[Bug]: DCP context window may leak pre-fork messages after `/fork message` (model issue uncertain)

[InCerryGit]

Bug Description

After using /fork message in an opencode session, subsequent DCP (Dynamic Context Pruning) calls sometimes include message content from before the fork point in the active conversation context. This means messages previously seen in the pre-fork session unexpectedly reappear. It is unclear whether the root cause is the model itself, DCP context boundary management, or some other backend process. Please investigate. Steps to reproduce and some environmental info are below.

Expected Behavior

After forking a conversation with /fork message, DCP should only consider messages from the forked session as context. Content from before the fork point should not bleed into new DCP calls in the forked session.

Debug Context Logs

It is difficult to reproduce, and no valid logs have been captured so far.

Tool Call Details

/fork message was used in an opencode session. On the first or subsequent DCP call after the fork, content from before the fork was included in the window — but it is not clear if this is a DCP, backend, or model-level issue.

DCP Version

@tarquinen/opencode-dcp@3.1.0

Opencode Version

1.3.2

Model

Other (specify in description)

Additional Context

To reproduce:

1. Start an opencode conversation for a while
2. Use /fork message to fork the session
3. Continue the conversation from the new fork
4. Observe if any DCP call includes messages from before the fork.

Root cause might be improper fork/context scoping, cache leakage, or model-side memory, but it's not yet clear.

[Tarquinen]

yea forking doesn't work with DCP, when opencode forks sessions it creates new session IDs and new message IDs for each message so I don't have an easy way of mapping the compression state from the parent session to the forked session. As far as I know there's no opencode events that fire on forking either, so I don't know if I would be able to do some kind of index based mapping either...

[InCerryGit]

yea forking doesn't work with DCP, when opencode forks sessions it creates new session IDs and new message IDs for each message so I don't have an easy way of mapping the compression state from the parent session to the forked session. As far as I know there's no opencode events that fire on forking either, so I don't know if I would be able to do some kind of index based mapping either...

Is there any way to handle it manually?