Aider agent harness

Capability: capability:platform

Harness: Aider (aider.chat)

Aider is an open-source (Apache 2.0 licensed), model-agnostic terminal pair programming agent harness. This guide documents how Aider operates as an agent harness for Apache Magpie for #317. Grounding is established by RFC-AI-0004 Principle 3 (Vendor Neutrality), which guarantees that Magpie adopters can drive all framework workflows using fully open-source, non-proprietary agent harnesses.

Aider provides an open-source execution stack with:

1. Model independence: Native connectivity to Anthropic Claude, OpenAI GPT, DeepSeek, Google Gemini, OpenRouter, and local open-weight inference runners (Ollama, llama.cpp, vLLM).
2. Git-native pairing: Tight integration with git working trees and incremental file editing.
3. Multi-model tiering: Built-in --architect mode splitting reasoning and code-editing responsibilities across specialized model tiers.

Harness contract

Invoke a Magpie skill

After running /magpie-setup to adopt the repository, the canonical .agents/skills/ links are active in your working tree. Aider does not discover skills on its own; load the one you need with /read-only or --read (see below).

Interactive terminal session

Start an interactive session within the adopted repository:

# Launch Aider with filtered environment variables and disabled auto-commits
source <framework>/tools/agent-isolation/agent-iso.sh
agent-iso aider --model <model> --no-auto-commits --no-dirty-commits

Progressive skill disclosure

To prevent context-window exhaustion across Magpie's 70+ workflows, load only the specific skill required for the current task:

/read-only .agents/skills/magpie-security-issue-triage/SKILL.md

Inside the session, prompt Aider to execute the workflow:

Follow the procedure in the loaded SKILL.md to triage inbound report #<issue-number>.
Adhere strictly to proposal-then-confirm discipline before applying any tracker mutations.

Aider reads the loaded SKILL.md, checks prerequisites, and presents proposed diagnostic steps before taking action.

Headless or automated execution

For non-interactive triage passes, automated sweeps, or headless scripting:

# Headless run using an explicit instruction text (-m) and target skill file
aider --model <model> \
  --read .agents/skills/magpie-list-skills/SKILL.md \
  --message "Execute the loaded list-skills procedure and summarize available workflows." \
  --no-auto-commits

Configuration and repository instructions

Repository configuration (.aider.conf.yml)

Configure repository defaults in .aider.conf.yml at the repository root to ensure all team members operate under Magpie's safety baseline:

# .aider.conf.yml
# 1. Enforce Human-in-the-loop confirmation
auto-commits: false
dirty-commits: false

# 2. Ingest repository safety instructions
read:
  - AGENTS.md

# 3. Model parameters (optional adopter defaults)
# model: <model>
# edit-format: diff

Protecting private and local state (.aiderignore)

Aider automatically constructs a repository map (repomap) to track codebase structure. To prevent Aider from reading private adopter tokens, test fixtures, or local overrides into the prompt context, configure .aiderignore at the repository root:

# .aiderignore
.env*
*.key
*.pem
*.token

Tool bridges and subshell execution

Magpie skills execute deterministic operations via language-agnostic scripts under tools/ (e.g. tools/cve-tool-vulnogram/, tools/github/, tools/privacy-llm/).

When a skill requires running a tool command:

• The operator or model proposes the subshell command (e.g. gh issue view <issue-number> --repo <tracker>).
• In an interactive session, execute the command directly via /run <cmd> or !<cmd>.
• Actions execute in the local project environment following standard subshell semantics.

Human-in-the-loop and git confirmation

Magpie enforces strict Human-in-the-Loop principles: no destructive action or git state mutation occurs without explicit human approval.

aider --no-auto-commits --no-dirty-commits

auto-commits: false
dirty-commits: false

When operating with auto-commits disabled:

• Diff review: Review proposed file changes using git diff or in-session /diff before staging.
• Explicit commit gating: Use /commit only after confirming the change satisfies review criteria.
• Write-access discipline: Outbound communications, issue state changes (gh issue close), and remote pushes (git push) must always remain gated on explicit human confirmation.

Multi-model routing and model floors

Aider supports multi-provider model routing, allowing adopters to calibrate model power against task complexity per docs/mode-economics.md.

Architect and editor model split

For complex triage and vulnerability assessment, use Aider's --architect mode to separate high-level reasoning from code editing:

# Reasoning model handles analysis; fast editor applies diffs
aider --architect \
  --model <reasoning-model> \
  --editor-model <editor-model> \
  --no-auto-commits

Local LLM connectivity (Ollama / vLLM)

For sovereign or air-gapped deployments where data cannot leave the local host (see docs/adapters/local-llm.md):

# Connect Aider to a local Ollama instance
export OLLAMA_API_BASE=http://127.0.0.1:11434
aider --model ollama_chat/<model> \
  --no-auto-commits

Clean-environment wrapper and isolation

To run Aider under Magpie's standard credential isolation policy:

source <framework>/tools/agent-isolation/agent-iso.sh
agent-iso aider --model <model> --no-auto-commits

The agent-iso launcher scrubs ambient cloud tokens while preserving local developer tooling (git, uv, gh, aider).

Verify

Verify that the Aider harness wiring conforms to framework standards:

# 1. Verify skill discovery topology
PYTHONUTF8=1 uv run --project tools/symlink-lint symlink-lint

# 2. Validate skill and tool metadata
PYTHONUTF8=1 uv run --project tools/skill-and-tool-validator --group dev skill-and-tool-validate

# 3. Check vendor neutrality score
PYTHONUTF8=1 uv run --project tools/vendor-neutrality-score vendor-neutrality-score

# 4. Check documentation table of contents and formatting
uv run prek run doctoc --all-files

See also

• docs/rfcs/RFC-AI-0004.md — normative principles for vendor neutrality and open-source harnesses.
• docs/vendor-neutrality.md — framework vendor neutrality index across agent harnesses.
• docs/adapters/local-llm.md — running open-weight models via local inference engines.
• docs/adapters/goose.md — Block's Goose agent harness guide.
• docs/adapters/add-a-harness.md — step-by-step guide for integrating agent harnesses.
• tools/agent-isolation/README.md — clean-environment launcher.
• Aider documentation — official Aider installation and usage guides.