Compaction permanently fails with 'Duplicate value for tool_call_id' after an interrupted tool run #40235

Summary

An aborted/interrupted tool run leaves a second tool part with the same callID in the message. The next auto-compaction then re-sends the full history to the provider and fails with 400 invalid_request_error: Duplicate value for 'tool_call_id', so compaction never completes (summary never written, token counter never resets). The session is effectively poisoned — every subsequent compaction fails the same way.

Root cause evidence (local sqlite DB, opencode v1.18.11)

One assistant message contains two tool parts sharing the same callID: a completed one (tool=todowrite) and an error one (tool=unknown, error: "Tool execution aborted", metadata.interrupted: true).

Message msg_fb9813f84... (session ses_0507e87b...):

• part prt_fb9814a85...: {"type":"tool","tool":"todowrite","callID":"functions.todowrite:1","state":{"status":"completed",...}}
• part prt_fb9816675...: {"type":"tool","tool":"unknown","callID":"functions.todowrite:1","state":{"status":"error","input":{},"raw":"","error":"Tool execution aborted","metadata":{"interrupted":true},...}}

Message msg_fc35e22d5... (same session):

• part prt_fc35e5d6d...: {"type":"tool","tool":"todowrite","callID":"functions.todowrite:0","state":{"status":"completed",...}}
• part prt_fc35e7502...: {"type":"tool","tool":"unknown","callID":"functions.todowrite:0","state":{"status":"error","input":{},"raw":"","error":"Tool execution aborted","metadata":{"interrupted":true},...}}

The unknown part with Tool execution aborted is written when a run is interrupted; it duplicates the callID of the tool call that was already recorded as completed.

Failure symptom

Two consecutive auto-compactions in the same session failed identically:

1. msg_fb9e6c8bf... (compaction, finish=None, err=true):
   Error from provider (DeepSeek): Duplicate value for 'tool_call_id' of functions.todowrite:1 in message[53] (400)
2. msg_fc45786f2... (compaction, finish=None, err=true):
   Error from provider (Console): Upstream request failed: [invalid_request_error] Duplicate value for 'tool_call_id' of functions.todowrite:0 in message[224] (400)

Log excerpt:

timestamp=...T21:18:22.176Z level=ERROR ... agent=compaction mode=primary
  error.error="AI_APICallError: Error from provider (Console): Upstream request failed: [invalid_request_error] Duplicate value for 'tool_call_id' of functions.todowrite:0 in message[224]"

Because compaction serializes the entire history head to the model, the duplicate is always present → compaction aborts before producing a summary → the message count / token counter keep growing and the session never recovers, even though compaction.auto and prune are enabled.

Expected behavior

• Tool parts must be deduplicated by callID before building model messages (drop the unknown/interrupted duplicate, keep the completed one), or
• interrupted runs should not write a second tool part with an existing callID at all.

Impact

Sessions that suffer an interrupted tool run become permanently un-compactable (until a fresh session). Auto-compaction then loops on the same 400, the token counter stays high, and the user is forced to /new.

Repro notes / workaround

• Took the evidence from ~/.local/share/opencode/opencode.db (tables message, part; parts of type tool with duplicate callID where one is tool="unknown", error="Tool execution aborted").
• Workaround for the affected session: delete the duplicate unknown/interrupted tool part (keep the completed one) while opencode is closed.