The work records cover 2 harnesses out of 16

`deja files`, `deja restore`, the after-compaction block and the moved-since annotation all read the same material: the paths a turn opened, the commands it ran, the spans an edit replaced. That material is extracted from **Claude and opencode only**. Counted on a 1150-session index, sessions carrying each record type: | harness | files | commands | edits | |---|---|---|---| | claude | yes | yes | yes | | opencode | yes | yes | yes | | codex | — | — | — | | cursor | — | — | — | | antigravity | — | — | — | | gemini, cline, copilot, … | — | — | — | So a Cursor or Codex user installs 0.16.5 and every one of those features is silent. That is worse than not having them: the commands exist, print nothing, and give no reason why. **The material is there in both stores I checked.** Cursor CLI transcripts carry `tool_use` blocks in the same shape Claude uses: ```json {"role":"assistant","message":{"content":[{"type":"tool_use","name":"CallDynamicTool","input":{…}}]}} ``` Codex rollouts carry `function_call` and `function_call_output` payloads alongside `message` and `reasoning`. Both are parsed today for their text and their tool calls are dropped on the floor. Worth doing in that order — Cursor first, since its shape is nearly identical to the Claude decoder that already exists — and worth checking the same question for the remaining twelve rather than assuming.