Capability Matrix
Capability
SubQ Code
Claude Code
Codex CLI
Headless flag
--json / -p
-p
exec --full-auto
Real-time streaming
Yes
Yes
No
System prompt override
Env var
--system-prompt
--instructions
Built-in worktree
Side agents
--worktree
No
Budget cap
No
--max-budget-usd
No
Bare/clean mode
No
--bare
No
Session ID control
Yes
--session-id
Rollout naming
Permission bypass
N/A
bypassPermissions
--full-auto
JSONL output
stdout
stream-json
Disk only
Disk session path
~/.subq/
~/.claude/
~/.codex/sessions/
Adapter Details
SubQ Code Adapter
Pi Agent
subq code --json "$PROMPT"
Real-time JSONL on stdout. Uses SUBQ_SYSTEM_INSTRUCTIONS_FILE for prompt override. Session ID passed directly. Streaming enrichment via Bun.JSONL.parseChunk().
--json
real-time
env override
Claude Code Adapter
Claude
claude -p "$PROMPT" --output-format stream-json --bare --permission-mode bypassPermissions
The --bare flag is critical: it skips hooks, plugins, and CLAUDE.md, ensuring clean evaluation. ANTHROPIC_API_KEY stripped from env to force Max subscription usage. Post-hoc session cross-reference via --session-id.
--bare
stream-json
key stripped
Codex CLI Adapter
Codex
codex exec --full-auto "$PROMPT" --json
No real-time streaming—JSONL written to ~/.codex/sessions/ post-hoc. Use proc.exited (not polling) to detect completion. Phase 7c—added after SubQ and Claude adapters prove the pattern.
proc.exited
disk JSONL
Phase 7c
The --bare flag: Claude Code must run with
--bare during evaluation. Without it, user-installed hooks and CLAUDE.md instructions interfere with enrichment—a user’s custom pre-commit hook could cause a false failure that has nothing to do with the agent’s capabilities.
ProcessSpawner Interface
interface
ProcessSpawner
spawn
(cmd: string[], opts: SpawnOpts) → SpawnedProcess
stdout
ReadableStream<Uint8Array> — incremental consumption
exited
Promise<number> — resolves with exit code
signal
AbortSignal — timeout via AbortController + setTimeout
-----
note
Bun.spawn() not available in Vitest — inject for testability