Harnesses

scrubjay is not Claude-only. A harness is the coding agent that produces and consumes sessions — Claude Code, opencode, or codex. Everything scrubjay does — sync your config in, relay each session out, and recall it later from any machine — is defined against a small seam so a new harness can be added without touching the archive, the logs/ catalogue, memory, or the readable Markdown layer that makes cross-harness recall work.

Which harnesses a machine syncs into is SCRUBJAY_HARNESSES in ~/.config/scrubjay/config (space-separated; default claude). Onboarding auto-detects the agents installed on the machine and sets this for you.

What's supported

	Claude Code	opencode	codex
Relay sessions to the archive	✅	✅	✅
Cross-harness recall (readable rendering)	✅	✅	✅
Read the archive from inside (/sjrecall & co.)	✅	✅	—
Config sync into the harness	✅	✅	—
Session hand-off (continue a session on another machine)	✅	✅	—

"Cross-harness recall" is the payoff of the shared shape: a session started in Claude Code on your laptop is searchable — and readable — from opencode on your desktop, and vice versa, because every harness's renderer emits the same # title / _N turns_ / ## User · ## Assistant Markdown.

opencode

opencode is a first-class harness: your history relays into the archive, you can recall it from inside opencode, and your setup follows you there too.

If opencode is on PATH when you onboard, it's detected automatically and there's nothing more to do. To add it to a machine that's already set up:

echo ': "${SCRUBJAY_HARNESSES:=claude opencode}"' >> ~/.config/scrubjay/config
bin/sync-config.sh

sync-config.sh then writes the following into ~/.config/opencode/ — additively, and it never rewrites a key it doesn't own (a setting you chose that the data repo never mentions always survives, and an unparseable opencode.json is refused rather than clobbered):

What	Where	From
shared + per-host settings	opencode.json (deep-merged under your own keys)	opencode/opencode.base.json + hosts/<host>/opencode/opencode.json
shared instructions	opencode.json → instructions[]	shared/AGENTS.md (by absolute path — live on git pull)
the lifecycle bridge	opencode.json → plugin[]	the app (sessions get logged + relayed)
the archive server	opencode.json → mcp.sjmcp	the app (/sjrecall, /sjfind, /sjbrowse work inside opencode)
the /sj* + your commands	commands/*.md	the app + claude-md/commands/ (yours override on a name clash)
agents	agent/*.md	claude-md/agents/ (translated) + opencode/agent/ (native)

The settings merge is the same model as Claude's settings.base.json + per-host overlay: shared defaults apply, the host overlay wins where it sets a key, and every key you set that the data repo doesn't mention is left untouched.

How Claude agents become opencode agents

claude-md/agents/*.md are authored for Claude, so scrubjay translates rather than copies them: it drops the Claude-only name:/model: lines, adds mode: subagent, and maps Claude's tools: allowlist onto opencode's permission map — allowing the listed tools and denying every other tool opencode knows, so a restricted agent can't quietly gain access it was never granted. Native agents you write for opencode go in opencode/agent/ and are used as-is (they win a name clash). Generated files are rewritten on every sync — edit the source in the data repo, never the copy.

Same-harness hand-off runs opencode import for you: bin/sj-resume.sh <handle> --import (the default) stages another machine's session and imports it into opencode in the destination project, leaving one command to run. See Session hand-off.

codex

codex relays and is recallable today — its sessions land in the archive with a readable rendering, so they show up in /sjrecall alongside everything else. Config sync into codex, reading the archive from inside codex, and hand-off into it are not built yet: codex keeps its config as TOML (where scrubjay's settings story is a jq merge) and resolves session ids through its own index, both of which need dedicated work. The details are tracked in bin/adapters/ROADMAP.md.

Adding a harness

The seam is bin/adapters/<harness>.sh: an adapter answers where the agent keeps its config, what a session's records are, how it renders to the shared Markdown shape, and how a session is resumed. The contract for implementers is bin/adapters/README.md; claude.sh is the reference implementation.