Activity

1. opencode-agent commented on Oct 12, 2026

   

   opencode-agentBot

   on Oct 12, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed report and the database scan. I was able to reproduce this on 2.0.27 with opencode/big-pickle.

   I ran a series of one-shot turns (opencode run -m opencode/big-pickle "Write a short haiku about number N, thinking carefully first."), then scanned the local session database. On a fresh database, 4 of 24 assistant messages with tokens.reasoning > 0 had only a text part. Their reasoning token counts were 1174, 461, 380 and 61. Their text was clean, with no inline chain-of-thought or <think> tags. The other turns had a normal reasoning part. The script below runs the turns and the scan.

   What I couldn't establish is where the reasoning gets lost. In the one turn I managed to capture on the wire, Zen streamed reasoning_content and OpenCode saved it correctly. I didn't capture an affected turn, so it's still open whether those responses contain reasoning that OpenCode drops or none at all. I also couldn't check the latest dev against this model, because Zen's free tier refuses requests from a source build. 2.0.27 is the latest release.

   Related issues, for reference:

   • desktop: reasoning part stops being recorded in a long session — chain-of-thought shows up as normal answer text #54298: reasoning part stops being recorded in long sessions, but there the chain-of-thought ends up in the answer text
   • reasoning: opencode zen (big-pickle) inline </think> leaks into text content instead of a reasoning part #54599: Big Pickle's inline </think> leaks into the text instead of becoming a reasoning part

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Repro for #54886: run N one-shot turns with opencode/big-pickle (free Zen model),
   # then scan the local DB for assistant messages with tokens.reasoning > 0 but no reasoning part.
   # Usage: OC="opencode" ./repro.sh 15   (or OC="bun ... packages/cli/src/index.ts")
   set -u
   N=${1:-15}
   OC=${OC:-opencode}
   mkdir -p /tmp/oc54886 && cd /tmp/oc54886
   for i in $(seq 1 "$N"); do
     timeout 150 $OC run --standalone -m opencode/big-pickle "Write a short haiku about number $((i+200)), thinking carefully first." >/dev/null 2>&1
   done
   python3 - <<'PY'
   import sqlite3, json, os
   c = sqlite3.connect(os.path.expanduser('~/.local/share/opencode/opencode.db'))
   cols = [r[1] for r in c.execute("pragma table_info(session_message)")]
   total = bad = 0
   for row in c.execute("select * from session_message"):
       d = dict(zip(cols, row))
       for k, v in list(d.items()):
           if isinstance(v, str) and v.startswith('{'):
               try: d.update(json.loads(v))
               except Exception: pass
       content, tok = d.get('content'), d.get('tokens')
       if not isinstance(content, list) or not isinstance(tok, dict) or not tok.get('reasoning'): continue
       total += 1
       if not any(p.get('type') == 'reasoning' for p in content):
           bad += 1
           print('MISSING reasoning part:', d.get('session_id'), 'tokens.reasoning =', tok['reasoning'], 'parts =', [p.get('type') for p in content])
   print(f'{bad}/{total} assistant messages with tokens.reasoning > 0 have no reasoning part')
   PY

   Run it: bash repro.sh
2. opencode-agent

   added

   reproduced

   on Oct 12, 2026