aider

aider is a terminal pair-programming agent driven by litellm. It differs from agedum's other harnesses in two ways that shape how agedum drives it:

• No native instruction discovery. aider reads neither AGENTS.md nor a CONVENTIONS.md on its own. Its one channel for standing context is --read FILE (a read-only file added to the chat) — the documented conventions mechanism. So agedum injects each scope's AGENTS.md as a --read argument, since aider reads no AGENTS.md natively.
• No skills mechanism. aider has nothing to load a skills tree into, so agedum does not inject skills for it (there is no SKILL.aider.md overlay). A project .agents/skills/ shows up in --dry-run as (not injected).

Wrapper resolution

Source	Injected at
project AGENTS.md	--read <compiled path> (appended arg)
project .agents/skills/	(not injected — aider has no skills mechanism)
global ~/.config/agents/AGENTS.md (+ optional AGENTS.aider.md overlay)	--read <compiled path> (appended arg)
global ~/.config/agents/skills/	(not injected — aider has no skills mechanism)

• Instructions — agedum compiles each scope's AGENTS.md to a file under the throwaway launch dir and appends --read <path> (project first, then global). The global one is the base merged with an optional AGENTS.aider.md overlay. Because the bwrap launch binds the whole real filesystem (--dev-bind / /), those paths resolve inside the namespace without a dedicated bind — Plan.binds stays empty, like kimi's agent file.
• extra_args: the --read flags above (one per scope that has an AGENTS.md).

agedum --wrapper aider -- aider                # drive aider with the same source
agedum --wrapper aider --dry-run -- aider      # show what would be injected

Wrapper mode does not disable git

Wrapper mode runs your literal command, so it does not append --no-git. agedum's launch namespace shares the real .git, and aider auto-commits by default — so to keep aider from committing into the real repo, pass --no-git yourself (agedum --wrapper aider -- aider --no-git) or use provider mode, which disables git by default (below).

Provider config

aider's knobs become appended CLI flags (like Cline and kimi), and CLI flags override any on-disk .aider.conf.yml:

{
  "harness": "aider",
  "slug": "aider-deepseek",
  "secretEnv": "DEEPSEEK_API_KEY",
  "config": {
    "model": "deepseek/deepseek-chat",
    "reasoningEffort": "high"
  }
}

config key	Appended
model	--model <model>
weakModel	--weak-model <model> (commit messages / history summarization)
editorModel	--editor-model <model> (the model used to apply edits)
reasoningEffort	--reasoning-effort <level>
git	default false → --no-git. true opts back into aider's git integration
autoCommits	when git is on, false → --no-auto-commits
yesAlways	--yes-always (true) — auto-confirm; useful with --run
baseUrl	sets OPENAI_API_BASE (env), for an OpenAI-compatible custom endpoint

The config above launches aider --model deepseek/deepseek-chat --reasoning-effort high --no-git.

Git integration is disabled by default. agedum's launch namespace shares the project's real, shared .git (see internals), and aider's default is --git with --auto-commits — so left alone it would commit into the real repo. agedum therefore appends --no-git unless the config sets git: true. With git re-enabled you can still suppress commits with autoCommits: false (→ --no-auto-commits). Enabling git in the shared namespace is a deliberate, hazardous opt-in.

The API key rides the environment. aider resolves credentials through litellm, which reads provider-specific environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY, DEEPSEEK_API_KEY, …). agedum exports secretEnv (and any requiredEnv) into the child under its own name via the normal required-env path, so no key flag is appended and no secret lands in argv. Set secretEnv to the variable litellm expects for your model's provider.

Custom endpoints. For an OpenAI-compatible endpoint, set baseUrl (→ OPENAI_API_BASE), pick an openai/<name> model, and put the key in OPENAI_API_KEY:

{
  "harness": "aider",
  "slug": "aider-local",
  "secretEnv": "OPENAI_API_KEY",
  "config": { "model": "openai/local-model", "baseUrl": "https://my.host/v1" }
}

--prompt/--run. aider's --message "<text>" runs a single message and exits (it disables chat mode), which is --run (aider --message "<text>"). aider has no "seed then stay interactive" mode, so --prompt is a fail-loud ProviderError — condash then falls back to spawn-and-type. See the prompt-seeding table.