opencode QA

QA the opencode coding agent itself. This skill maps each QA need to a tested helper script and a deep reference. Every script ships a --self-test that asserts its scenario against the live machine, so the scripts are both the QA tools and their own regression checks.

Verified against opencode v1.17.7 (bun 1.3.12, macOS). Confirm the installed version with opencode --version; the surface is stable but always sanity check a flag with opencode <cmd> --help.

Golden rules (read before running anything)

• READS of the live DB are safe and intended. Investigating sessions (Case D)

only reads ~/.local/share/opencode/opencode.db.

• Anything that SPAWNS opencode (serve, run, the TUI) must use an isolated XDG

sandbox so QA never writes junk sessions into the real DB. The bundled scripts already do this; if you run opencode by hand for QA, set XDG_DATA_HOME / XDG_CONFIG_HOME / XDG_STATE_HOME / XDG_CACHE_HOME to temp dirs first.

• Global text search over the part table is a multi-GB scan. Always scope it

(--session, --recent, or --since). The text script refuses an unbounded scan on purpose.

• The opencode source repo (packages/opencode) tests itself with bun test

and CANNOT run tests from the repo root. See references/testing-harness.md.

Setup

Scripts live next to this file under scripts/. Invoke them from this skill directory (or with their absolute path):

cd <this-skill-dir>                        # .agents/skills/opencode-qa
bash scripts/lib/common.sh --self-check    # confirm the harness + deps

Docker is the default QA surface. Run QA inside a disposable container that has the latest opencode and a copy of your config, with the host untouched: script/agent/qa-docker.sh (see references/docker-qa.md). The local scripts below are the fallback for when Docker is unavailable or on Windows.

common.sh provides the shared harness (DB path, SQL escaping, isolated XDG sandbox, free port, server start/stop, and an EXIT-trap cleanup). It requires opencode, sqlite3, curl, jq, and tmux on PATH.

Router: pick your case

Case A: CLI / terminal works

The canonical scriptable, non-interactive entry is opencode run. JSON mode emits one event per line so you can assert on it.

# stream structured events (types: text, tool_use, step_start, step_finish, reasoning, error)
opencode run "list files in src" --format json
# run a slash command
opencode run --command commit
# resume the last session
opencode run -c "continue"
# target an already-running server instead of booting one
opencode run "explain auth" --attach http://127.0.0.1:4096 -p "$OPENCODE_SERVER_PASSWORD"

Other QA-useful commands: opencode db path, opencode debug paths, opencode session list --format json, opencode models --verbose. Full flag detail in references/cli-commands.md.

Case B: a specific hook, action, or event

opencode publishes lifecycle events over Server-Sent Events at GET /event. Plugins observe the same events via the event hook, so seeing an event on the wire proves a hook would fire.

# prove the SSE plumbing works (isolated server, asserts server.connected)
bash scripts/sse-hook-probe.sh --self-test

# watch a REAL server for a specific event while you trigger an action
bash scripts/sse-hook-probe.sh --attach http://127.0.0.1:4096 \
  --password "$OPENCODE_SERVER_PASSWORD" --directory "$PWD" \
  --event message.part.updated --timeout 30

Trigger an action over HTTP (fire-and-forget so the stream is not blocked):

curl -X POST -u opencode:$OPENCODE_SERVER_PASSWORD -H 'Content-Type: application/json' \
  -d '{"parts":[{"type":"text","text":"say hi"}]}' \
  "http://127.0.0.1:4096/session/<ses_id>/prompt_async?directory=$PWD"

A real prompt needs a configured provider, so run the watch-and-trigger pattern against your real server, not the isolated sandbox. Event-type catalog, the 21 plugin hook points, and how to load a local plugin: references/events-hooks.md. Server start, auth, and routes: references/server-api.md.

Case C: the TUI

bash scripts/tui-smoke.sh --self-test

This launches the TUI under tmux in an isolated sandbox, confirms it renders (capture-pane), confirms send-keys reaches the composer, tears the tmux session down, and verifies the real DB session count is unchanged.

Honest verdict: tmux is fine for SMOKE (did it boot, render, accept a key) but fragile for asserting conversation output (the TUI is a 60fps full-screen app). For real behavior assertions use Case A (opencode run), Case B (server API + SSE), or the TUI control HTTP API (POST /tui/append-prompt, POST /tui/submit-prompt, POST /tui/execute-command). Details and the manual tmux recipe: references/tui-tmux.md.

Case D: investigate sessions in the DB

Read-only against the live SQLite DB. The session table is small (title and id lookups are instant); message text lives in the multi-GB part table, so text search must be scoped.

# by id
bash scripts/db-session-by-id.sh ses_3a4ee6335ffedFB8f76BPU1Eb3
# by title / name (newest first; second arg = limit)
bash scripts/db-session-by-name.sh "auth refactor" 20
# by message text - scope with --session, --recent N, or --since "<window>"
bash scripts/db-session-by-text.sh --session ses_3a4e... "ULTRAWORK"
bash scripts/db-session-by-text.sh --recent 50 "permission denied"
bash scripts/db-session-by-text.sh --since "7 days" --limit 50 "TODO"
# export an entire session as clean JSON
bash scripts/export-roundtrip.sh ses_3a4e... > session.json

Ad hoc queries: opencode db "<SQL>" --format json. Schema, tested query shapes with timings, the legacy message/part vs V2 session_message distinction, and the 25 GB caveat: references/db-investigation.md.

Scripts index

Run any script with --self-test to verify it against the live machine, or -h for usage. DB-read scripts are read-only; serve/sse/tui scripts use an isolated sandbox and clean up on exit.

Risks and caveats

• 25 GB part table: never run an unbounded text scan. Use --session,

--recent, or --since. A naive JOIN ... WHERE session.time_created >= X scans oldest-first and can take ~50s; the scripts use an IN-subquery on the newest sessions (~20ms).

• opencode export writes its banner to STDERR; pipe with 2>/dev/null before

jq or you will get a parse error.

• The server enforces auth only when OPENCODE_SERVER_PASSWORD is set;

otherwise it runs unsecured. Authenticated calls use -u opencode:$PASS. Unauthenticated calls to a secured server return HTTP 401.

• Installed binary vs dev source: cite dev source paths for internals but

verify flags against the installed opencode <cmd> --help.

• Isolation: any QA that spawns opencode must use an isolated XDG sandbox so it

never pollutes the real DB. Prove it by comparing sqlite3 "$(opencode db path)" "SELECT count(*) FROM session" before and after.

• TUI output assertions are fragile; use the API for real assertions.

References

• references/cli-commands.md - every QA-relevant opencode subcommand and flag
• references/db-investigation.md - DB schema, tested queries, the 25 GB caveat
• references/server-api.md - server start, auth, route catalog, /doc
• references/events-hooks.md - SSE endpoints, event types, plugin hooks
• references/tui-tmux.md - tmux recipe, isolation, TUI control API
• references/testing-harness.md - how opencode tests itself (bun test)
• references/sdk.md - the @opencode-ai/sdk client (reference only)
• references/docker-qa.md - run QA in a disposable Docker container (default; local is the fallback)