Fix workspace replay, stale cache reads and search state; refresh image/video models

.claude/skills/nexus-model-updates/SKILL.md

@@ -36,20 +36,24 @@ python3 .claude/skills/nexus-model-updates/scripts/check_model_registry.py \
    models live in the image adapters' own catalogs, not in `<Provider>Models.ts`,
    the structural gate does not see them, and the committed tool catalogs embed
    their enum.
-5. **A Live / realtime voice model** (Google `*-live*`, OpenAI `gpt-realtime*`)
+5. **A video generation model** — follow `protocols/add-video-model.md`. They
+   live in one static array in `src/services/llm/types/VideoTypes.ts`. The
+   structural gate does not see them, and an id missing from that array is
+   rejected even when the provider lists it.
+6. **A Live / realtime voice model** (Google `*-live*`, OpenAI `gpt-realtime*`)
    — follow `protocols/add-realtime-voice-model.md`. These ids answer only over
    a WebSocket session, so they do not belong in `<Provider>Models.ts`, the
    structural gate does not see them, and the provider smoke lane cannot prove
    them; the realtime smoke lane drives the shipped session instead.
-6. **Before calling any of the above done** — run `protocols/verify-model.md`. You
+7. **Before calling any of the above done** — run `protocols/verify-model.md`. You
    MUST get a zero exit from
    `scripts/check_model_registry.py --repo-root . <provider>`, and you MUST NOT
    report a model as working on the strength of a registry entry: the entry is a
    claim about an id, and only a live call tests it.
-7. **End of a session that used this skill** — run `protocols/self-refine.md`.
+8. **End of a session that used this skill** — run `protocols/self-refine.md`.
 
 ## Map
-- `protocols/` the procedures: add-model, add-image-model,
+- `protocols/` the procedures: add-model, add-image-model, add-video-model,
   add-realtime-voice-model, change-default, add-provider-registry,
   verify-model, self-refine.
 - `references/` mechanism, read on demand: `registry-anatomy.md` (what a

.claude/skills/nexus-model-updates/protocols/add-image-model.md

@@ -75,6 +75,11 @@ endpoint actually accepts, proven by one live generation through the adapter.
    the resulting 400 looks like a broken model.
 
 ## Guidelines
+- Pattern: Google's image docs show only Interactions API samples, but
+  `GeminiImageAdapter` calls `generateContent`. Before you add a Gemini image
+  id, check that `supportedGenerationMethods` in the keyed Google listing
+  (`references/provider-listings.md`) includes `generateContent`. Without a
+  key, record the entry as unverified.
 - Pattern: take the output format from the response's `media_type`, not from
   what was requested — Gemini returns JPEG regardless.
 - Pattern: read the wire enum, not the UI enum. Google's `imageConfig.imageSize`

.claude/skills/nexus-model-updates/protocols/add-video-model.md

@@ -0,0 +1,64 @@
+# Protocol: add-video-model
+
+Context: video generation models are in neither `<Provider>Models.ts` nor
+an image adapter. One static array, `VIDEO_MODELS` in
+`src/services/llm/types/VideoTypes.ts`, holds them. It decides what the video
+settings offer and what `generateVideo` accepts. A model missing from that
+array is rejected, even when the provider's live listing has it. The
+`generateVideo` tool schema carries no model enum, so tool catalogs do not
+change, and the structural gate does not see these entries either.
+
+## Mission
+A catalog entry that sends only aspect ratios, resolutions and durations the
+provider accepts, on a transport the adapter already uses.
+
+## Steps
+
+1. **Check the transport first.** `GoogleVideoAdapter` uses
+   `models/<id>:predictLongRunning`. Google's newer video models are served
+   only by the Interactions API (`POST /v1beta/interactions`), and that is
+   new adapter work, not a catalog entry. `OpenRouterVideoAdapter` speaks
+   OpenRouter's unified `POST /api/v1/videos`, so any model in its video
+   catalog uses the same request shape.
+
+2. **Source the facts.**
+
+   ```bash
+   curl -s https://openrouter.ai/api/v1/videos/models | python3 -c '
+   import json,sys
+   for m in json.load(sys.stdin)["data"]:
+       print(m["id"], m.get("supported_resolutions"), m.get("supported_aspect_ratios"),
+             m.get("supported_durations"), "audio", m.get("generate_audio"),
+             "frames", m.get("supported_frame_images"))'
+   ```
+
+   For Google direct, the parameter table on ai.google.dev/gemini-api/docs/video
+   is the source, and `/docs/deprecations` gives shutdown dates. Read it per
+   model: Veo on the Gemini API accepts `16:9`/`9:16` only, even though other
+   routes add `1:1`.
+
+3. **Fit the entry to the unions.** `VideoAspectRatio` and `VideoResolution`
+   are narrow, closed unions. Declare the intersection of the listing and the
+   union. A model with no value in the intersection (for example a `2K`-only
+   model) needs a type change and is not a catalog-only addition. OpenRouter's
+   resolution enum is case-sensitive (`4K`), and the adapter maps the internal
+   `4k` to it. Keep that mapping when you touch `submitJob`.
+
+4. **`supportsReferenceImage` means `input_references`** (reference-to-video)
+   on OpenRouter, not a first-frame image. A model whose listing names only
+   `supported_frame_images` and whose vendor does not document reference
+   images gets `false`.
+
+5. **Defaults.** `resolveDefaultVideoSelection` falls back to the first entry of
+   the first configured provider in `VIDEO_PROVIDER_PRIORITY`, and the shipped
+   `defaultVideoModel` in `src/types/llm/ProviderTypes.ts` is an independent
+   value. Keep each provider's first entry stable unless you mean to move the
+   fallback. When the shipped default moves, the `shipped video default` test in
+   `tests/unit/VideoGenerationService.test.ts` must still pass.
+
+6. **Live call.** `tests/debug/video-generation-live-smoke.test.ts` drives the
+   real adapters. Video generation is billed per second, so run it only with
+   the owner's go-ahead. Without that run, report the entry as unverified.
+
+## Next
+`verify-model.md` steps 1–3, then `self-refine.md`.

.claude/skills/nexus-model-updates/refinement-log.md

@@ -131,3 +131,5 @@ Append-only record of changes made by `protocols/self-refine.md`. Newest on top.
 - 2026-10-09 | Cost audit found a model cache-read rate still using the usual discount ratio despite a published exception. | Corrected the existing model price from the provider table and updated the regression expectation; no workflow change. | Files: AnthropicModels.ts, LLMCostCalculator.cacheRates.test.ts.
 
 - 2026-10-13: Authenticated model metadata can expose a lower output ceiling than streaming documentation. Used the keyed listing and kept transport-specific caps separate; no protocol change needed.
+
+- 2026-10-14 | Updating the image and video catalogs, the skill had no route for video models: they live in a single static `VIDEO_MODELS` array in `src/services/llm/types/VideoTypes.ts` that the gate does not see. Several problems in that file were only visible in the provider's listing: Veo routes declared `1:1`, which no Veo route accepts; the OpenRouter adapter sent `4k` to a case-sensitive `4K` enum; and Google's Veo 3.1 previews shut down on 2026-10-22 with an Interactions-API-only successor. Gemini image docs now show only Interactions samples while the adapter calls `generateContent`, and there was no Google key available to check `supportedGenerationMethods`. | Added `protocols/add-video-model.md` and a SKILL.md workflow row, plus a guideline in `add-image-model.md` to confirm `generateContent` before adding a Gemini image id. | Files: `SKILL.md`, `protocols/add-video-model.md`, `protocols/add-image-model.md`.

.claude/skills/nexus-storage/references/failure-modes.md

@@ -26,6 +26,23 @@ default, so an unknown type is dropped in silence during replay.
 Fix: write through the repository (event first, then cache) and add the `case` to
 the applier. Re-verify by rebuilding the cache again and confirming survival.
 
+## "After Rebuild cache the old value is still shown until reload"
+
+Two separate causes; check which before fixing.
+
+- A read answered from the repository `QueryCache` (shared by every repository,
+  60 s TTL) rather than SQLite. Anything that wipes or replays the database
+  (`StorageMaintenanceService.rebuildCache`, startup replay, `sync()`) must call
+  `getQueryCache().clear()` after the replay settles. Compare
+  `SELECT ... FROM workspaces` with `workspaceService.getWorkspace` in the
+  running plugin: a mismatch is this cause.
+- A replayed event resolved a UNIQUE constraint with `INSERT OR REPLACE`, which
+  deletes the *other* row (same name, different id) together with its edits.
+  Appliers must decide name collisions explicitly; `WorkspaceEventApplier`
+  keeps both rows and disambiguates the later-created name. Replay shards are
+  listed in filename order, so any "first wins" rule has to compare timestamps
+  or ids, never arrival order.
+
 ## "A list tool returns an entity, but loading it by ID or name says not found"
 
 Check `pluginStorage.migration.state` before blaming the point lookup. Affected

.claude/skills/nexus-storage/refinement-log.md

@@ -153,3 +153,9 @@ to 0). Edit `.skills/` and run `npm run sync:skills`; never edit a mirror. |
 ## 2026-10-13 — thinking disclosure implementation
 
 Targeted chat message writes continue through the existing adapter/repository event path; no schema or event-type additions. Missing targets fail explicitly. Existing storage guidance applied without a protocol correction; no change.
+
+## 2026-10-14 — Rebuild left stale reads; replay deleted same-name workspaces
+
+- Observation: "Rebuild cache" left `getWorkspace` answering from the QueryCache, and `WorkspaceEventApplier` replayed a same-name, different-id `workspace_created` with INSERT OR REPLACE, deleting the other workspace. Neither symptom was in failure-modes.md, and the "first wins" rule was order-dependent because shards replay in listing order.
+- Change: added the "old value is still shown until reload" entry to references/failure-modes.md.
+- Files: references/failure-modes.md, refinement-log.md

.cline/skills/nexus-model-updates/SKILL.md

@@ -36,20 +36,24 @@ python3 .claude/skills/nexus-model-updates/scripts/check_model_registry.py \
    models live in the image adapters' own catalogs, not in `<Provider>Models.ts`,
    the structural gate does not see them, and the committed tool catalogs embed
    their enum.
-5. **A Live / realtime voice model** (Google `*-live*`, OpenAI `gpt-realtime*`)
+5. **A video generation model** — follow `protocols/add-video-model.md`. They
+   live in one static array in `src/services/llm/types/VideoTypes.ts`. The
+   structural gate does not see them, and an id missing from that array is
+   rejected even when the provider lists it.
+6. **A Live / realtime voice model** (Google `*-live*`, OpenAI `gpt-realtime*`)
    — follow `protocols/add-realtime-voice-model.md`. These ids answer only over
    a WebSocket session, so they do not belong in `<Provider>Models.ts`, the
    structural gate does not see them, and the provider smoke lane cannot prove
    them; the realtime smoke lane drives the shipped session instead.
-6. **Before calling any of the above done** — run `protocols/verify-model.md`. You
+7. **Before calling any of the above done** — run `protocols/verify-model.md`. You
    MUST get a zero exit from
    `scripts/check_model_registry.py --repo-root . <provider>`, and you MUST NOT
    report a model as working on the strength of a registry entry: the entry is a
    claim about an id, and only a live call tests it.
-7. **End of a session that used this skill** — run `protocols/self-refine.md`.
+8. **End of a session that used this skill** — run `protocols/self-refine.md`.
 
 ## Map
-- `protocols/` the procedures: add-model, add-image-model,
+- `protocols/` the procedures: add-model, add-image-model, add-video-model,
   add-realtime-voice-model, change-default, add-provider-registry,
   verify-model, self-refine.
 - `references/` mechanism, read on demand: `registry-anatomy.md` (what a

.cline/skills/nexus-model-updates/protocols/add-image-model.md

@@ -75,6 +75,11 @@ endpoint actually accepts, proven by one live generation through the adapter.
    the resulting 400 looks like a broken model.
 
 ## Guidelines
+- Pattern: Google's image docs show only Interactions API samples, but
+  `GeminiImageAdapter` calls `generateContent`. Before you add a Gemini image
+  id, check that `supportedGenerationMethods` in the keyed Google listing
+  (`references/provider-listings.md`) includes `generateContent`. Without a
+  key, record the entry as unverified.
 - Pattern: take the output format from the response's `media_type`, not from
   what was requested — Gemini returns JPEG regardless.
 - Pattern: read the wire enum, not the UI enum. Google's `imageConfig.imageSize`

.cline/skills/nexus-model-updates/protocols/add-video-model.md

@@ -0,0 +1,64 @@
+# Protocol: add-video-model
+
+Context: video generation models are in neither `<Provider>Models.ts` nor
+an image adapter. One static array, `VIDEO_MODELS` in
+`src/services/llm/types/VideoTypes.ts`, holds them. It decides what the video
+settings offer and what `generateVideo` accepts. A model missing from that
+array is rejected, even when the provider's live listing has it. The
+`generateVideo` tool schema carries no model enum, so tool catalogs do not
+change, and the structural gate does not see these entries either.
+
+## Mission
+A catalog entry that sends only aspect ratios, resolutions and durations the
+provider accepts, on a transport the adapter already uses.
+
+## Steps
+
+1. **Check the transport first.** `GoogleVideoAdapter` uses
+   `models/<id>:predictLongRunning`. Google's newer video models are served
+   only by the Interactions API (`POST /v1beta/interactions`), and that is
+   new adapter work, not a catalog entry. `OpenRouterVideoAdapter` speaks
+   OpenRouter's unified `POST /api/v1/videos`, so any model in its video
+   catalog uses the same request shape.
+
+2. **Source the facts.**
+
+   ```bash
+   curl -s https://openrouter.ai/api/v1/videos/models | python3 -c '
+   import json,sys
+   for m in json.load(sys.stdin)["data"]:
+       print(m["id"], m.get("supported_resolutions"), m.get("supported_aspect_ratios"),
+             m.get("supported_durations"), "audio", m.get("generate_audio"),
+             "frames", m.get("supported_frame_images"))'
+   ```
+
+   For Google direct, the parameter table on ai.google.dev/gemini-api/docs/video
+   is the source, and `/docs/deprecations` gives shutdown dates. Read it per
+   model: Veo on the Gemini API accepts `16:9`/`9:16` only, even though other
+   routes add `1:1`.
+
+3. **Fit the entry to the unions.** `VideoAspectRatio` and `VideoResolution`
+   are narrow, closed unions. Declare the intersection of the listing and the
+   union. A model with no value in the intersection (for example a `2K`-only
+   model) needs a type change and is not a catalog-only addition. OpenRouter's
+   resolution enum is case-sensitive (`4K`), and the adapter maps the internal
+   `4k` to it. Keep that mapping when you touch `submitJob`.
+
+4. **`supportsReferenceImage` means `input_references`** (reference-to-video)
+   on OpenRouter, not a first-frame image. A model whose listing names only
+   `supported_frame_images` and whose vendor does not document reference
+   images gets `false`.
+
+5. **Defaults.** `resolveDefaultVideoSelection` falls back to the first entry of
+   the first configured provider in `VIDEO_PROVIDER_PRIORITY`, and the shipped
+   `defaultVideoModel` in `src/types/llm/ProviderTypes.ts` is an independent
+   value. Keep each provider's first entry stable unless you mean to move the
+   fallback. When the shipped default moves, the `shipped video default` test in
+   `tests/unit/VideoGenerationService.test.ts` must still pass.
+
+6. **Live call.** `tests/debug/video-generation-live-smoke.test.ts` drives the
+   real adapters. Video generation is billed per second, so run it only with
+   the owner's go-ahead. Without that run, report the entry as unverified.
+
+## Next
+`verify-model.md` steps 1–3, then `self-refine.md`.

.cline/skills/nexus-model-updates/refinement-log.md

@@ -131,3 +131,5 @@ Append-only record of changes made by `protocols/self-refine.md`. Newest on top.
 - 2026-10-09 | Cost audit found a model cache-read rate still using the usual discount ratio despite a published exception. | Corrected the existing model price from the provider table and updated the regression expectation; no workflow change. | Files: AnthropicModels.ts, LLMCostCalculator.cacheRates.test.ts.
 
 - 2026-10-13: Authenticated model metadata can expose a lower output ceiling than streaming documentation. Used the keyed listing and kept transport-specific caps separate; no protocol change needed.
+
+- 2026-10-14 | Updating the image and video catalogs, the skill had no route for video models: they live in a single static `VIDEO_MODELS` array in `src/services/llm/types/VideoTypes.ts` that the gate does not see. Several problems in that file were only visible in the provider's listing: Veo routes declared `1:1`, which no Veo route accepts; the OpenRouter adapter sent `4k` to a case-sensitive `4K` enum; and Google's Veo 3.1 previews shut down on 2026-10-22 with an Interactions-API-only successor. Gemini image docs now show only Interactions samples while the adapter calls `generateContent`, and there was no Google key available to check `supportedGenerationMethods`. | Added `protocols/add-video-model.md` and a SKILL.md workflow row, plus a guideline in `add-image-model.md` to confirm `generateContent` before adding a Gemini image id. | Files: `SKILL.md`, `protocols/add-video-model.md`, `protocols/add-image-model.md`.

.cline/skills/nexus-storage/references/failure-modes.md

@@ -26,6 +26,23 @@ default, so an unknown type is dropped in silence during replay.
 Fix: write through the repository (event first, then cache) and add the `case` to
 the applier. Re-verify by rebuilding the cache again and confirming survival.
 
+## "After Rebuild cache the old value is still shown until reload"
+
+Two separate causes; check which before fixing.
+
+- A read answered from the repository `QueryCache` (shared by every repository,
+  60 s TTL) rather than SQLite. Anything that wipes or replays the database
+  (`StorageMaintenanceService.rebuildCache`, startup replay, `sync()`) must call
+  `getQueryCache().clear()` after the replay settles. Compare
+  `SELECT ... FROM workspaces` with `workspaceService.getWorkspace` in the
+  running plugin: a mismatch is this cause.
+- A replayed event resolved a UNIQUE constraint with `INSERT OR REPLACE`, which
+  deletes the *other* row (same name, different id) together with its edits.
+  Appliers must decide name collisions explicitly; `WorkspaceEventApplier`
+  keeps both rows and disambiguates the later-created name. Replay shards are
+  listed in filename order, so any "first wins" rule has to compare timestamps
+  or ids, never arrival order.
+
 ## "A list tool returns an entity, but loading it by ID or name says not found"
 
 Check `pluginStorage.migration.state` before blaming the point lookup. Affected

.cline/skills/nexus-storage/refinement-log.md

@@ -153,3 +153,9 @@ to 0). Edit `.skills/` and run `npm run sync:skills`; never edit a mirror. |
 ## 2026-10-13 — thinking disclosure implementation
 
 Targeted chat message writes continue through the existing adapter/repository event path; no schema or event-type additions. Missing targets fail explicitly. Existing storage guidance applied without a protocol correction; no change.
+
+## 2026-10-14 — Rebuild left stale reads; replay deleted same-name workspaces
+
+- Observation: "Rebuild cache" left `getWorkspace` answering from the QueryCache, and `WorkspaceEventApplier` replayed a same-name, different-id `workspace_created` with INSERT OR REPLACE, deleting the other workspace. Neither symptom was in failure-modes.md, and the "first wins" rule was order-dependent because shards replay in listing order.
+- Change: added the "old value is still shown until reload" entry to references/failure-modes.md.
+- Files: references/failure-modes.md, refinement-log.md