Import foreign agent sessions and MCP servers into omp
as native JSONL sessions and mcp.json entries — so conversations started
in OpenCode, Google Antigravity, or Meta Muse Code (and MCP servers
configured there) keep working after you switch to omp.
One shared provenance rule across every importer: imported tool and
output history stays plain text — never synthesized toolCall blocks,
so replays can't fake tool provenance.
| Skill | Imports from | How |
|---|---|---|
skills/opencode-import |
OpenCode opencode.db — 1.X stores (session/message/part) and 2.X stores (session_v2/session_message) |
In-repo converter (src/core.ts) + /opencode-import extension (adapters/omp). Reads the DB with sqlite3 -readonly; idempotent per-runtime registry |
skills/antigravity-import |
Antigravity CLI brain/<uuid>/.system_generated/logs/transcript_full.jsonl (+ titles from conversation_summaries.db) |
scripts/antigravity_to_omp.py (stdlib only). Sidesteps the protobuf steps.step_payload blobs in conversations/*.db, which have no stable schema |
skills/muse-code-import |
Muse sessions/<yyyy>/<mm>/<dd>/<uuid>/session.jsonl (+ session-index.db for list/title/workspace) |
scripts/muse_to_omp.py (stdlib only). Maps framed run/task events in sequence order; reasoning bodies are server-side encrypted and task/output chunks duplicate tool results — both skipped with counts |
skills/mcp-import |
~/.config/muse/settings.json, ~/.gemini/mcp_config.json |
scripts/mcp_to_omp.py (stdlib only). Adds names missing from ~/.omp/agent/mcp.json; skips source-disabled entries by default and anything omp already discovers itself (opencode.json, codex, gemini settings, claude, cursor, windsurf, vscode need no import — check /mcp list first) |
The OpenCode core was originally sourced from
pi-opencode-session-import
(MIT, see LICENSE.upstream) and is maintained here.
git clone https://github.com/Twilight0/omp-import-skill.git
cd omp-import-skill
# skills (pick the ones you need)
ln -s "$PWD/skills/opencode-import" ~/.omp/agent/skills/opencode-import
ln -s "$PWD/skills/antigravity-import" ~/.omp/agent/skills/antigravity-import
ln -s "$PWD/skills/muse-code-import" ~/.omp/agent/skills/muse-code-import
ln -s "$PWD/skills/mcp-import" ~/.omp/agent/skills/mcp-import
# /opencode-import extension (required by the opencode-import skill)
ln -s "$PWD/adapters/omp" ~/.omp/agent/extensions/opencode-importRestart the session afterwards (skills and extensions load at startup).
Requires sqlite3 and bun on PATH (bun test runs the opencode core suite).
/opencode-import # pick an OpenCode session (v1 + v2)
/opencode-import list [search] # browse (--cwd, --since, --limit)
/opencode-import all --dry-run # preview bulk import
/opencode-import all # bulk import, idempotent
scripts/antigravity_to_omp.py <uuid> --dry-run # preview
scripts/antigravity_to_omp.py <uuid> --cwd /path/to/project
scripts/muse_to_omp.py --list # browse muse sessions
scripts/muse_to_omp.py <uuid> --dry-run # preview
scripts/mcp_to_omp.py --dry-run # preview MCP merge
scripts/mcp_to_omp.py --source muse --dry-run # one source
scripts/mcp_to_omp.py # merge into ~/.omp/agent/mcp.json
omp --resume <written-file> # open an imported sessionConverted sessions land in ~/.omp/agent/sessions/<encoded-cwd>/ with a
parentSession pointer back at the source (opencode:<id>,
antigravity:<uuid>, muse:<uuid>); MCP merges back up the target with a
timestamped .bak first. Run /mcp reload (or restart) to pick up new MCP
servers.
skills/opencode-import/SKILL.md # opencode workflow + safety rules
skills/antigravity-import/SKILL.md # mapping table + workflow
skills/muse-code-import/SKILL.md # mapping table + workflow
skills/mcp-import/SKILL.md # MCP import workflow + safety rules
src/core.ts # opencode conversion core (v1 + v2)
src/core.test.ts # `bun test` (15 tests)
adapters/omp/index.ts # /opencode-import extension
scripts/antigravity_to_omp.py # brain transcript -> omp JSONL
scripts/muse_to_omp.py # muse session.jsonl -> omp JSONL
scripts/mcp_to_omp.py # foreign MCP defs -> mcp.json merge
LICENSE # MIT, this project
LICENSE.upstream # MIT attribution (opencode core)
- Antigravity
brain/logs can lag a live conversation; import idle ones. - Assistant
GENERICrows mix prose and tool output; kept as text, truncatable via--max-text-chars. - Muse reasoning bodies are server-side encrypted and
task/outputchunks duplicate committed tool results; both are skipped, counts shown. - MCP merge skips source-disabled entries unless
--include-disabled; same-name entries already in the target are never touched. - Stock-weight models keep portable reasoning; abliterated-weight reasoning never replays onto other hosts (enforced at import time by keeping it inside plain assistant text, never tool payloads).