[big-pickle] Agente entra en loop de texto sin emitir tool calls (se atasca narrando la acción a tomar) #46030

Activity

1. github-actions commented on Aug 28, 2026

   

   github-actionsBot

   on Aug 28, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • Bug: OpenCode enters infinite loop after tool calls complete (Zen/big-pickle) #26220: Bug: OpenCode enters infinite loop after tool calls complete (Zen/big-pickle) — same model, similar loop behavior
   • big-pickle got stuck in an autoregressive loop during a brainstorming task #42956: big-pickle got stuck in an autoregressive loop during a brainstorming task — same model, stuck generating text instead of acting
   • subagent: infinite loop of identical tool calls for ~50min with no loop protection, uncontrollable token burn #45442: subagent: infinite loop of identical tool calls for ~50min with no loop protection — related loop/no-progress behavior

   If your issue is distinct (text narration loop specifically, not tool-call repetition), please clarify how it differs from the above.
2. aentrepreneur commented on Aug 31, 2026

   

   aentrepreneur

   on Aug 31, 2026

   Author

   More actions

   Clarifying how this differs from the linked duplicates, plus fresh log evidence from the local setup (wanted to add data before checking the bot's dedup suggestions).

   How this is distinct

   • big-pickle got stuck in an autoregressive loop during a brainstorming task #42956 (big-pickle stuck in autoregressive loop during brainstorming): same model, but the stuck phase there is generation during an open-ended creative task. This report is the closing/persistence phase — the model has already decided on concrete actionable steps (write backup, edit, then a tool call) and still cannot emit the tool.
   • subagent: infinite loop of identical tool calls for ~50min with no loop protection, uncontrollable token burn #45442 (infinite loop of identical tool calls): here the model emits no tool calls at all — it loops on text generation / re-requesting the LLM instead of calling tools. The opposite failure of the tool-call-repetition loop.
   • Bug: OpenCode enters infinite loop after tool calls complete (Zen/big-pickle) #26220 (infinite loop after tool calls complete): that is post-tool stagnation. This report is pre-tool — narration without ever issuing the call.

   So the distinguishing signature: repeated LLM stream calls with zero tool emissions, and it does not self-recover (needs a manual abort).

   Log evidence (2026-08-28, opencode 1.18.22, Ubuntu 24.04/WSL2)

   From ~/.local/share/opencode/log/opencode.log, all in a single run a9c7879c, model opencode/big-pickle, agent plan:

   1. Narration loop without tool calls (00:58–01:01):

   • 8 consecutive message=stream events for big-pickle within ~3 minutes.
   • Between those streams: 0 real tool calls (only the initial shell tool using shell before the window).
   • The first stream (00:58:08) was cut and needed a manual abort (level=ERROR ... error=Aborted at 00:58:11) — it did not exit the loop on its own.

   2. Recurring burst (16:05–16:09):

   • ~28 consecutive big-pickle streams in 4 minutes with only 1 permission=bash event and 0 errors → chained text generation without acting.

   3. Global frequency (that day):

   • 523 stream LLM calls for big-pickle in this single run (all other runs that day total 13). Peak of 10 streams/minute at 16:06.
   • Streams are 3–15 s apart, i.e. short generation turns with no tool side effects — matching the "narration/no-op" signature.

   Volunteer data for whoever triages: happy to share the session/snapshot or a larger log excerpt. This is the same platform as #42950 (which just got a Windows confirmation), so it may be worth correlating the two — both involve opencode/big-pickle (and adjacent free models) stalling or degrading on the built-in provider.
3. aentrepreneur commented on Aug 31, 2026

   

   aentrepreneur

   on Aug 31, 2026

   Author

   More actions

   Adding the root cause hypothesis and a minimal reproduction — this is what I believe the runtime/loop actually does, so it can be fixed faster. Based on the same log data.

   Observed mechanism (what the logs show)

   Following one stream in the burst, the loop advances one integer per iteration:

   loop step=47   →  stream (big-pickle)  →  no tool-call part  →  loop step=48  →  stream …
   loop step=48   →  stream (big-pickle)  →  no tool-call part  →  loop step=49  →  stream …
   … (repeats)
   

   • Each agent-loop iteration issues exactly one LLM stream call for big-pickle.
   • The response contains no tool part (no tool-call), only text; so the runtime has nothing to execute and just iterates the loop again.
   • Timing: loop step=N → stream is 3–15 s apart; a truly-working step would show a tool execution/permission event between iterations. In the bursts there is none.
   • Global proof: the whole closing run did 556 loop steps (max step=69 per-part counter; 100s of stream calls). A closing task of this size (backup + edit + write + mine) should be well under ~40 steps. The rest was consumed by narration with no-op turns.

   Root cause hypothesis

   The model enters a degenerate text-generation attractor: with the decision already made and the correct next tool known, it keeps producing prose that describes the tool call instead of the tool call itself. Because the assistant turn ends with a text part and no tool part, the runtime correctly has nothing to run and re-enters the loop — so the harness is obeying the agent-loop contract, but the model never completes a turn with a tool call. Two contributing factors to check:

   1. Context pressure at end-of-session: this appeared at the closing/persistence phase of a long session (many prior steps, big accumulated context). Compare with the mimicry in opencode/big-pickle: intermittent socket disconnects cause silent drop (no UI error) + repeated Aborted in logs (v1.18.18, Linux/WSL2) #42950 (free-model degradation) — likely correlated.
   2. No loop protection for "text-only turns": there is protection for repeated tool calls (subagent: infinite loop of identical tool calls for ~50min with no loop protection, uncontrollable token burn #45442) but I don't see one for repeated tool-less turns. A guard that detects N consecutive assistant turns with no tool call (e.g. "no-progress" / "narration-only") and stops/retries would bound the blast radius regardless of model.

   Minimal reproduction

   1. Any session where the agent has a decided, single next tool call (e.g. closing a session: write file X, then a tool call), ideally late in a long session.
   2. Let big-pickle drive the last steps.
   3. Observe: assistant emits text narrating the action; no tool part; loop step monotonically increments; repeats until manually aborted.

   Suggested fix / ask

   1. Loop guard: detect N consecutive assistant turns that return text-only with no tool call and no final answer, then surface a warning / stop (mirroring the existing tool-call loop protection).
   2. Log the no-op: log a no-tool-part signal so it's countable in a dashboard (like the stream-error buckets).
   3. Verify whether end-of-session context length triggers it (compare a long vs short session).

   Happy to share the full session snapshot or a redacted log excerpt on request.
4. aentrepreneur commented on Sep 13, 2026

   

   aentrepreneur

   on Sep 13, 2026

   Author

   More actions

   Update (v1.18.30, 2026-09-12) — no sustained narration-loop bursts observed today; loop-guard ask still open

   Same environment, today's full-day log (3 sessions, opencode/big-pickle, main agents):

   • 274 stream calls ↔ 1,568 tool/permission events → healthy tool interleaving; no pure-narration signature.
   • Peak minutes 10–12 streams/min, isolated (not the sustained 4-min bursts at up to 10/min seen on 2026-08-28, run a9c7879c).
   • error=Aborted: 2, both preceded by message=cancel (user cancels).
   • Zero transport errors (stream error/rate-limit/cannot-connect) — consistent with the discriminator I documented in this thread: the loop failure produces clean logs, so today's data doesn't confirm or deny the loop; it just shows no burst occurred.

   The loop-guard ask remains unanswered and stands: detect N consecutive assistant turns that return a text-only part (no tool call, no final answer) and stop/surface a warning, mirroring the existing repeated-tool-call protection (#45442). A no-tool-part log signal would make this countable (like the stream-error buckets).

   Suggestion: keep open until a loop guard lands or maintainer triage decides it should move to a feature request. I'll re-run the burst detector after any loop-side patch and report delta.
5. aentrepreneur commented on Oct 21, 2026

   

   aentrepreneur

   on Oct 21, 2026

   Author

   More actions

   Status check on this one — no maintainer reply yet, so adding current context rather than new noise.

   • Environment still pinned to 1.19.3 (Ubuntu 24.04 / WSL2, opencode/big-pickle). Latest full-day logs from my side remain clean of the narration-loop class, so I cannot claim a V2 repro myself.
   • Question: is the narration-loop guard planned for V2, or should [big-pickle] Agente entra en loop de texto sin emitir tool calls (se atasca narrando la acción a tomar) #46030 be treated as a V1-only report and closed on that basis?

   If it is in scope for 2.x, I can upgrade and run a controlled retest (same prompt shape, same model, 3 sessions) and post logs there.
6. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   More actions

   ¡Gracias por el reporte detallado!

   Por la versión que indicas (1.18.22), esto es OpenCode v1. v1 solo recibe correcciones de bugs críticos (pérdida de datos, problemas de seguridad, cuelgues sin alternativa). Este comportamiento malgasta turnos y tokens, pero tiene una alternativa: interrumpir el patrón y pedir la herramienta directamente. Por eso es poco probable que se corrija en v1.

   Te recomendamos probar OpenCode v2 (paquete npm @opencode/cli, comando opencode), donde se hace el desarrollo activo.

   Issues relacionados con el mismo síntoma (el modelo anuncia la llamada a la herramienta pero no la emite):

   • agent: tool call looping - model repeats "let me make the call / let me go" without emitting the call #57501: el mismo bucle de "let me make the call" sin emitir la llamada
   • Agent enters an infinite "intent-narration" loop: announces tool calls but repeatedly fails to actually invoke them #46492: bucle de "narración de intención" con opencode/big-pickle
   • AI stuck in an infinite repetition loop and fails to execute commands #46239: el asistente se queda atascado repitiendo texto sin ejecutar comandos
   • System prompt: tell the model its between-tool-call text is never read #59012: propuesta abierta de cambiar el system prompt, que podría reducir este tipo de narración

   Si crees que es crítico, o si también te pasa en v2, responde aquí indicando la versión de v2 y los pasos para reproducirlo.
7. opencodejr

   added

   pending close

   triaging

   on Oct 28, 2026
8. opencodejr

   closed this as not plannedon Oct 29, 2026
9. opencodejr commented on Oct 29, 2026

   

   opencodejrBot

   on Oct 29, 2026

   More actions

   Cerramos este issue como no planificado. El reporte es sobre OpenCode v1 (1.18.22), que solo recibe correcciones de bugs críticos, y desde nuestro último comentario no ha habido respuesta que indique que sea crítico o que también ocurra en v2.

   Si te pasa lo mismo con opencode/big-pickle en OpenCode v2 (@opencode/cli), abre un nuevo issue con la versión exacta y los pasos para reproducirlo. ¡Gracias por tomarte el tiempo de reportarlo!