The cutover's first real lesson ships: tool_choice "required" was a
workaround for a backend that ignored it, and became an unbreakable
tool loop on a backend that obeys it every request (~80 s arithmetic
turns, observed). The health check now learns which server answers
behind OLLAMA_HOST (/props is llama-server's own surface) and only
Ollama gets the advisory nudge. Probed unforced on llama-server: 3/3
tool calls via the --jinja template.
Also carries BACKEND_SLOT_PINNING (default off) for the P3 pilot.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Each pipeline phase owns one llama-server slot (steward 0, orchestrator
1, synthesizer 2), carried as id_slot in extra_body through the same
mechanism tool_choice already uses, so the phase's stable prompt prefix
stays in that slot's KV cache and a turn re-prefills only its new
tokens. Off by default; a no-op on the Claude backend and ignored by
Ollama, so the flag is safe on any backend and the cutover itself stays
a pure env swap.
The merge helper preserves existing extra_body keys — mutation-checked
(dropping the merge fails exactly the test written for it).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three native-API touchpoints converted — steward /api/generate to
/v1/chat/completions, embeddings /api/embeddings to /v1/embeddings,
health /api/tags to /v1/models — so the backend behind OLLAMA_HOST is
swappable by env alone. This makes the serving plan's "tatlock needs
zero changes" claim true for the llama-server cutover and for forge
after it. EMBEDDING_HOST (default: OLLAMA_HOST) lets gen and embed
point at different servers, which the boilerroom stack needs.
The dead OllamaClient goes with it: a native-API client nothing
imported, whose presence would make the post-cutover "no native
endpoints" grep lie.
Contract tests rewritten to mirror the new requests and extended with
the embeddings shape (dim must match the configured Qdrant dimension).
All pass against live Ollama's /v1 — deployable before any cutover.
Both new assertions mutation-checked via env overrides (bogus model,
wrong dim: each fails). The Anthropic contract now skips on 401: the
configured key is deliberately revoked per workspace D-11, which is
"fallback disabled", not a boundary break. Embedding continuity across
backends was measured separately: same nomic bytes, cosine 1.0000.
663 unit tests pass; ruff clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rolls back the claudification backend preference: PREFER_CLOUD_BACKEND now
defaults to false, resolve_backend() picks Ollama first and uses Claude when
explicitly preferred or when the new Ollama startup health check fails. The
Steward retries mid-request failures on the other backend in both directions.
Also hardens the fallback itself: Anthropic SDK imports are lazy so a broken
anthropic package degrades to Ollama-only instead of crashing at import time
(root cause of the production outage since April), anthropic is pinned to a
pydantic-ai-1.27-compatible range, ANTHROPIC_MODEL defaults to claude-sonnet-5
(sonnet-4-20250514 retired 2026-06-15), sampling parameters are stripped from
Claude calls (Sonnet 5 rejects them), and the Steward timeout is configurable
(STEWARD_TIMEOUT, default 60s) since gemma4 needs ~35s warm for analysis.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PydanticAI handles tool_choice natively for Anthropic. The extra_body
hack caused an infinite tool call loop where Claude kept calling the
same tool because tool_choice was forced to "any".
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
AnthropicModel doesn't accept api_key directly; it must be passed
through an AnthropicProvider instance.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
All agents now prefer Claude API when ANTHROPIC_API_KEY is configured,
with automatic fallback to Ollama when offline or unconfigured. New
src/anthropic/ module provides model selection via get_model() factory.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
jpmschweitzerandClaude Fable 5