The release audit swept every agent.run and raw chat call: the two
streaming orchestrate paths never carried a binding (they predate the
v2.6.0 pinning), and the biographer and housekeeper ran bare. Through
the wrapper a session-less call takes idle slots only and 503s once
four sessions are resident — so the streaming paths join
tatlock-orchestrate, and the experts share tatlock-experts at rank 35:
between the librarian and the phases, the two least-used consumers
trade the spare slot by rank instead of anyone hitting an empty pool.
Rank mutation shown to fail its test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The flavor probe gains a third answer: the wrapper names itself on
/health, so behind OLLAMA_HOST tatlock now distinguishes boilerroom,
a bare llama-server, and Ollama. Through the wrapper each pipeline
phase is a named session with the decided eviction ranking —
tatlock-steward/-orchestrate/-synthesize at 40, librarian at 30,
lower parks sooner (webber will sit at 20; Open WebUI stays
session-less and can never evict anyone). Against a bare llama-server
the raw id_slot pins survive unchanged, Ollama gets neither, and an
unprobed flavor sends nothing rather than guessing — the backend
stays swappable by env alone. tool_choice through the wrapper follows
the llama-server rule, since that is who answers.
The wrapper's balancing and compaction-due signals are read
everywhere: an httpx response hook on the provider covers every
PydanticAI call, streams included, and the steward's raw call reads
the body extras. Acting on compaction_due is a future ticket — the
signal just must not pass silently.
Verified end to end against the live wrapper: the dev server probed
flavor=boilerroom, a full pipeline turn answered in 5.7 s, and
GET /sessions showed all three phase sessions resident at rank 40
with engine-reported occupancies. The demonstration also filled the
production slot map — session-less prod delegations would have 503d
— cleared by a wrapper restart and filed as boilerroom T-11 (sessions
need an exit). A latent test flaw surfaced too: the ollama
tool_choice test relied on the dev backend probing as ollama; it now
pins the flavor it claims to test.
27 selector tests (9 new), 677 total green; three mutations shown to
fail their tests (librarian rank, the wrapper branch, the no-nudge
set); the new wrapper contract class runs 10/10 against the live
boundary.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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>
Mechanical only, and separated from the judgment calls that follow so the
reviewable changes are not buried in a 98-file whitespace diff.
227 automatic fixes: 60 blank lines carrying whitespace, 60 unsorted import
blocks, 34 Optional[X] to X | None, 28 unused imports, 16 deprecated typing
imports, 12 datetime.timezone.utc to datetime.UTC, and assorted smaller
modernisations. Then `ruff format` over src and tests: 98 files reformatted,
35 already conforming.
No file among the unused-import findings defines __all__ or is an __init__.py,
so nothing here removes a re-export.
`make test`: 658 passed, unchanged from HEAD.
Two things observed while verifying, neither addressed here:
`pytest tests/` cannot collect — tests/e2e/test_orchestration_e2e.py uses an
`e2e` marker that is not registered, and the config is strict about markers.
This fails identically at HEAD, so it predates this change; `make test` passes
because it ignores tests/e2e, tests/integration and tests/contracts.
test_tatlock_tool_call_logging_calculator is flaky. It failed once in a full run
with these changes and passed on the next, passes in isolation with them, and
fails in isolation at HEAD. It is order- or timing-dependent, not a regression
from this commit — established by running the full suite both ways rather than
by reasoning about which change could have caused it.
Co-Authored-By: Claude <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>
jpmschweitzerandClaude Fable 5