release v2.7.0

Named sessions through the boilerroom wrapper, ranked by use:
pipeline phases 40, experts 35, librarian 30 — lower parks sooner.
Balancing and compaction signals read on every response.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.env.example

@@ -8,11 +8,16 @@ API_HOST=0.0.0.0
 API_PORT=8000
 API_PREFIX=/v1
 
-# Ollama Configuration (local - primary backend)
+# Local backend (primary). The var name is legacy: any OpenAI-compatible
+# server works — Ollama, a bare llama-server, or the boilerroom wrapper
+# (http://127.0.0.1:8602 on tower-of-joy). The flavor is probed at
+# startup; through the wrapper, BACKEND_SLOT_PINNING=true sends named
+# sessions with the T-4 eviction ranking instead of raw slot pins.
 OLLAMA_HOST=http://localhost:11434
 OLLAMA_DEFAULT_MODEL=gemma4:e2b
 OLLAMA_TIMEOUT=120
 STEWARD_TIMEOUT=60
+BACKEND_SLOT_PINNING=false
 
 # Anthropic Configuration (Claude - cloud fallback)
 # Set ANTHROPIC_API_KEY to keep the Claude fallback available: it is used

.gitea/workflows/build.yml

@@ -10,12 +10,20 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Create Gitea Release
+        # Idempotent: a re-fired tag finds its release already present and
+        # says so instead of failing on the 409.
         run: |
+          api="${{ github.server_url }}/api/v1/repos/${{ github.repository }}/releases"
+          auth='Authorization: token ${{ secrets.GITHUB_TOKEN }}'
+          if curl -sf -H "$auth" "$api/tags/${{ github.ref_name }}" > /dev/null; then
+            echo "release for ${{ github.ref_name }} already exists — nothing to do"
+            exit 0
+          fi
           curl -sf -X POST \
-            -H "Authorization: token ${{ secrets.GITHUB_TOKEN }}" \
+            -H "$auth" \
             -H "Content-Type: application/json" \
             -d '{"tag_name": "${{ github.ref_name }}", "name": "Release ${{ github.ref_name }}", "body": "Automated release for ${{ github.ref_name }}"}' \
-            "${{ github.server_url }}/api/v1/repos/${{ github.repository }}/releases"
+            "$api"
 
   build:
     runs-on: ubuntu-latest

.pql/changelog/ticket_history/2026-09.sql

@@ -0,0 +1,2 @@
+INSERT INTO ticket_history (ticket_record_id, field, old_value, new_value, changed_by, changed_at, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G99D07EM5DZWTS0WD9V8PSKM', 'status', 'backlog', 'in_progress', NULL, '2026-09-12 13:50:48', '2026-09-12 13:50:48.289', '2026-09-12 13:50:48.289', NULL, '2e88d155325b3e704787f48b2aa97591', 2) ON CONFLICT(hash) DO NOTHING;
+INSERT INTO ticket_history (ticket_record_id, field, old_value, new_value, changed_by, changed_at, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G99D07EM5DZWTS0WD9V8PSKM', 'status', 'in_progress', 'done', NULL, '2026-09-12 14:13:29', '2026-09-12 14:13:29.027', '2026-09-12 14:13:29.027', NULL, '5b9dd9a62002db0fb4dc58f7f9ffcbb6', 2) ON CONFLICT(hash) DO NOTHING;

.pql/changelog/ticket_idmap/2026-09.sql

@@ -0,0 +1,3 @@
+INSERT INTO ticket_idmap (record_id, ticket_id, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G94KC9XM3PW91C6B7VJA6VRM', 'T-3', '2026-09-11 21:09:10.267', '2026-09-11 21:09:10.267', NULL, 'a1ccf4d118877c230dffb961a7531cb4', 2) ON CONFLICT(record_id) DO UPDATE SET ticket_id=excluded.ticket_id, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= ticket_idmap.updated_at;
+INSERT INTO ticket_idmap (record_id, ticket_id, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G99D07EM5DZWTS0WD9V8PSKM', 'T-4', '2026-09-12 08:20:21.494', '2026-09-12 08:20:21.494', NULL, '5b28e07bdc3be1f9ed0ed71d2da60c72', 2) ON CONFLICT(record_id) DO UPDATE SET ticket_id=excluded.ticket_id, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= ticket_idmap.updated_at;
+INSERT INTO ticket_idmap (record_id, ticket_id, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G99D07Z3FZY8EBT9PJ1RCGHW', 'T-5', '2026-09-12 08:20:21.625', '2026-09-12 08:20:21.625', NULL, '0a417012779f612a4dad1a7dbd0787e3', 2) ON CONFLICT(record_id) DO UPDATE SET ticket_id=excluded.ticket_id, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= ticket_idmap.updated_at;

.pql/changelog/tickets/2026-09.sql

@@ -0,0 +1,5 @@
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G94KC9XM3PW91C6B7VJA6VRM', 'story', NULL, 'Stream the synthesis phase — TTFT equals wall time today', 'Nothing streams until steward-orchestrate-synthesize completes, so desklock''s time-to-first-word is the full pipeline (TTFT mean 14.9s, max 26.2s — measured 2026-09-11, workspace note cladmin/agent/notes/2026-09-11-serving-baseline.md). Stream the synthesis phase out as it generates: a first-class goal of the serving rework independent of KV reuse (workspace D-60). The backend now streams /v1 SSE natively (llama-gen).', 'backlog', 'high', NULL, NULL, NULL, '2026-09-11 21:09:10.253', '2026-09-11 21:09:10.253', NULL, 'f7d31ca4fe747835a32744c310f51674', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= tickets.updated_at;
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G99D07EM5DZWTS0WD9V8PSKM', 'story', NULL, 'Adopt boilerroom named sessions with the usage ranking', 'Replace raw id_slot pinning (BACKEND_SLOT_PINNING) with the wrapper''s session extension fields, now that tatlock routes through boilerroom:8602 (since 2026-09-12). Ranking by likelihood of use, decided 2026-09-12 — lower eviction_order parks sooner: tatlock pipeline phases highest (eviction_order 40; e.g. sessions tatlock-steward/tatlock-orchestrate/tatlock-synthesize), librarian-routed calls as session librarian (30), webber will sit at 20 (workspace T-137), Open WebUI stays session-less by design (idle slots only, can never evict — and it is being deprecated for an own-webui build). Read the x-boilerroom-balancing and compaction_due signals; the compaction endpoint expects the client transcript (tatlock''s Redis/Qdrant context is the natural producer — boilerroom D-5 names tatlock the first producer).', 'backlog', 'high', NULL, NULL, NULL, '2026-09-12 08:20:21.493', '2026-09-12 08:20:21.493', NULL, '3fb8c350876fc056648a19e0f2b7718c', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= tickets.updated_at;
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G99D07Z3FZY8EBT9PJ1RCGHW', 'bug', NULL, 'The container logs nowhere: stdout empty, /app/logs mount untouched', 'docker logs tatlock is 0 lines and the mounted ~/docker-data/tatlock/logs has been empty since 2025-12 — the running app writes logs to neither, which made the 2026-09-12 wrapper-cutover verification blind (flavor detection had to be proven behaviorally). Dev mode logs to build/logs/server.log via make run; production should log to stdout for docker logs at minimum.', 'backlog', 'medium', NULL, NULL, NULL, '2026-09-12 08:20:21.624', '2026-09-12 08:20:21.624', NULL, '3fd294efc9afa857aef15810d9003298', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= tickets.updated_at;
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G99D07EM5DZWTS0WD9V8PSKM', 'story', NULL, 'Adopt boilerroom named sessions with the usage ranking', 'Replace raw id_slot pinning (BACKEND_SLOT_PINNING) with the wrapper''s session extension fields, now that tatlock routes through boilerroom:8602 (since 2026-09-12). Ranking by likelihood of use, decided 2026-09-12 — lower eviction_order parks sooner: tatlock pipeline phases highest (eviction_order 40; e.g. sessions tatlock-steward/tatlock-orchestrate/tatlock-synthesize), librarian-routed calls as session librarian (30), webber will sit at 20 (workspace T-137), Open WebUI stays session-less by design (idle slots only, can never evict — and it is being deprecated for an own-webui build). Read the x-boilerroom-balancing and compaction_due signals; the compaction endpoint expects the client transcript (tatlock''s Redis/Qdrant context is the natural producer — boilerroom D-5 names tatlock the first producer).', 'in_progress', 'high', NULL, NULL, NULL, '2026-09-12 08:20:21.493', '2026-09-12 13:50:48.289', NULL, '62abeb4da2f1f7aedb8d92884e7e148c', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= tickets.updated_at;
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06G99D07EM5DZWTS0WD9V8PSKM', 'story', NULL, 'Adopt boilerroom named sessions with the usage ranking', 'Replace raw id_slot pinning (BACKEND_SLOT_PINNING) with the wrapper''s session extension fields, now that tatlock routes through boilerroom:8602 (since 2026-09-12). Ranking by likelihood of use, decided 2026-09-12 — lower eviction_order parks sooner: tatlock pipeline phases highest (eviction_order 40; e.g. sessions tatlock-steward/tatlock-orchestrate/tatlock-synthesize), librarian-routed calls as session librarian (30), webber will sit at 20 (workspace T-137), Open WebUI stays session-less by design (idle slots only, can never evict — and it is being deprecated for an own-webui build). Read the x-boilerroom-balancing and compaction_due signals; the compaction endpoint expects the client transcript (tatlock''s Redis/Qdrant context is the natural producer — boilerroom D-5 names tatlock the first producer).', 'done', 'high', NULL, NULL, NULL, '2026-09-12 08:20:21.493', '2026-09-12 14:13:29.027', NULL, '89ce1be323ebb26f00c63569f197cab1', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at >= tickets.updated_at;

CHANGELOG.md

@@ -7,6 +7,48 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [2.7.0] - 2026-09-13
+
+### Added
+
+- Named sessions through the boilerroom wrapper (T-4): the backend
+  flavor probe now recognizes the wrapper by its `/health` service
+  name, and each pipeline phase sends `session`/`eviction_order`
+  extension fields instead of raw `id_slot` — `tatlock-steward`,
+  `tatlock-orchestrate`, `tatlock-synthesize` at rank 40 and the
+  librarian at 30, per the 2026-09-12 usage ranking (lower parks
+  sooner; webber will sit at 20, Open WebUI stays session-less).
+  Against a bare llama-server the old slot pins still apply, and
+  Ollama gets neither, so the backend remains swappable by env alone.
+  The wrapper's `x-boilerroom-balancing` and compaction-due signals
+  are read on every response (an httpx hook on the provider, body
+  extras on the steward's raw call) and logged — acting on
+  `compaction_due` is a future ticket. `get_model_info()` reports the
+  probed `local_flavor`. Wire-pinned by a new boilerroom wrapper
+  contract class in `make test-contracts`.
+- Every generation call is named: the streaming orchestrate paths and
+  the biographer/housekeeper experts (shared session `tatlock-experts`,
+  rank 35) — through the wrapper a session-less call takes idle slots
+  only and would 503 once four sessions are resident, so the release
+  audit swept every `agent.run` and raw chat call under a session.
+
+## [2.6.0] - 2026-09-11
+
+### Added
+
+- `BACKEND_SLOT_PINNING` (default off): pins each pipeline phase to a
+  llama-server slot (steward=0, orchestrator=1, synthesizer=2) so a
+  phase's stable prompt prefix stays in that slot's KV cache. No-op on
+  Claude; ignored by Ollama.
+
+### Fixed
+
+- Tool-calling no longer loops on llama-server backends: the health
+  check detects the server flavor, and `tool_choice: "required"` (an
+  advisory nudge on Ollama, an every-request mandate on llama-server)
+  is sent only to Ollama. On llama-server the model calls tools
+  unforced via its chat template.
+
 ## [2.5.0] - 2026-09-11
 
 ### Added

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tatlock"
-version = "2.5.0"
+version = "2.7.0"
 description = "OpenAI-compatible API with Ollama backend"
 requires-python = ">=3.12"
 dependencies = [

src/agents/biographer/agent.py

@@ -191,9 +191,12 @@ async def run_biographer(
     )
 
     try:
+        from src.anthropic.model_selector import with_phase_binding
+
         result = await agent.run(
             prompt,
             message_history=message_history,
+            model_settings=with_phase_binding(None, "experts"),
         )
 
         logger.info(

src/agents/housekeeper/agent.py

@@ -206,12 +206,12 @@ async def run_housekeeper(
 
     try:
         # Temperature 0.1 for slight exploration (skipped on Claude backend)
-        from src.anthropic.model_selector import get_sampling_settings
+        from src.anthropic.model_selector import get_sampling_settings, with_phase_binding
 
         result = await agent.run(
             prompt,
             message_history=message_history,
-            model_settings=get_sampling_settings(0.1),
+            model_settings=with_phase_binding(get_sampling_settings(0.1), "experts"),
         )
 
         logger.info(

src/agents/librarian/agent.py

@@ -272,11 +272,14 @@ async def run_librarian(
     )
 
     try:
+        from src.anthropic.model_selector import with_phase_binding
+
         # One shared library-desk connection for all tool calls in this run
         async with library_client_session():
             result = await agent.run(
                 prompt,
                 message_history=message_history,
+                model_settings=with_phase_binding(None, "librarian"),
             )
 
         logger.info(

src/agents/steward/agent.py

@@ -162,19 +162,37 @@ class StewardAgent:
         OpenAI-compatible server (Ollama, llama-server) can sit behind
         OLLAMA_HOST without this method knowing which.
         """
+        from src.anthropic.model_selector import phase_extra_body
+
+        payload: dict = {
+            "model": self.ollama_model,
+            "messages": [{"role": "user", "content": prompt}],
+            "temperature": 0.3,  # Lower = more consistent
+            "top_p": 0.9,
+        }
+        # Session through the wrapper, id_slot 0 against a bare
+        # llama-server, nothing elsewhere (repo T-4).
+        payload.update(phase_extra_body("steward"))
+
         async with httpx.AsyncClient(timeout=self.timeout) as client:
             response = await client.post(
                 f"{self.ollama_host}/v1/chat/completions",
-                json={
-                    "model": self.ollama_model,
-                    "messages": [{"role": "user", "content": prompt}],
-                    "temperature": 0.3,  # Lower = more consistent
-                    "top_p": 0.9,
-                },
+                json=payload,
             )
 
             response.raise_for_status()
             result = response.json()
+
+            # The wrapper reports on session-carrying responses (its
+            # D-4/D-5); this raw call bypasses the provider hook, so
+            # the body extras are read here. Acting on compaction_due
+            # is a later ticket — the signal must not pass silently.
+            balancing = result.get("balancing")
+            if balancing:
+                logger.info("backend_balancing", session="tatlock-steward", moves=balancing)
+            if result.get("compaction_due"):
+                logger.warning("backend_compaction_due", session="tatlock-steward")
+
             return result["choices"][0]["message"]["content"].strip()
 
     async def analyze(self, query: str, conversation_history: list[dict] | None = None) -> str:

src/agents/tatlock.py

@@ -404,10 +404,13 @@ class TatlockAgent(AgentInterface):
             # with async context managers inside generators
             # The StreamingCoordinator will handle word-by-word streaming
             # Pass message_history to maintain conversation context and tracker for tool logging
+            from src.anthropic.model_selector import with_phase_binding
+
             result = await self.agent.run(
                 user_message,
                 message_history=message_history if message_history else None,
                 deps=tracker,
+                model_settings=with_phase_binding(None, "orchestrate"),
             )
             final_text = result.output
 
@@ -536,13 +539,13 @@ class TatlockAgent(AgentInterface):
 
         # Run with scoped tools and tracker
         # Force tool_choice to make LLM actually call tools
-        from src.anthropic.model_selector import get_tool_choice_settings
+        from src.anthropic.model_selector import get_tool_choice_settings, with_phase_binding
 
         result = await scoped_agent.run(
             enriched_message,
             message_history=pydantic_history if pydantic_history else None,
             deps=tool_tracker,
-            model_settings=get_tool_choice_settings(),
+            model_settings=with_phase_binding(get_tool_choice_settings(), "orchestrate"),
         )
 
         logger.info(
@@ -625,10 +628,13 @@ class TatlockAgent(AgentInterface):
         # Use run() instead of run_stream() to avoid Ollama 400 bug
         # with streaming + tool calls (PydanticAI issues #1292, #2256)
         # We yield the final response in chunks to maintain streaming interface
+        from src.anthropic.model_selector import with_phase_binding
+
         result = await scoped_agent.run(
             enriched_message,
             message_history=pydantic_history if pydantic_history else None,
             deps=tool_tracker,
+            model_settings=with_phase_binding(None, "orchestrate"),
         )
 
         # Stream the final response in chunks to maintain UX
@@ -725,13 +731,13 @@ class TatlockAgent(AgentInterface):
                 pydantic_history.append(ModelResponse(parts=[TextPart(content=content)]))
 
         # Run with scoped tools and tracker
-        from src.anthropic.model_selector import get_tool_choice_settings
+        from src.anthropic.model_selector import get_tool_choice_settings, with_phase_binding
 
         result = await scoped_agent.run(
             enriched_message,
             message_history=pydantic_history if pydantic_history else None,
             deps=tool_tracker,
-            model_settings=get_tool_choice_settings(),
+            model_settings=with_phase_binding(get_tool_choice_settings(), "orchestrate"),
         )
 
         # Extract tool calls and results from the agent's messages
@@ -888,9 +894,12 @@ class TatlockAgent(AgentInterface):
                 pydantic_history.append(ModelResponse(parts=[TextPart(content=content)]))
 
         # Run synthesis
+        from src.anthropic.model_selector import with_phase_binding
+
         result = await synthesis_agent.run(
             synthesis_prompt,
             message_history=pydantic_history if pydantic_history else None,
+            model_settings=with_phase_binding(None, "synthesize"),
         )
 
         logger.info(

src/anthropic/model_selector.py

@@ -28,6 +28,13 @@ logger = get_logger(__name__)
 # Cached health check results (set once at startup)
 _claude_available: bool | None = None
 _ollama_available: bool | None = None
+# Which server answers behind OLLAMA_HOST: "boilerroom", "llama-server"
+# or "ollama". They disagree on tool_choice semantics and on which
+# extension fields do anything (see get_tool_choice_settings and
+# phase_extra_body), so the health check probes for the wrapper's name
+# on /health, then for /props — served by llama-server only — and
+# caches the answer here.
+_local_flavor: str | None = None
 
 
 async def check_ollama_health() -> bool:
@@ -40,7 +47,7 @@ async def check_ollama_health() -> bool:
     Returns:
         True if Ollama is reachable and OLLAMA_DEFAULT_MODEL is pulled.
     """
-    global _ollama_available
+    global _ollama_available, _local_flavor
 
     host = str(config.OLLAMA_HOST).rstrip("/")
     model = config.OLLAMA_DEFAULT_MODEL
@@ -53,12 +60,34 @@ async def check_ollama_health() -> bool:
             response.raise_for_status()
             names = [m.get("id", "") for m in response.json().get("data", [])]
 
+            # The boilerroom wrapper names itself on /health; a bare
+            # llama-server answers /health without a service field and
+            # serves /props (its own surface); Ollama answers neither.
+            is_wrapper = False
+            try:
+                health = await client.get(f"{host}/health")
+                is_wrapper = (
+                    health.status_code == 200 and health.json().get("service") == "boilerroom"
+                )
+            except (httpx.HTTPError, ValueError):
+                is_wrapper = False
+
+            if is_wrapper:
+                _local_flavor = "boilerroom"
+            else:
+                try:
+                    props = await client.get(f"{host}/props")
+                    _local_flavor = "llama-server" if props.status_code == 200 else "ollama"
+                except httpx.HTTPError:
+                    _local_flavor = "ollama"
+
         if model in names or f"{model}:latest" in names:
             _ollama_available = True
             logger.info(
                 "ollama_health_check_passed",
                 host=host,
                 model=model,
+                flavor=_local_flavor,
             )
             return True
 
@@ -244,19 +273,97 @@ def get_model(prefer_cloud: bool | None = None) -> AnthropicModel | OpenAIChatMo
 
 def get_tool_choice_settings() -> ModelSettings:
     """
-    Get model_settings for forcing tool calls on the first request.
+    Get model_settings for tool calling on the orchestration phase.
 
-    For Claude: PydanticAI handles tool_choice natively, so no extra_body needed.
-    For Ollama: Pass tool_choice="required" via extra_body to force tool calling.
+    Claude: PydanticAI handles tool_choice natively — no extra_body.
+
+    Ollama: tool_choice="required" via extra_body. Advisory there (Ollama
+    ignores it), but it nudges gemma4 to actually call tools, which the
+    persona-suppression gotcha made necessary.
+
+    llama-server: NO tool_choice. It enforces "required" on every request
+    in the run, so after a tool returns, the next generation is again
+    forced to call a tool — an unbreakable tool loop (~80 s turns,
+    observed at cutover). Its --jinja template renders tool definitions
+    the way gemma4 was trained, and the model calls tools reliably
+    unforced (probed 3/3).
+
+    boilerroom: same as llama-server — the wrapper forwards to it, and
+    its /props passthrough is what lets this detection keep working.
     """
     from pydantic_ai.settings import ModelSettings
 
     if resolve_backend() == "claude":
         # PydanticAI's Anthropic model handles tool_choice internally
         return ModelSettings()
-    else:
-        # Ollama needs explicit tool_choice via extra_body
-        return ModelSettings(extra_body={"tool_choice": "required"})
+    if _local_flavor in ("llama-server", "boilerroom"):
+        return ModelSettings()
+    # Ollama needs explicit tool_choice via extra_body
+    return ModelSettings(extra_body={"tool_choice": "required"})
+
+
+# Phase bindings for the local backend (this repo's T-4, decided
+# 2026-09-12): through the boilerroom wrapper each phase is a named
+# session with an eviction rank — lower parks sooner, so the most-used
+# context keeps its KV warm. The pipeline phases outrank the librarian,
+# which outranks webber (20, filed in its own repo); Open WebUI stays
+# session-less and can never evict anyone (wrapper D-4).
+# phase -> (direct llama-server slot, wrapper session name, eviction rank)
+_PHASE_BINDINGS: dict[str, tuple[int | None, str, int]] = {
+    "steward": (0, "tatlock-steward", 40),
+    "orchestrate": (1, "tatlock-orchestrate", 40),
+    "synthesize": (2, "tatlock-synthesize", 40),
+    # Experts (biographer, housekeeper) share one session between the
+    # librarian and the phases: the two least-used consumers trade the
+    # spare slot by rank instead of anyone hitting an empty pool — a
+    # session-less call 503s once four sessions are resident (D-4's
+    # idle-only rule), so every generation call here must carry a name.
+    "experts": (None, "tatlock-experts", 35),
+    "librarian": (None, "librarian", 30),
+}
+
+
+def phase_extra_body(phase: str) -> dict[str, str | int]:
+    """
+    The extension fields a pipeline phase sends to the local backend.
+
+    Through the boilerroom wrapper: a named session with its eviction
+    rank — the wrapper owns the name-to-slot map and reports parks and
+    restores on the response. Against a bare llama-server: the raw
+    id_slot pin this replaces (the librarian floats there, as it always
+    did). Ollama and Claude get nothing, and an unprobed flavor sends
+    nothing rather than guessing. Gated by BACKEND_SLOT_PINNING like
+    the pinning it grew out of.
+    """
+    if not config.BACKEND_SLOT_PINNING or resolve_backend() == "claude":
+        return {}
+    slot, session, rank = _PHASE_BINDINGS[phase]
+    if _local_flavor == "boilerroom":
+        return {"session": session, "eviction_order": rank}
+    if _local_flavor == "llama-server" and slot is not None:
+        return {"id_slot": slot}
+    return {}
+
+
+def with_phase_binding(settings: ModelSettings | None, phase: str) -> ModelSettings | None:
+    """
+    Merge a phase's local-backend binding into model settings.
+
+    Session fields through the wrapper, id_slot against a bare
+    llama-server, settings untouched everywhere else — one call site
+    stays correct on any backend (see phase_extra_body).
+    """
+    extra = phase_extra_body(phase)
+    if not extra:
+        return settings
+
+    from pydantic_ai.settings import ModelSettings
+
+    merged = dict(settings or {})
+    extra_body = dict(merged.get("extra_body") or {})
+    extra_body.update(extra)
+    merged["extra_body"] = extra_body
+    return ModelSettings(**merged)
 
 
 def get_sampling_settings(temperature: float) -> ModelSettings:
@@ -291,5 +398,6 @@ def get_model_info() -> dict:
         "claude_configured": bool(config.ANTHROPIC_API_KEY),
         "ollama_available": is_ollama_available(),
         "ollama_model": config.OLLAMA_DEFAULT_MODEL,
+        "local_flavor": _local_flavor,
         "prefer_cloud": config.PREFER_CLOUD_BACKEND,
     }

src/core/config.py

@@ -147,6 +147,10 @@ class Config(BaseSettings):
         default=None,
         description="OpenAI-compatible embeddings host; falls back to OLLAMA_HOST so gen and embed can live on different servers",
     )
+    BACKEND_SLOT_PINNING: bool = Field(
+        default=False,
+        description="Pin each pipeline phase to a llama-server slot (steward=0, orchestrator=1, synthesizer=2) so a phase's stable prompt prefix stays in that slot's KV cache; no-op on Claude, ignored by Ollama",
+    )
 
     # Redis Memory Database
     REDIS_MEMORY_DB: int = Field(default=1, description="Redis database number for memory cache")

src/ollama/provider.py

@@ -8,6 +8,7 @@ This provider sanitizes messages to use empty strings instead of null.
 
 from typing import Any
 
+import httpx
 from openai import AsyncOpenAI
 from pydantic_ai.providers.ollama import OllamaProvider
 
@@ -17,6 +18,27 @@ from src.core.logging_config import get_logger
 logger = get_logger(__name__)
 
 
+async def _log_wrapper_signals(response: httpx.Response) -> None:
+    """
+    Surface the boilerroom wrapper's per-response signals (repo T-4).
+
+    balancing names the sessions parked or restored to serve this
+    request (wrapper D-4 — never silent); compaction-due says a session
+    crossed its authored context budget (wrapper D-5 — a signal, not an
+    interrupt). Reading them is this ticket; producing the compacted
+    transcript is a later one. The headers are absent on any other
+    backend, so the hook costs nothing there.
+    """
+    balancing = response.headers.get("x-boilerroom-balancing")
+    if balancing and balancing != "[]":
+        logger.info("backend_balancing", moves=balancing)
+    if response.headers.get("x-boilerroom-compaction-due") == "true":
+        logger.warning(
+            "backend_compaction_due",
+            hint="GET /sessions on the wrapper names the session over budget",
+        )
+
+
 class TatlockOllamaProvider(OllamaProvider):
     """
     Custom OllamaProvider with message sanitization for Tatlock agents.
@@ -56,6 +78,15 @@ class _SanitizedAsyncOpenAI(AsyncOpenAI):
         # configured Ollama timeout instead of the SDK default (~600s),
         # so one stuck request cannot eat the whole delegation budget.
         kwargs.setdefault("timeout", float(config.OLLAMA_TIMEOUT))
+        # The response hook reads the wrapper's balancing/compaction
+        # headers on every call this client makes, streams included.
+        kwargs.setdefault(
+            "http_client",
+            httpx.AsyncClient(
+                event_hooks={"response": [_log_wrapper_signals]},
+                timeout=float(config.OLLAMA_TIMEOUT),
+            ),
+        )
         super().__init__(api_key="ollama", **kwargs)
 
     @property

tests/anthropic/test_model_selector.py

@@ -67,7 +67,10 @@ class TestGetModel:
 
 
 class TestToolChoiceSettings:
-    def test_ollama_forces_tool_choice(self, local_first):
+    def test_ollama_forces_tool_choice(self, local_first, monkeypatch):
+        # Pin the flavor: the session-scoped app init probes the real
+        # dev backend, and this test's answer must not depend on it.
+        monkeypatch.setattr(model_selector, "_local_flavor", "ollama")
         settings = model_selector.get_tool_choice_settings()
         assert settings.get("extra_body") == {"tool_choice": "required"}
 
@@ -91,3 +94,114 @@ class TestGetModelInfo:
         info = model_selector.get_model_info()
         assert info["backend"] == "claude"
         assert info["model"] == config.ANTHROPIC_MODEL
+
+
+class TestPhaseBinding:
+    def test_disabled_returns_settings_unchanged(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", False)
+        base = model_selector.get_tool_choice_settings()
+        assert model_selector.with_phase_binding(base, "orchestrate") is base
+        assert model_selector.with_phase_binding(None, "synthesize") is None
+
+    def test_wrapper_flavor_sends_session_and_rank(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(model_selector, "_local_flavor", "boilerroom")
+        settings = model_selector.with_phase_binding(
+            model_selector.get_tool_choice_settings(), "orchestrate"
+        )
+        extra_body = settings["extra_body"]
+        assert extra_body["session"] == "tatlock-orchestrate"
+        assert extra_body["eviction_order"] == 40
+        assert "id_slot" not in extra_body
+
+    def test_wrapper_ranking_matches_the_decision(self, local_first, monkeypatch):
+        # Decided 2026-09-12: pipeline over librarian over webber (20,
+        # filed in its own repo); lower eviction_order parks sooner.
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(model_selector, "_local_flavor", "boilerroom")
+        ranks = {
+            phase: model_selector.phase_extra_body(phase)["eviction_order"]
+            for phase in ("steward", "orchestrate", "synthesize", "experts", "librarian")
+        }
+        assert ranks["steward"] == ranks["orchestrate"] == ranks["synthesize"] == 40
+        assert ranks["experts"] == 35
+        assert ranks["librarian"] == 30
+        assert ranks["librarian"] < ranks["experts"] < ranks["steward"]
+        assert ranks["librarian"] > 20  # webber's rank stays below
+
+    def test_wrapper_session_names_are_stable(self, local_first, monkeypatch):
+        # The wrapper's map keys on these; renaming one orphans its slot.
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(model_selector, "_local_flavor", "boilerroom")
+        names = {
+            phase: model_selector.phase_extra_body(phase)["session"]
+            for phase in ("steward", "orchestrate", "synthesize", "experts", "librarian")
+        }
+        assert names == {
+            "steward": "tatlock-steward",
+            "orchestrate": "tatlock-orchestrate",
+            "synthesize": "tatlock-synthesize",
+            "experts": "tatlock-experts",
+            "librarian": "librarian",
+        }
+
+    def test_direct_llama_server_keeps_id_slot(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(model_selector, "_local_flavor", "llama-server")
+        settings = model_selector.with_phase_binding(None, "synthesize")
+        assert settings["extra_body"] == {"id_slot": 2}
+
+    def test_librarian_floats_on_direct_llama_server(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(model_selector, "_local_flavor", "llama-server")
+        assert model_selector.phase_extra_body("librarian") == {}
+
+    def test_ollama_flavor_gets_no_fields(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(model_selector, "_local_flavor", "ollama")
+        assert model_selector.phase_extra_body("steward") == {}
+
+    def test_unprobed_flavor_sends_nothing_rather_than_guessing(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(model_selector, "_local_flavor", None)
+        assert model_selector.phase_extra_body("orchestrate") == {}
+
+    def test_merge_preserves_existing_extra_body(self, local_first, monkeypatch):
+        from pydantic_ai.settings import ModelSettings
+
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(model_selector, "_local_flavor", "boilerroom")
+        base = ModelSettings(extra_body={"marker": 1})
+        merged = model_selector.with_phase_binding(base, "steward")
+        assert merged["extra_body"]["marker"] == 1
+        assert merged["extra_body"]["session"] == "tatlock-steward"
+
+    def test_claude_backend_never_bound(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(config, "PREFER_CLOUD_BACKEND", True)
+        base = model_selector.get_tool_choice_settings()
+        assert model_selector.with_phase_binding(base, "orchestrate") is base
+
+
+class TestLocalFlavorToolChoice:
+    def test_llama_server_flavor_sends_no_tool_choice(self, local_first, monkeypatch):
+        monkeypatch.setattr(model_selector, "_local_flavor", "llama-server")
+        settings = model_selector.get_tool_choice_settings()
+        assert not settings.get("extra_body")
+
+    def test_boilerroom_flavor_sends_no_tool_choice(self, local_first, monkeypatch):
+        # The wrapper forwards to llama-server, which enforces
+        # tool_choice — the nudge through it is the tool-loop incident.
+        monkeypatch.setattr(model_selector, "_local_flavor", "boilerroom")
+        settings = model_selector.get_tool_choice_settings()
+        assert not settings.get("extra_body")
+
+    def test_ollama_flavor_keeps_advisory_required(self, local_first, monkeypatch):
+        monkeypatch.setattr(model_selector, "_local_flavor", "ollama")
+        settings = model_selector.get_tool_choice_settings()
+        assert settings["extra_body"] == {"tool_choice": "required"}
+
+    def test_unknown_flavor_defaults_to_ollama_semantics(self, local_first, monkeypatch):
+        monkeypatch.setattr(model_selector, "_local_flavor", None)
+        settings = model_selector.get_tool_choice_settings()
+        assert settings["extra_body"] == {"tool_choice": "required"}

tests/contracts/test_service_contracts.py

@@ -114,16 +114,18 @@ class TestLocalBackendContract:
         )
 
     async def test_openai_compat_tool_calling(self):
-        # Mirrors the request PydanticAI's OpenAIChatModel sends for the
-        # orchestration phase, including the extra_body tool_choice.
+        # Mirrors the orchestration-phase request on llama-server: tools
+        # attached, NO tool_choice. The backend must call the tool unforced
+        # — "required" is deliberately absent because llama-server enforces
+        # it on every request in a run, which turns the tool loop
+        # unbreakable (observed at cutover: ~80 s turns).
         response = await _post_or_skip(
             f"{OLLAMA}/v1/chat/completions",
-            "ollama",
+            "local-backend",
             {
                 "model": config.OLLAMA_DEFAULT_MODEL,
                 "messages": [{"role": "user", "content": "What is 6 * 7? Use the calculator."}],
                 "tools": [CALCULATOR_TOOL],
-                "tool_choice": "required",
                 "stream": False,
             },
             timeout=config.OLLAMA_TIMEOUT,
@@ -252,3 +254,62 @@ class TestRedisContract:
             await client.delete("contract-test-key")
         finally:
             await client.aclose()
+
+
+@pytest.mark.contract
+class TestBoilerroomWrapperContract:
+    """Boundary: the boilerroom wrapper surface session adoption rides on
+    (repo T-4). Skips when OLLAMA_HOST is not the wrapper — a direct
+    engine or Ollama is a valid deployment, not a violation.
+    """
+
+    async def _skip_unless_wrapper(self) -> None:
+        response = await _get_or_skip(f"{OLLAMA}/health", "local-backend")
+        try:
+            service = response.json().get("service")
+        except ValueError:
+            service = None
+        if service != "boilerroom":
+            pytest.skip("OLLAMA_HOST is not the boilerroom wrapper")
+
+    async def test_wrapper_names_itself_on_health(self):
+        # Mirrors check_ollama_health()'s wrapper probe: the flavor
+        # detection keys on this exact field.
+        await self._skip_unless_wrapper()
+
+    async def test_props_passes_through_for_flavor_detection(self):
+        # tatlock treats a non-200 /props as Ollama and applies the
+        # tool_choice nudge — enforced by llama-server into a tool
+        # loop. The wrapper passing /props through is what keeps that
+        # detection safe (boilerroom T-10).
+        await self._skip_unless_wrapper()
+        response = await _get_or_skip(f"{OLLAMA}/props", "local-backend")
+        assert response.status_code == 200
+
+    async def test_session_fields_answer_with_signals(self):
+        # Mirrors phase_extra_body() through the wrapper: a session-
+        # carrying request must come back with the balancing header and
+        # the compaction signal in the body (wrapper D-3/D-4/D-5).
+        await self._skip_unless_wrapper()
+        response = await _post_or_skip(
+            f"{OLLAMA}/v1/chat/completions",
+            "local-backend",
+            {
+                "model": config.OLLAMA_DEFAULT_MODEL,
+                # No max_tokens: gemma4 spends its first tokens thinking,
+                # and a tight cap returns empty content (not a wrapper
+                # fault — the sibling passthrough test runs uncapped too).
+                "messages": [{"role": "user", "content": "Reply with the single word: pong"}],
+                "session": "contract-probe",
+                "eviction_order": 1,
+            },
+            timeout=config.OLLAMA_TIMEOUT,
+        )
+        assert response.status_code == 200
+        assert "x-boilerroom-balancing" in response.headers
+        assert response.headers.get("x-boilerroom-compaction-due") in ("true", "false")
+        body = response.json()
+        assert isinstance(body.get("compaction_due"), bool)
+        assert isinstance(body.get("balancing"), list)
+        # The extension fields must not leak into the engine's answer
+        assert body["choices"][0]["message"]["content"].strip()