agent: empty <conversation-checkpoint> replies after compaction/revert are marked succeeded (no tools, no output) #57240

Summary

After a compaction (and again after a revert to just-after that compaction), the model role-flips into OpenCode's compaction summarizer: it first emits the <conversation-checkpoint>…</checkpoint> serialization itself as a user-visible answer, then degenerates to answering every following turn with an empty <conversation-checkpoint><summary>(no content)</summary></checkpoint>. No tool parts, no substantive text — yet every step is stored finish=stop and closes idle outcome=succeeded, so nothing looks broken from the outside. Related: #50814 (flattened [Assistant tool call] transcripts after compaction). This is a distinct degenerate loop observed in the same session.

Environment

• opencode version: 2.0.31 (install/channel: latest)
• OS: Windows 11 Home 10.0.26200 (win32 x64)
• Terminal: TERM=xterm-256color; desktop client (TERM_PROGRAM not set)
• Shell: C:\Windows\system32\cmd.exe
• Active plugins: none (no plugin key in ~/.config/opencode/opencode.json, no local plugin dirs)
• Models: stepfun-ai/step-5-preview via a custom @ai-sdk/openai-compatible provider (NVIDIA NIM endpoint); mimo/mimo-v2.6-flash also affected earlier in the same session

Reproduction

1. Run a long build session until auto-compaction fires (session ses_eca4fc28dffdfCI4tEPJ6I7UhI; compaction record at seq 37478, 2026-10-19 15:47 local, 148 KB).
2. Revert the session to just after that compaction (the revert removed seq 37485–37608, i.e. the polluted turns — verified by diffing session rows before/after).
3. Send a message that touches the compaction summary (e.g. asking how summary formatting is produced).
4. Keep sending normal follow-up turns.

Observed (all verified against the session SQLite DB, not the TUI rendering):

After the revert the session produced 0 real tool calls. It was effectively dead while every turn reported success.

Expected Behavior

• The assistant answers user turns normally and emits structured tool parts; it must not reproduce compaction serialization as chat output.
• A turn with zero tool parts and no substantive text must not be recorded as a successful turn.

Actual Behavior

• The model emits the compaction scaffold (<conversation-checkpoint> / <summary> / <recent-context> / </checkpoint>) as its answer, then collapses to empty checkpoint blocks for every subsequent turn.
• <conversation-checkpoint> appears in none of the 80 stored compaction records, nor in instruction_blob, kv, or any config file on this machine (verified by search). The model reproduces the server-side serialization from context/memory, then imitates its own prior degenerate output.
• The single full checkpoint it emitted (seq 37655) also re-asserted fabricated statistics from the compaction: model emits tool calls as literal text after compaction — tools never run, session returns no output #50814 failure mode ("23 files / 41,857 chars → 10–14K tokens per turn"; the real measurement is 9 files / 210,189 chars), promoting a hallucinated number into a checkpoint-shaped "authoritative" summary that later turns then inherit.

Additional Context

• Evidence collected from a read-only copy of the session DB: session_message rows (type, content part types, finish, tokens, snapshot), type='compaction' rows, instruction_blob, kv.
• Recovery dead-ends observed:
  • Reverting to just-after the compaction keeps the polluted compaction record as the context base; the model relapses.
  • /compact is not a user message, so revert/undo cannot target it.
  • POST /api/session/{id}/fork copies projected history before a message, so any fork point after the first polluted compaction carries the pollution into the child session — forking cannot produce a clean continuation of this session at all.
• Suggested mitigations:
  1. Integrity check: an assistant message whose text/reasoning contains [Assistant tool call]:, [Tool result]:, or <conversation-checkpoint> while having zero tool parts should fail the step (or at least warn in the TUI + log), not close idle/succeeded.
  2. Make the compaction scaffold visibly non-imitable / out-of-band, since the model demonstrably role-flips into "summarizer" once the conversation touches compaction.
  3. Structural representation of prior tool calls in post-compaction prompts (tracked in compaction: model emits tool calls as literal text after compaction — tools never run, session returns no output #50814).

Happy to provide the full DB rows (seq 37655, 37677–37713) if useful.