Runtime support

Reflectors, curators, and mergers run in the harnesses you already use. honed connects its tools (submit_reflection, or commit_change and finish) to each run as an MCP server served from your own process, and turns off your own customizations — hooks, skills, context files — where the harness allows it.

Harness Reflect Hand off a session Curate / merge How
Codex yes yes yes codex exec --json (exec fork to hand off, exec resume for later turns); read-only sandbox to reflect, workspace-write to edit
Claude Code yes yes yes claude -p; Read/Grep/Glob, plus Edit/Write to edit
Pi yes yes yes pi -p --mode json; read/grep/find/ls, plus edit/write; tools through a run-only extension
OpenCode yes yes yes opencode run --format json; isolated run config, restricted permissions, per-run MCP tools
Antigravity no no no reads MCP servers only from your own configuration, so honed cannot connect its tools for one run

Sessions can be indexed from Claude Code, Codex, Pi, OpenCode, and Antigravity, and exported to any of their formats; --best-effort keeps what a format cannot represent as labelled text.

Notes:

  • Agents never run commands while curating or merging; every commit goes through commit_change.
  • Handed-off sessions are installed as copies in the harness's home (or home=), redacted (see redaction); the original transcript is never changed.
  • timeout= bounds each turn: the subprocess is killed for Claude Code and Pi, and for Codex the turn is abandoned and its client closed. Either way the run fails with a timeout error.
  • Each run's token usage comes from the harness: Claude Code's JSON result (usage, total_cost_usd), Pi's assistant messages (usage, one per model call), Codex's thread token usage (its running total, so each turn counts the difference), and OpenCode's JSON step_finish events (tokens, cost). See token usage.