AGENTS.md

@@ -35,7 +35,7 @@ code.
 | `events.py` | In-memory decision-event broker | Pure stdlib (`queue`, `collections.deque`). Thread-safe. No Textual import. The SSE endpoint lives in `dispatcher.py` and calls `events.subscribe()`/`publish_decision()`. |
 | `config.py` | Pydantic models + YAML loader | All config models inherit `StrictModel` (`extra="forbid"`). Unknown keys fail at load. |
 | `capabilities.py` | Request-side capability detection (`tools`, `images`, `json_mode`, `reasoning`) | Reads from the OpenAI-format request body, not from the classifier. |
-| `context_prune.py` | Relevance-based context pruning (`pinch`) | Ships **disabled** (`pinch.enabled: false`). Imports `PinchConfig` from `config.py` (no circular import: `config.py` doesn't import `context_prune`). |
+| `context_prune.py` | Relevance-based context pruning (`pinch`) | Ships **enabled** by default (`pinch.enabled: true`; `pinch.relevance.enabled: true`; budget-gated: the embed only fires when the conversation exceeds `budget_tokens`). Imports `PinchConfig` from `config.py` (no circular import: `config.py` doesn't import `context_prune`). |
 | `logs.py` | Structured logging (logfmt, journald, ContextVar trace ids) | `logs.bind()` exists for StreamingResponse generators that lose the ContextVar. |
 
 ### Pure scoring/routing modules (no I/O)
@@ -161,6 +161,47 @@ PYTHONPATH=src python -m router_cli "Refactor this Django view"
 PYTHONPATH=src python -m poller && PYTHONPATH=src python -m tier
 ```
 
+## Gitea / `tea` CLI (not `gh`)
+
+This repo is hosted on **self-hosted Gitea** (`git.adlee.work/alee/6krrt`).
+The GitHub CLI `gh` is **not installed** and will fail — use `tea` instead.
+
+- **Login:** `tea login` as `alee`. The `alee` login is NOT the default tea
+  login — always pass `--remote origin` so tea resolves the right context.
+
+- **Remote:** `origin` resolves to `git.adlee.work/alee/6krrt`.
+
+- **List open PRs:** `tea pulls --remote origin`
+
+- **PR metadata (JSON):**
+  ```
+  tea pulls <PR_INDEX> --remote origin --fields index,state,draft,title,mergeable,base,head,body -o json
+  ```
+  Use the `base` and `head` from this output for local diff commands.
+
+- **PR creation / update / maintenance:** use the `tea pr` family (e.g. `tea pr create --base main --title ... --description ...`). Confirm against `tea pr --help` for the current flag set.
+
+- **Post a comment to a PR** (uses the `tea api` proxy to Gitea's REST API):
+  ```
+  tea api --remote origin "/repos/{owner}/{repo}/issues/<PR_INDEX>/comments" -F body=@-
+  ```
+  - **CRITICAL:** use `-F` (typed field), **not** `-f`. `-f key=@file` writes the literal string `@file`; only `-F` reads stdin or the contents of a file.
+  - The endpoint **must** include the `/repos/` prefix — omitting it returns a 404.
+  - `tea api` already prints JSON to stdout — do **not** pass `-o json` on top of that (you would write the body to a file literally named `json`).
+  - Fallback if stdin is awkward: `-F body=@/tmp/review-body.md`.
+
+- **Getting a PR diff:** prefer local `git diff origin/<base>...origin/<head>`
+  (base and head from the PR JSON). In tea v0.14.1 the `diff`/`patch` fields
+  return empty even when requested — do not rely on them.
+
+- **Gitea source links** (not GitHub `blob` links):
+  `https://git.adlee.work/alee/6krrt/src/commit/<FULL-SHA>/<path>#L<start>-L<end>`
+  (segment is `src/commit`, not `blob`).
+
+- **Code-review plugin:** this repo ships a project-local fork of the code-review
+  plugin (`code-review-tea`) that already wires these `tea` calls. Prefer
+  invoking it (e.g. `/code-review-tea:code-review`) over hand-rolling.
+
 ## What's NOT built yet (open items)
 
 1. **Leaderboard priors unfilled** — `leaderboards.yaml` ships empty.

config/config.yaml

@@ -166,12 +166,11 @@ pinch:
   # they carry the bulk of a long agent session's tokens and are least needed
   # in full by the time the next turn is answered.
   #
-  # It does NOT touch the classifier's input, and it defaults off — enable it
-  # only if long sessions are shipping more prompt tokens than you want to pay
-  # for. Tool results can only be dropped safely because tool outputs are
-  # idempotent enough for a placeholder; a wrong guess here loses context, so
-  # start conservative (large budget, small reduction).
-  enabled: false
+  # On by default. Tool results can only be dropped safely because tool
+  # outputs are idempotent enough for a placeholder; a wrong guess here loses
+  # context, so start conservative (large budget, small reduction) and watch
+  # route_decisions / pinch stats on real traffic before widening it.
+  enabled: true
   budget_tokens: 50000
   keep_last_turns: 4
   max_summarize_chars: 4000
@@ -180,7 +179,7 @@ pinch:
     # project — and specifically requires pinch.enabled too, since this has no
     # effect otherwise. Ship it, watch route_decisions / pinch stats on real
     # traffic, then decide the default.
-    enabled: false
+    enabled: true
     # An EMBEDDING model, not a chat model — this must not point at
     # classifier.model or verification.model. Pull one on the same Ollama:
     #   ollama pull nomic-embed-text
@@ -207,13 +206,11 @@ session_cache:
   staleness_minutes: 20
 
 circuit_breaker:
-  # Passive availability circuit breaker. Off by default, matching every other
-  # new-and-unproven knob in this project. When enabled, a model that returns
-  # 5xx is temporarily excluded from routing with exponential backoff; recovery
-  # is passive (the next real request becomes the probe once the cooldown
-  # passes). Unlike most new knobs this one has a low-risk failure mode even
-  # when wrong, so it's a reasonable candidate to flip on sooner.
-  enabled: false
+  # Passive availability circuit breaker, on by default. When enabled, a model
+  # that returns 5xx is temporarily excluded from routing with exponential
+  # backoff; recovery is passive (the next real request becomes the probe once
+  # the cooldown passes). It has a low-risk failure mode even when wrong.
+  enabled: true
   initial_cooldown_seconds: 30
   max_cooldown_seconds: 600
   backoff_multiplier: 2.0

src/config.py

@@ -306,12 +306,12 @@ class PinchRelevanceConfig(StrictModel):
 
     When enabled (and ``pinch.enabled`` is also true), the dispatcher embeds
     the current-turn query with the old tool-result candidates and trims the
-    least relevant first, so a relevant-but-old result survives. Off by
+    least relevant first, so a relevant-but-old result survives. On by
     default; any failure reverts to uniform trimming. This must point at an
     EMBEDDING model, never ``classifier.model`` or ``verification.model``.
     """
 
-    enabled: bool = False
+    enabled: bool = True
     # OpenAI-compatible embeddings endpoint on the same local Ollama.
     model: str = "nomic-embed-text"
     base_url: str = "http://localhost:11434/v1"
@@ -343,7 +343,7 @@ class PinchConfig(StrictModel):
     are summarized or dropped (they carry the bulk of a long session's tokens).
     """
 
-    enabled: bool = False
+    enabled: bool = True
     budget_tokens: int = 50000
     # How many recent user turns (plus their assistant replies and tool results)
     # are protected from pruning.
@@ -406,10 +406,10 @@ class CircuitBreakerConfig(StrictModel):
 
     When enabled, a model that returns 5xx is temporarily skipped by routing
     (with exponential backoff). Recovery is passive: a real request that would
-    have picked it becomes the probe once the cooldown passes. Off by default.
+    have picked it becomes the probe once the cooldown passes. On by default.
     """
 
-    enabled: bool = False
+    enabled: bool = True
     initial_cooldown_seconds: int = 30
     max_cooldown_seconds: int = 600
     backoff_multiplier: float = 2.0

src/dispatcher.py

@@ -59,6 +59,7 @@ from capabilities import detect_capabilities, iter_image_url_values
 from config import FlexPreference, RouterConfig, load_config
 from context_prune import (
     extract_text,
+    estimate_tokens,
     order_by_relevance,
     prune_context,
     trim_candidates,
@@ -1818,7 +1819,7 @@ def _embed_for_relevance(
     return order_by_relevance(embeddings[0], embeddings[1:])
 
 
-def _relevance_order_for(messages: list[dict], cfg) -> Optional[list[int]]:
+def _relevance_order_for(messages: list[dict], cfg, budget_tokens: int) -> Optional[list[int]]:
     """Compute the relevance_order for prune_context, or None (uniform).
 
     When pinch.relevance is off, or candidate count is below min_candidates,
@@ -1831,6 +1832,11 @@ def _relevance_order_for(messages: list[dict], cfg) -> Optional[list[int]]:
     candidates, _, _ = trim_candidates(messages, cfg.pinch.keep_last_turns)
     if len(candidates) < rel.min_candidates:
         return None
+    # Skip the embed when the conversation is under budget — prune_context would
+    # be a no-op, so ranking is pointless.
+    orig_tokens = sum(estimate_tokens(extract_text(m)) for m in messages)
+    if orig_tokens <= budget_tokens:
+        return None
     query = _last_user_text(messages)
     candidate_texts = [_text_only(messages[i]) for i in candidates]
     return _embed_for_relevance(query, candidate_texts, cfg)
@@ -2275,7 +2281,7 @@ def chat_completions(body: dict[str, Any], background: BackgroundTasks):
                 budget_tokens=cfg.pinch.budget_tokens,
                 keep_last_turns=cfg.pinch.keep_last_turns,
                 max_summarize_chars=cfg.pinch.max_summarize_chars,
-                relevance_order=_relevance_order_for(messages, cfg),
+                relevance_order=_relevance_order_for(messages, cfg, budget_tokens=cfg.pinch.budget_tokens),
             )
             logs.debug(
                 "pinch",
@@ -2507,7 +2513,7 @@ def chat_completions(body: dict[str, Any], background: BackgroundTasks):
                 budget_tokens=cfg.pinch.budget_tokens,
                 keep_last_turns=cfg.pinch.keep_last_turns,
                 max_summarize_chars=cfg.pinch.max_summarize_chars,
-                relevance_order=_relevance_order_for(messages, cfg),
+                relevance_order=_relevance_order_for(messages, cfg, budget_tokens=cfg.pinch.budget_tokens),
             )
             logs.debug(
                 "pinch",

tests/test_admin_config.py

@@ -20,6 +20,7 @@ import threading
 from pathlib import Path
 
 import pytest
+import yaml
 from fastapi import FastAPI
 from starlette.testclient import TestClient
 
@@ -138,7 +139,8 @@ def test_config_POST_creates_backup_before_write(client, tmp_path):
     # The backup captured the sentinel comment and the PRE-write value.
     backup_text = backups[0].read_text()
     assert _SENTINEL in backup_text
-    assert re.search(r"^\s*enabled:\s*false\s*$", backup_text, re.MULTILINE) is not None
+    backup_yaml = yaml.safe_load(backup_text)
+    assert backup_yaml["circuit_breaker"]["enabled"] is True
 
 
 def test_config_concurrent_writes_are_atomic_no_zero_byte_backups(tmp_path):

tests/test_admin_health.py

@@ -11,6 +11,8 @@ from __future__ import annotations
 
 import os
 import sqlite3
+import subprocess
+import sys
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
 

tests/test_chat_completions.py

@@ -104,6 +104,9 @@ def router(tmp_path, monkeypatch):
     # The local vision fallback is off unless a test opts in; without this it
     # would fire for every image request and try to reach localhost.
     monkeypatch.setattr(dispatcher.cfg.local_vision, "enabled", False)
+    # Isolate the shared module-level session cache: a real cache keyed on the
+    # same fingerprint leaks a classification from one test into the next.
+    monkeypatch.setattr(dispatcher.cfg.session_cache, "enabled", False)
     monkeypatch.setenv("NEURALWATT_API_KEY", "test-key")
 
     calls = []

tests/test_config_endpoints.py

@@ -237,7 +237,7 @@ def test_nonpositive_local_timeout_is_rejected(raw):
 
 def test_the_shipped_config_loads_the_pinch_section(raw):
     loaded = RouterConfig(**raw)
-    assert loaded.pinch.enabled is False
+    assert loaded.pinch.enabled is True
     assert loaded.pinch.budget_tokens > 0
     assert loaded.pinch.keep_last_turns > 0
 
@@ -246,7 +246,7 @@ def test_pinch_defaults_when_absent(raw):
     cfg = copy.deepcopy(raw)
     cfg.pop("pinch")
     loaded = RouterConfig(**cfg)
-    assert loaded.pinch.enabled is False
+    assert loaded.pinch.enabled is True
     assert loaded.pinch.budget_tokens == 50000
     assert loaded.pinch.keep_last_turns == 4
 
@@ -294,7 +294,7 @@ def test_pinch_max_summarize_chars_accepts_default(raw):
 
 def test_pinch_relevance_defaults_load(raw):
     loaded = RouterConfig(**raw)
-    assert loaded.pinch.relevance.enabled is False
+    assert loaded.pinch.relevance.enabled is True
     assert loaded.pinch.relevance.model == "nomic-embed-text"
     assert loaded.pinch.relevance.min_candidates == 2
 
@@ -303,7 +303,7 @@ def test_pinch_relevance_defaults_when_pinch_absent(raw):
     cfg = copy.deepcopy(raw)
     cfg.pop("pinch")
     loaded = RouterConfig(**cfg)
-    assert loaded.pinch.relevance.enabled is False
+    assert loaded.pinch.relevance.enabled is True
     assert loaded.pinch.relevance.timeout_seconds == 10
 
 
@@ -323,7 +323,7 @@ def test_nonpositive_pinch_relevance_min_candidates_is_rejected(raw):
 
 def test_circuit_breaker_defaults_load(raw):
     loaded = RouterConfig(**raw)
-    assert loaded.circuit_breaker.enabled is False
+    assert loaded.circuit_breaker.enabled is True
     assert loaded.circuit_breaker.initial_cooldown_seconds == 30
     assert loaded.circuit_breaker.max_cooldown_seconds == 600
     assert loaded.circuit_breaker.backoff_multiplier == 2.0

tests/test_dispatcher_helpers.py

@@ -120,7 +120,7 @@ def test_embed_for_relevance_request_exception_returns_none(monkeypatch):
 
 def test_relevance_order_for_returns_none_when_disabled():
     cfg = _cfg_with_relevance(enabled=False)
-    assert dispatcher._relevance_order_for([], cfg) is None
+    assert dispatcher._relevance_order_for([], cfg, budget_tokens=100) is None
 
 
 def test_relevance_order_for_below_min_candidates_skips_embedding(monkeypatch):
@@ -138,7 +138,27 @@ def test_relevance_order_for_below_min_candidates_skips_embedding(monkeypatch):
         {"role": "tool", "name": "read", "content": "x" * 6000},
         {"role": "assistant", "content": "a"},
     ]
-    assert dispatcher._relevance_order_for(messages, cfg) is None
+    assert dispatcher._relevance_order_for(messages, cfg, budget_tokens=100) is None
+    assert called["n"] == 0
+
+
+def test_relevance_order_for_skips_embedding_when_under_budget(monkeypatch):
+    called = {"n": 0}
+
+    def fake_post(*a, **k):
+        called["n"] += 1
+        return _FakeResp(status_code=200, json_payload={"data": None})
+
+    monkeypatch.setattr(dispatcher.requests, "post", fake_post)
+    cfg = _cfg_with_relevance(enabled=True, min_candidates=2)
+    # Two short tool-results -> token estimate under budget -> no embed call.
+    messages = [
+        {"role": "user", "content": "q"},
+        {"role": "tool", "name": "read", "content": "short"},
+        {"role": "tool", "name": "read", "content": "also short"},
+        {"role": "assistant", "content": "a"},
+    ]
+    assert dispatcher._relevance_order_for(messages, cfg, budget_tokens=50000) is None
     assert called["n"] == 0