compatibility with opencode-dcp(dynamic context prune)

[chenbo-again]

I wonder if there is some compatible issue with dcp? because they both manipulate session context, and I also notice the failure rate of dcp tool calling is increased, and sometimes, summary is even higher than compressed after compressed.

[Ultronen]

There is a concrete interaction worth checking, but the symptoms alone do not establish an incompatibility.

In Hindsight's current standalone @vectorize-io/opencode-hindsight plugin, hooks.ts at f7dd3f4 injects recalled memories into the system prompt and also appends recalled memories to the compaction context. DCP's plugin entry point also registers a system-prompt transform and a message transform. These shared surfaces make a controlled comparison useful; they do not prove that either plugin breaks the other.

One detail matters when isolating this: autoRecall: false does not disable Hindsight's compaction recall in that source version. I invoked its actual hooks with mocked OpenCode/Hindsight clients. Disabling autoRecall suppressed the system-prompt recall, and disabling autoRetain suppressed idle retention, but invoking the compaction hook with both flags disabled still called recall and appended <hindsight_memories> to its context. This was a hook-level test, not an end-to-end test with DCP or your model.

Consequently, adding memory during compaction is one possible contributor to a larger context afterwards; it is not proof of the cause of your larger summary or failed tool calls. To isolate the interaction, compare fresh sessions running DCP alone, Hindsight alone, and both, with the same task/model/settings. Remove the Hindsight plugin entry for the DCP-only case rather than relying on those two flags. Use a disposable memory bank with identical starting contents for each Hindsight case, since earlier runs can change later recall results.

Please include the exact OpenCode, DCP and Hindsight package versions/name, plugin order, the failing tool name and error, and the before/after token measurements you are comparing. The package name matters because the standalone OpenCode plugin and the newer coding-agents integration are separate implementations. A tool argument/schema failure needs different evidence from a connection failure. Those details would let us investigate a specific interaction instead of inferring one from context growth alone.