feat(llm): native Anthropic Messages API wire protocol "anthropic"

CONTRIBUTORS.md

@@ -34,6 +34,33 @@ else's environment — that led to a fix landing in the repository.
   hardcoded `~/.claude/settings.json` that ignored `CLAUDE_CONFIG_DIR`. A/B/A
   probe-script reproduction isolating schema-vs-path, root-cause to the exact
   functions, and a verified fix delivered as a pull request.
+- **Damien Moon** ([@DamieMoon](https://github.com/DamieMoon)) — PR
+  [#46](https://github.com/GottZ/ctx/pull/46): on a CPU embed backend one
+  large block turned every following `ctx query` into a 500 after the
+  client's 120 s deadline — the background backfill only ran inside the
+  dream loop (off by default), and the query path's inline backfill had a
+  count cap but no time budget, so the slow block ate the caller's whole
+  deadline and was memoized as a backend outage. Server and llama.cpp logs,
+  a live before/after on the affected host, and the shipped fix: a
+  dedicated embed-backfill arm that runs regardless of Dream, a wall-clock
+  `sync_budget` for the query path, and a `caller_timeout` memo class that
+  hands an overrunning block to the arm. Hardened with the arm staying
+  inside the iterated tenant's scope window, opt-outs that restore the
+  pre-PR paths, a pick that skips blocks no trusted backend can serve,
+  no idle wait after a parked block, hot interval changes, and a test
+  that pins the arm's wiring into `Run`. The review of this PR also
+  surfaced a cross-tenant egress path in the existing dream-loop backfill,
+  fixed separately in v5.23.1. Also PR
+  [#47](https://github.com/GottZ/ctx/pull/47): a third wire protocol,
+  `anthropic` — the native Messages API with prompt caching on the stable
+  system prefix, effort control, SSE streaming with tool use for the web
+  chat, Migration 155, probe, seed and wizard support — so a hosted-Claude
+  install no longer needs a sampling-stripping sidecar in front of the
+  OpenAI-compatible layer. Hardened with a closing call that keeps the tool
+  definitions under `tool_choice: none`, a per-effort thinking reserve on
+  `max_tokens`, validation of keys the Messages API rejects, no
+  whitespace-only text blocks, and a tool-loop replay in the API's block
+  order.
 - **TurgutKural** ([@TurgutKural](https://github.com/TurgutKural)) — PR
   [#9](https://github.com/GottZ/ctx/pull/9): `/health` reported every cloud
   inference backend as down. The probe demanded HTTP 200 on `GET /`, which only

README.md

@@ -73,7 +73,7 @@ Running the server (Go daemon + PostgreSQL 18 + pgvector) is `docker compose up
 
 > **Upgrading from v4? Reach v4.38.0 first.** v5.0.0 removes the `CTX_{CHAT,CHAT_FALLBACK,EMBED,DREAM,DREAM_EMBED,RERANK}_*` connection variables: backends are configured in the pool (`ctx backends`) and nowhere else. **v4.38.0 is the last release that still reads those variables**, and one boot on it carries a `.env`-only configuration into the pool for you, then names every variable and leftover settings row you still have. Going straight from an older v4 to v5 is not fatal down to v4.19.0 — the env bridge is gone by then and you re-enter your topology by hand with `ctx backends seed` — but **below v4.19.0 it is a hard stop**: v5 folds migrations 001–113 into a single baseline and refuses a database that never applied 113. Migration path and rollback notes: [operations](docs/operations.md#migration-133-the-backend-tuple-rows-are-deleted--and-the-upgrade-hop-that-has-to-come-first).
 
-**Models:** the tested reference stack is qwen3.6:27b for chat/synthesis/dream/digest, Qwen3-Embedding-8B (1024-dim, CPU-sized) for embeddings, and a bge-reranker-v2-m3 cross-encoder sidecar for rerank — but ctx is engine-agnostic: any OpenAI-compatible or Ollama API works, models are per-backend configuration. CPU-only setups work; a GPU shortens chat latency. Details: [Models & hardware](docs/operations.md#models--hardware).
+**Models:** the tested reference stack is qwen3.6:27b for chat/synthesis/dream/digest, Qwen3-Embedding-8B (1024-dim, CPU-sized) for embeddings, and a bge-reranker-v2-m3 cross-encoder sidecar for rerank — but ctx is engine-agnostic: any OpenAI-compatible or Ollama API works, and the chat-class roles can run on the native Anthropic Messages API (`protocol: "anthropic"`); models are per-backend configuration. CPU-only setups work; a GPU shortens chat latency. Details: [Models & hardware](docs/operations.md#models--hardware).
 
 ## How LLMs use ctx
 

docker-compose.yml

@@ -362,8 +362,10 @@ services:
       # ── status.* (1) ──
       CTX_STATUS_CHANNEL_PROBE_INTERVAL: ${CTX_STATUS_CHANNEL_PROBE_INTERVAL:-}
 
-      # ── embed_backfill.* (4) ──
+      # ── embed_backfill.* (6) ──
       CTX_EMBED_BACKFILL_SYNC_CAP: ${CTX_EMBED_BACKFILL_SYNC_CAP:-}
+      CTX_EMBED_BACKFILL_SYNC_BUDGET: ${CTX_EMBED_BACKFILL_SYNC_BUDGET:-}
+      CTX_EMBED_BACKFILL_INTERVAL: ${CTX_EMBED_BACKFILL_INTERVAL:-}
       CTX_EMBED_BACKFILL_MAX_TOKENS: ${CTX_EMBED_BACKFILL_MAX_TOKENS:-}
       CTX_EMBED_BACKFILL_BACKOFF_BASE: ${CTX_EMBED_BACKFILL_BACKOFF_BASE:-}
       CTX_EMBED_BACKFILL_BACKOFF_CAP: ${CTX_EMBED_BACKFILL_BACKOFF_CAP:-}