feat: janitor role — session checkpoint compaction

New cortex/janitor.py runs before each orchestrator dispatch. When a session
exceeds 20 user turns or ~12K estimated tokens, the oldest half is summarized
by the janitor role model and replaced with a compact checkpoint message.
Fail-safe: always returns original history if the model call fails.

Config: JANITOR_TURN_THRESHOLD, JANITOR_TOKEN_THRESHOLD in .env.
Assign Gemma E4B or Haiku 4.5 to the janitor role for effectively-free compaction.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cortex/config.py

@@ -71,13 +71,20 @@ class Settings(BaseSettings):
     role_chat: str = "claude_cli"
     role_orchestrator: str = "gemini_api"
     role_distill: str = "claude_cli"
+    role_janitor: str = "claude_cli"   # assign a cheap/fast model: Haiku 4.5, local Gemma E4B
     role_coder: str = "claude_cli"
     role_research: str = "gemini_api"
 
     # Comma-separated list of standard roles shown in the model settings UI.
     # Add custom roles here to extend the UI without code changes.
-    # Example: DEFINED_ROLES=chat,orchestrator,distill,coder,research,medical
-    defined_roles: str = "chat,orchestrator,distill,coder,research"
+    # Example: DEFINED_ROLES=chat,orchestrator,distill,janitor,coder,research,medical
+    defined_roles: str = "chat,orchestrator,distill,janitor,coder,research"
+
+    # Session checkpoint compaction ("janitor") thresholds.
+    # Compaction fires when EITHER threshold is exceeded.
+    # Override in .env: JANITOR_TURN_THRESHOLD=15  JANITOR_TOKEN_THRESHOLD=8000
+    janitor_turn_threshold: int = 20    # user turns (each turn = 1 user + 1 assistant message)
+    janitor_token_threshold: int = 12000  # estimated tokens (chars / 4 heuristic)
 
     # Memory tier token budgets — soft caps used during distillation
     # Override in .env: MEMORY_BUDGET_LONG=4000 etc.

cortex/janitor.py

@@ -0,0 +1,117 @@
+"""
+Session checkpoint compaction ("janitor").
+
+Called before each orchestrator run. When a session exceeds the configured turn
+or token threshold, the oldest half of the history is summarized by the janitor
+role model and replaced with a compact checkpoint message. This keeps the token
+count passed to the orchestrator lean while preserving a faithful record of what
+happened earlier in the session.
+
+The janitor role should be assigned a cheap, fast model — a small local model
+(Gemma E4B) or a lightweight cloud model (Haiku 4.5). It has no tools and the
+task is simple enough that quality matters less than speed and cost.
+
+Thresholds (configurable in .env):
+  JANITOR_TURN_THRESHOLD  — compact after N user turns  (default: 20)
+  JANITOR_TOKEN_THRESHOLD — compact after ~N estimated tokens (default: 12000)
+"""
+
+import logging
+
+from config import settings
+
+logger = logging.getLogger(__name__)
+
+_SYSTEM = "You are a concise summarizer. Write only the summary — no preamble, no labels."
+
+_PROMPT_TMPL = """\
+Summarize the conversation below in 3–8 sentences. Capture what was discussed, \
+any decisions or conclusions reached, and key specifics (names, values, file paths, etc.). \
+Write only the summary paragraph.
+
+CONVERSATION:
+{conversation}"""
+
+
+def _format_messages(messages: list[dict]) -> str:
+    lines = []
+    for m in messages:
+        role = m.get("role", "unknown").upper()
+        content = (m.get("content") or "").strip()
+        if not content:
+            continue
+        # Cap individual messages so the prompt stays manageable for small models
+        if len(content) > 600:
+            content = content[:600] + "…"
+        lines.append(f"[{role}]: {content}")
+    return "\n".join(lines)
+
+
+async def maybe_checkpoint(session_id: str) -> list[dict]:
+    """
+    Load the session, compact if thresholds are exceeded, and return the
+    message list to use for the upcoming orchestrator run.
+
+    Always returns a list — returns the original (unchanged) list if:
+      - the session does not exist yet
+      - thresholds are not met
+      - the janitor model call fails (fail-safe: never discard history)
+    """
+    from session_store import load, save
+
+    messages = load(session_id)
+    if not messages:
+        return []
+
+    turn_count = sum(1 for m in messages if m["role"] == "user")
+    estimated_tokens = sum(len(m.get("content") or "") for m in messages) // 4
+
+    if (turn_count < settings.janitor_turn_threshold
+            and estimated_tokens < settings.janitor_token_threshold):
+        return messages
+
+    # Walk back to a clean turn boundary so we never split mid-exchange.
+    # midpoint lands on an "assistant" message boundary.
+    midpoint = len(messages) // 2
+    while midpoint > 0 and messages[midpoint - 1].get("role") != "assistant":
+        midpoint -= 1
+
+    if midpoint < 4:
+        # Too short to compact meaningfully — threshold likely set very low
+        return messages
+
+    old_messages = messages[:midpoint]
+    recent_messages = messages[midpoint:]
+
+    conversation_text = _format_messages(old_messages)
+    summary_prompt = _PROMPT_TMPL.format(conversation=conversation_text)
+
+    try:
+        from llm_client import complete as llm_complete
+        summary, backend = await llm_complete(
+            system_prompt=_SYSTEM,
+            messages=[{"role": "user", "content": summary_prompt}],
+            role="janitor",
+        )
+
+        checkpoint_msg = {
+            "role": "assistant",
+            "content": (
+                f"[Session checkpoint — {len(old_messages)} messages summarized "
+                f"via {backend}]\n\n{summary.strip()}"
+            ),
+        }
+
+        compacted = [checkpoint_msg] + recent_messages
+        save(session_id, compacted)
+
+        logger.info(
+            "Janitor: session=%s compacted %d→%d messages (turns=%d ~%d tokens) via %s",
+            session_id, len(messages), len(compacted), turn_count, estimated_tokens, backend,
+        )
+        return compacted
+
+    except Exception as exc:
+        # Fail-safe: never lose history because the janitor model is unavailable
+        logger.warning("Janitor skipped for session %s: %s", session_id, exc)
+        return messages

cortex/routers/orchestrator.py

@@ -257,6 +257,7 @@ async def _run_job(job_id: str, req: OrchestrateRequest, user: str) -> None:
 
     try:
         from session_store import load as load_session, save as save_session, generate_session_id
+        from janitor import maybe_checkpoint as janitor_checkpoint
 
         tier = req.tier or settings.default_tier
         role_cfg = model_registry.get_role_config(user, req.chat_role)
@@ -272,7 +273,8 @@ async def _run_job(job_id: str, req: OrchestrateRequest, user: str) -> None:
         )
 
         session_id = req.session_id or generate_session_id()
-        history = load_session(session_id)
+        # Compact old session turns before dispatching — no-op on new sessions or short ones.
+        history = await janitor_checkpoint(session_id) if req.session_id else load_session(session_id)
         session_messages = history or None
 
         orch_model = model_registry.get_model_for_role(user, "orchestrator")

documentation/TODO__Agents.md

@@ -44,7 +44,7 @@ automatically. Remaining work is quality/reliability parity, not ground-up desig
 - [x] Retry logic on transient API errors (connection timeout, 429, 503) — 2026-05-09
   - `_chat_with_retry()` helper in `openai_orchestrator.py`; 3 attempts, exponential backoff (1s, 2s)
   - Retries on `APIConnectionError` and `APIStatusError` with status 429/500/502/503/504
-- [ ] Test end-to-end with Gemma 4 E4B and 26B A4B on scott_gaming
+- [x] Test end-to-end with Gemma 4 E4B and 26B A4B on scott_gaming — 2026-06-17
 - [ ] Review `ARCH__FUTURE.md` agent architecture ideas before finalising design
 - Reference: `docs/OPEN_WEBUI_API.md`, `documentation/ARCH__FUTURE.md` §1
 
@@ -211,43 +211,28 @@ Upload an image or document inline and have it flow into context.
 - [x] Text/code files read as UTF-8, injected as fenced code block in message
 - [x] Thumbnail/filename shown above sent message in UI
 
-### [Intelligence] Session checkpoint compaction — "janitor" role
-Proactive in-session context pruning using a cheap/fast model to keep expensive
-model costs down as sessions grow. Not continuous per-token — checkpoint-triggered.
+### [Intelligence] Session checkpoint compaction — "janitor" role ✅ — 2026-06-17
+Proactive in-session context pruning using a cheap/fast model. Fires before each
+orchestrator run; compacts oldest half of history when either threshold is exceeded.
 
-**Design:**
-- New `janitor` role in the model registry (alongside `chat`, `orchestrator`, `distill`)
-  - Assign a cheap/fast model: Haiku 4.5, local Gemma E4B, or similar
-  - Falls back to the `distill` role model if `janitor` is not configured
-- Trigger condition (either/or): session exceeds N turns (e.g. 20) OR estimated token
-  count exceeds a threshold (e.g. 12K tokens of history)
-- On trigger: call janitor model with the oldest half of session history; ask it to
-  write a compact "what we've established so far" summary block (3–8 sentences)
-- Replace the compacted turns with a single synthetic `assistant` message:
-  `[Session checkpoint — {N} turns summarized]: {summary}`
-- The remaining recent turns stay untouched — only the stale prefix is replaced
-- Token estimate: count chars / 4 as a cheap heuristic; no exact tokenizer needed
+- [x] **`cortex/janitor.py`** — `maybe_checkpoint(session_id)` — loads session,
+      checks `janitor_turn_threshold` (default 20) and `janitor_token_threshold`
+      (default 12000 estimated tokens); finds a clean turn boundary; calls janitor
+      role model with the oldest half; replaces compacted messages with a single
+      `[Session checkpoint — N messages summarized via {backend}]` assistant message;
+      fail-safe returns original messages if model call fails — 2026-06-17
+- [x] **`cortex/config.py`** — `janitor_turn_threshold`, `janitor_token_threshold`,
+      `role_janitor` settings; `janitor` added to `defined_roles` — 2026-06-17
+- [x] **`cortex/routers/orchestrator.py`** — calls `janitor_checkpoint(session_id)`
+      before dispatching to either orchestrator engine; no-op on new sessions — 2026-06-17
+- [x] **`model_registry.py`** — `janitor` already in `REQUIRED_ROLES`,
+      `ROLE_DEFAULT_TOOLS` (no tools), and `_ROLE_LAST_RESORT` from earlier session
 
-**Files to change:**
-- `model_registry.py` — add `janitor` to `ROLE_DEFAULT_TOOLS` (empty list — no tools)
-  and to the roles UI in `settings/models`
-- `session_store.py` — add `maybe_checkpoint(session_id)` that checks turn count /
-  estimated tokens and calls the janitor model if threshold is exceeded
-- `openai_orchestrator.py` — call `maybe_checkpoint()` at the start of each run,
-  before building the active tool list and context
-- `orchestrator_engine.py` — same, before building the Gemini context
-- Settings UI — expose janitor turn/token thresholds as configurable values
-  (default: 20 turns or 12K history tokens)
+**To configure:** assign Gemma E4B (local, free) or Haiku 4.5 to the `janitor` role
+in Settings → Model Registry. Thresholds overridable in `.env`:
+`JANITOR_TURN_THRESHOLD=15  JANITOR_TOKEN_THRESHOLD=8000`
 
-**Economics:**
-- Haiku 4.5: ~$0.80/1M input — compacting 10K tokens costs ~$0.008
-- Saves 8–12K tokens on every subsequent Sonnet/Opus call in that session
-- Break-even after 1–2 expensive model calls post-checkpoint
-- Local janitor (Gemma E4B) = effectively free; ideal default when available
-
-**Not needed yet** — most sessions are short enough that existing `_compact_messages()`
-heuristic handles the worst cases. Priority rises with dev-agent pipeline work where
-aider tool results can be very large.
+**Deferred:** Settings UI sliders for thresholds (low value — .env is sufficient)
 
 ### [UX] Token streaming for orchestrator final response ✅ — 2026-06-16
 Text appears token-by-token while the model is generating, instead of waiting for the