opencode.md

OpenCode adapter

The adapter discovers local OpenCode sessions and normalizes a conservative subset into AEP v0.1. OpenCode persists each conversation as many small JSON files under a storage/ tree, defaulting to ${XDG_DATA_HOME:-~/.local/share}/opencode/storage:

storage/session/<projectID>/<sessionID>.json   session info
storage/message/<sessionID>/<messageID>.json   one message
storage/part/<messageID>/<partID>.json         one message part

The adapter targets this file-based JSON storage layout. Message and part identifiers are creation-ordered ascending ULIDs — OpenCode's Identifier generator (verified against sst/opencode at tag v0.15.0) prefixes each id with a big-endian millisecond timestamp, so a plain lexicographic sort is chronological. A session is therefore the incremental unit: the cursor records the highest message identifier already normalized and resumes past it. Recent OpenCode builds are migrating this data into a single SQLite database; that backing store is out of scope for v0.1 and is neither read nor written.

Tool parts are updated in place as a call moves through pending → running → completed/error. The cursor only advances across a contiguous prefix of fully terminal messages: the moment a message still holds a non-terminal tool part (or cannot be read), the watermark freezes at the previous message so that message — and everything after it — is re-read on the next import. Re-reads are safe because the store deduplicates unchanged events and inserts only the newly terminal tool.completed / tool.failed outcome. No failure-or-recovery signal is lost to a session that was still in flight at first import.

OpenCode's on-disk layout is an application detail rather than a stable integration interface, so this adapter treats every mapping as a version-observed capability: unknown message roles and part types are counted and skipped, fields are accessed defensively, and no semantic event is guessed. Anonymized fixtures guard the supported subset.

Preview and import

Preview the exact metadata-only discovery plan without opening a database:

mise exec -- cargo run -p autophagy-cli -- --output json \
  import --adapter opencode --dry-run

Import the default storage tree:

mise exec -- cargo run -p autophagy-cli -- \
  import --adapter opencode

Pass another storage/ root after import when needed. Repeat --project /exact/working/directory for selection. Filtered imports use independent cursor scopes and do not hide later unfiltered history.

Privacy and indexing

Prompt text, assistant messages, tool input, and tool output are omitted from event metadata by default. --include-content opts into local persistence under opencode.content. Tool inputs remain structural AEP evidence but enter FTS only with --index-tool-input. Use --index-metadata opencode.tool only when the content is locally approved for search. Change the content policy only with a fresh database because it changes canonical event bodies.

Default credential rules redact recognized secrets before persistence. Add repeatable --exclude-path GLOB values for project or artifact exclusions. Project filters, exclusion globs, and the privacy-policy version are part of the cursor scope, so changing policy safely rescans instead of losing skipped data.

Incremental and cross-adapter guarantees

The per-session cursor advances only across a contiguous prefix of terminal messages, so an unchanged reimport of a finished session reads no message records and inserts nothing. A session that was still in flight re-reads only its deferred tail until those tool calls terminate, and the store deduplicates the unchanged events. The session-start event is emitted once and tracked in cursor state. Event IDs are keyed on message and part identifiers (not positional indices), so rescans and out-of-order part writes stay idempotent.

Every native adapter runs through the same conformance harness. It requires a non-empty initial import, zero rejected or conflicting fixture evidence, a zero-read unchanged reimport, and an unchanged stored event count. CLI tests also import several adapters into one database and verify their source provenance remains distinct.

Capability matrix

OpenCode record shape	AEP v0.1 event	Notes
Session info	session.started	Native session ID and working directory retained
Message role user	prompt.submitted	Text collected from text parts, omitted unless opted in
Message role assistant text parts	decision.recorded	Emitted only when text parts are non-empty
tool part state.status completed	tool.called + tool.completed	Self-contained pair; file input preserved
tool part state.status error	tool.called + tool.failed	Failure carried from the error state
tool part state.status pending / running	tool.called	No terminal outcome is invented; the message is deferred and re-read until the tool reaches completed/error, whose event then lands
reasoning, file, step-start, step-finish, snapshot, patch, agent parts	Dropped	Not normalized into events
Unknown message role	Unsupported/skipped	Counted without semantic inference
SQLite-backed storage	Not read	v0.1 targets the file-based JSON layout only
User correction or rejection	Not inferred	Requires explicit future source evidence

Every emitted event includes opencode.project_id; message-derived events additionally carry opencode.message_id, opencode.role, and the tool name and call ID for tool events.