summarize: raise output-token budget to 16k for reasoning models

Reasoning models (MiniMax M-series, Claude with thinking) emit a variable
<think> block before the JSON, overrunning the ~4000-token provider
default and truncating the structured output. Forward an explicit
MAX_OUTPUT_TOKENS=16000 ceiling (override via PODMIND_MAX_OUTPUT_TOKENS);
non-thinking models like the DeepSeek default still stop at ~1.3k, so
cost is unchanged for them.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: 2 people

bin/summarize.py

@@ -8,6 +8,7 @@
 import argparse
 import asyncio
 import json
+import os
 import sys
 from pathlib import Path
 
@@ -22,6 +23,15 @@
 RESULTS = Path("/tmp/podmind-results")
 RESULTS.mkdir(exist_ok=True)
 
+# Output-token ceiling for the summary call. The structured JSON is only ~1-2k
+# tokens, but reasoning models (MiniMax M-series, Claude with thinking) spend a
+# large, variable amount of the budget on a <think> block first — a 4000 cap
+# truncated the JSON for ~25-50% of episodes on MiniMax-M3 (some needed >9000).
+# This is a ceiling, not a target: non-thinking models (DeepSeek) still stop at
+# ~1.3k, so the higher cap doesn't change their cost. Override per-vault if
+# needed.
+MAX_OUTPUT_TOKENS = int(os.environ.get("PODMIND_MAX_OUTPUT_TOKENS", "16000"))
+
 PROMPT = """You are extracting a structured summary of a podcast episode for a knowledge wiki.
 
 Read the meta and transcript below. Produce a single JSON object with these exact fields. Do not wrap in markdown fences. Do not include any text before or after the JSON.
@@ -104,7 +114,7 @@ async def process_one(client: httpx.AsyncClient, provider: llm.LLMProvider, item
     async with sem:
         prompt = build_prompt(item)
         try:
-            content, usage = await provider.achat(client, prompt)
+            content, usage = await provider.achat(client, prompt, max_tokens=MAX_OUTPUT_TOKENS)
         except llm.LLMError as e:
             print(f"  [error] {item['raw_dir']}: {e}", file=sys.stderr)
             return item["idx"], False, llm.Usage()

tests/test_review_fixes.py

@@ -79,6 +79,53 @@ def test_placeholder_in_meta_not_substituted(self, tmp_path: Path):
         assert "raw_dir: test-show/2026-01-01-test-ep" in prompt
 
 
+# ---------- summarize output-token budget (reasoning-model fix) ----------
+
+class TestMaxOutputTokens:
+    def test_default_budget_is_16000(self):
+        import summarize
+        assert summarize.MAX_OUTPUT_TOKENS == 16000
+
+    def test_env_override(self, monkeypatch):
+        import importlib
+        import summarize
+        monkeypatch.setenv("PODMIND_MAX_OUTPUT_TOKENS", "5000")
+        importlib.reload(summarize)
+        try:
+            assert summarize.MAX_OUTPUT_TOKENS == 5000
+        finally:
+            monkeypatch.delenv("PODMIND_MAX_OUTPUT_TOKENS", raising=False)
+            importlib.reload(summarize)  # restore default for other tests
+
+    def test_process_one_forwards_budget_to_achat(self, tmp_path, monkeypatch):
+        import asyncio
+
+        import summarize
+        from podmind import llm, paths
+
+        d = paths.EPISODES_DIR / "test-show" / "2026-01-01-budget-ep"
+        d.mkdir(parents=True, exist_ok=True)
+        (d / "meta.json").write_text(json.dumps({"title": "T", "show": "S", "duration_sec": 60}))
+        (d / "transcript.md").write_text("hello world transcript")
+        monkeypatch.setattr(summarize, "RESULTS", tmp_path)  # don't touch /tmp/podmind-results
+        item = {"raw_dir": "test-show/2026-01-01-budget-ep", "idx": "01",
+                "show_slug": "test-show", "date": "2026-01-01"}
+
+        seen = {}
+
+        class FakeProvider:
+            async def achat(self, client, prompt, **kwargs):
+                seen["max_tokens"] = kwargs.get("max_tokens")
+                return '{"hook": "x", "takeaways": []}', llm.Usage(10, 0, 5)
+
+        async def go():
+            return await summarize.process_one(None, FakeProvider(), item, asyncio.Semaphore(1))
+
+        idx, ok, _usage = asyncio.run(go())
+        assert ok is True
+        assert seen["max_tokens"] == summarize.MAX_OUTPUT_TOKENS == 16000
+
+
 # ---------- tick_finalize.to_episode resilience ----------
 
 class TestToEpisodeResilience: