Skip to main content

Codegen agents

Eva Board's autonomous loop shells out to a coding-agent CLI to actually edit code, verify diffs, and produce review verdicts. The same CLI handles all three roles (implementer, verifier, reviewer) — there is no separate LLM provider to configure.

The agent is pluggable. Claude Code is the default; you can swap in any CLI that reads a prompt from stdin and makes file edits in its working directory.

Built-in presets​

Pick a preset in Settings → Coding agent in the app. Each user can pick their own — server-side env defaults are only used when a user hasn't selected one.

PresetCODEGEN_AGENTNotes
Claude Codeclaude-codeDefault. Best long-context reasoning today; needs the claude CLI on PATH and claude login.
Codex CLIgenericOpenAI Codex CLI; pair with CODEGEN_COMMAND=codex and CODEGEN_ARGS=--auto-approve.
AidergenericPair with CODEGEN_COMMAND=aider and CODEGEN_ARGS=--yes,--no-stream,--message-file,-.
OpenHandsgenericPair with CODEGEN_COMMAND=openhands and any non-interactive flags your install requires.
ClinegenericWrap the Cline CLI shim; pair with CODEGEN_COMMAND=cline.
CustomgenericAny CLI that reads a prompt from stdin and edits files in its working directory.

The wrappers live in the standalone teslashibe/codegen-go module — see generic.go and claude.go.

Claude Code (default)​

npm install -g @anthropic-ai/claude-code
claude login    # opens a browser for OAuth

# In Eva Board's environment:
CODEGEN_AGENT=claude-code
# Optional model override; defaults to the Claude Code default if empty.
CODEGEN_MODEL=claude-sonnet-4-5

The claude binary must be on the API process's PATH. Under systemd that means setting Environment=PATH=... (see Self-hosting → Sample systemd unit) or symlinking into /usr/local/bin.

Generic CLI​

Any CLI that reads a prompt from stdin and makes file edits in its working directory works.

CODEGEN_AGENT=generic
CODEGEN_COMMAND=/usr/local/bin/codex
CODEGEN_ARGS=--auto-approve,--model,gpt-4o

Examples:

# OpenAI Codex CLI
CODEGEN_AGENT=generic
CODEGEN_COMMAND=codex
CODEGEN_ARGS=--auto-approve

# Aider
CODEGEN_AGENT=generic
CODEGEN_COMMAND=aider
CODEGEN_ARGS=--yes,--no-stream,--message-file,-

Aider's --message-file - flag tells it to read the prompt from stdin, which is what the wrapper supplies. Tune flags so the CLI runs non-interactively and exits when done.

Per-user overrides​

Server-side CODEGEN_AGENT, CODEGEN_COMMAND, and CODEGEN_ARGS act as defaults for users who have not picked an agent in the Settings UI. Once a user selects a preset (Claude Code, Codex, Aider, OpenHands, Cline, or Custom) those values are persisted in board_settings and override the env defaults for that user.

The precedence rule lives in backend/internal/board/cards_handler.go (resolveCodegenAgent).

Tunables​

VariableDefaultDescription
CODEGEN_TIMEOUT30mPer-invocation timeout (Go duration, e.g. 45m).
CODEGEN_MAX_OUTPUT_BYTES10485760Cap on captured combined stdout/stderr (10 MiB). Negative disables the cap.

Bump CODEGEN_TIMEOUT for large refactors; the agent will get killed mid-run if a single invocation exceeds it, which counts as a failed verification iteration.

What the agent sees​

Each invocation runs in the per-card git worktree at <repo>/../worktrees/<shortID> with the prompt on stdin. The prompt includes:

  • The card title, description, and acceptance criteria
  • The current diff against the base branch (after the first iteration)
  • Any verifier-failure feedback from the previous iteration
  • Any reviewer-suggested changes from the previous review cycle
  • Any user-supplied feedback queued via SubmitFeedback

The agent has full read/write access to the worktree and can run shell commands within it. It does not have access to the host filesystem outside the worktree — the wrapper sets chdir and that's the only sandboxing primitive Eva Board relies on. Run untrusted prompts in a container or VM if that matters to you.