\Throwable)`, log, and return. Apply the same isolation in `AiContextBuildSystemPromptSubscriber` for both the system-prompt and loop-awareness listeners. - In the router, skip vanished or throwing types (`hasDefinition()` / catch `\Throwable`), drop a routing map that names a vanished type, and never cache a map built while a type was throwing. - Add `ai_context` and `ai_context_subcontext` to `EXCLUDED_REQUEST_TAGS` as named constants and use them in `AiContextSubcontextResolver`. - Add an in-flight boolean on the subscriber so a nested CCC chat cannot re-enter selection. - Drop disabled/unavailable types from contention only when more than one type claimed the request, so a lone claim still reports `type_disabled` per the #3586444+s status contract. - Add regression tests for omitted / `null` `push_enabled` resolving to the type default. - Reset path, entity and bundle attributes this module applied for a previous AI call at the start of every routed call. Only keys this service wrote are cleared; values other modules set through `setPath()` / `setEntity()` / `setEntityBundle()` are left alone. - Replay a stored invocation result only when its status is `pushed`. - Catch usage-tracking failures on their own in both subscribers so the push still completes and the `pushed` marker is written. ## Split off Not in this issue: - Duplicate `consumer_id` rows and validator index mapping: #3586541+s - Usage-tracking upsert race, cron usage purge cap, and `site:install --existing-config` orphaning from a beta4 export: 1.0.1 under #3586389+s (the `--existing-config` conversion belongs with #3586470+s) - `ListAiContextItems::clearEntity()` also removing the current call's bundle hint: one line on #3586389+s, not its own issue ## Tests - Kernel: a consumer type that throws in `create()`, `resolveConsumerId()` or `getRoutingRequestTags()` leaves the provider call intact and logs. - Kernel: a stale routing-map type id is skipped, logged once, and the map is rebuilt on the next request. - Kernel: a nested chat tagged `ai_context` / `ai_context_subcontext` does not route or select again; a real conditional-subcontext call through `ProviderProxy` is not handled. - Kernel: selection failures on both paths leave the request and agent prompt intact; loop-awareness failures are logged. - Kernel: usage-tracking failures still push on the generic path (marker written, one block after replay) and still append on the agent path. - Kernel: disabled / unavailable type out of contention; omitted / `null` `push_enabled` inherit the type default; path and entity hints do not leak between calls, including declined calls; route entity wins again; negative result not replayed. - Unit: router dead / vanished / throwing types, incomplete map not cached; resolver reset keeps values another module set and records only keys actually written. ## Out of scope - Changing shipped Agent / Automator routing tags - Redesigning conditional subcontext LLM calls (decode robustness follow-up #3586542+s) - Beta5 upgrade-path items already on #3586470+s (10028 mode flip, settings import, post-update 0006) - Other public API freeze items on #3586444+s (`getSettings()`, `getResultForSelection()`, `fromArray()` `entity_id`, clone `getCacheableMetadata()`) - Consumer-description tooltip XSS (already `strip_tags()` on the pill) - Persistent selection-result caching (#3586474+s) - Per-item `isEnabledAndAvailable()` lookups in the selector match loop (#3586475+s) - Review nits not filed: double enabled/available evaluation per request, per-request memo for an incomplete routing map, misleading "Ambiguous" log when both claiming types are dead, logging `$e->getMessage()` in the new catch blocks ## Follow-up issues - #3586541+s - #3586542+s ## Related issues - #3586404+s - #3586444+s - #3586461+s - #3586470+s - #3586389+s ## AI usage - [x] AI assisted issue