Host support: Aider — explicit read-only instruction loading and configuration-safe content installation

[ScriptedAlchemy]

Scope / research evidence

P2 feature expansion, limited initially to verified content and command configuration. Research date 2026-09-06; Agent Bundle baseline 5fc8de37c9695b816d55bdcf18a6bcb1dfee69f5. Official Aider docs were read; no native session or model invocation was run.

Ponytail lists Aider among instruction-file consumers. The documented supported mechanism is explicit read-only context, not proof that every file named AGENTS.md is auto-discovered. No general native plugin-package manifest, arbitrary MCP-client registration, or full lifecycle-hook protocol was established by the official references below. Do not invent those APIs or advertise an unofficial fork/bridge as stock Aider support.

Primary documentation and configuration references

• Coding conventions and read-only context
• YAML configuration schema/options, discovery and precedence
• CLI/environment configuration overview
• In-chat command reference
• Lint/test command integration
• Environment-file and credential configuration
• Git behavior and integration
• Official source

Pin actual Aider/package version and parser behavior before shipping an automated installer. These references establish a small host tier, not a negative assertion that no future native extension capability can exist.

Proposed first projection

Use the current canonical content renderer to produce a compact authored instruction document and supporting files. Keep a small plugin small; it does not need an MCP server, JSX conversion, Python extension, or persistent process.

consumer-project/
  .aider.conf.yml                  # optional receipt-owned read-list patch
  agent-instructions/review.md
  agent-instructions/resources/...

Immediate operator use requires no config mutation:

aider --read ./agent-instructions/review.md

Inside a session use the native /read-only command; the conventions guide also illustrates its /read shorthand. Do not use /add for guidance intended to remain outside the editable target set.

Persistent configuration can extend the user's existing read list:

read:
  - AGENTS.md
  - agent-instructions/review.md

Only include AGENTS.md when it exists and the operator intends it as context. Do not overwrite that file with a host-generated substitute. A references link is not proof that all supporting content is automatically loaded; preserve explicit loading semantics and test the intended instruction set.

Scope, precedence and update behavior

Aider reads YAML from home, repository root and current directory, with later files taking priority. --config selects a single configuration rather than merging the normal chain. Consequently a generated --config wrapper can accidentally hide the user's model, permission or read-list settings; it must not be the default installation shortcut.

Normalize scalar/list forms of read without dropping unrelated entries or duplicating a previously installed file. Resolve relative paths in the host's real working/configuration context, including nested execution and paths with spaces. Editing or removing a disk config does not automatically revoke content already loaded into a running conversation; document when a new session is necessary rather than claim runtime hot-unload.

Aider can create commits and manages its own history/configuration. An Agent Bundle install is a file/config binding operation, not authorization to commit user changes, modify model settings, change shell environment or purge .aider histories. Use receipt ownership and refuse to clobber edited/foreign content.

Executable integration boundary

Aider documents lint and test commands that feed results back into its editing loop. Lint commands receive file paths; test commands have their own invocation shape, with optional automatic testing after edits. These are not pre-tool authorization hooks, and they do not register canonical tools with an MCP client.

For example, an explicitly authored test integration might use:

test-cmd: "node ./scripts/project-check.mjs"
auto-test: false

Do not turn auto-test on, replace an existing command, or append an arbitrary shell pipeline just to fit a plugin. Initially document explicit execution of a compiled CLI as an operator command. If canonical hook/MCP declarations target this profile, report unsupported capability until an authoritative host contract and native proof justify a mapping.

Acceptance

• A content-only artifact installs/relocates without any MCP, Flight, SQLite or Python plugin runtime.
• Native read-only loading includes the intended guidance without making it an editable target.
• Existing AGENTS.md and read-list entries are preserved; scalar/list, duplicate and missing-file cases are explicit.
• Home/repository/cwd and --config precedence are tested using the real native parser/CLI.
• Nested working directories, Windows/POSIX paths, spaces and relocation do not introduce source-checkout paths.
• If lint/test integration is selected, actual argv, exit status and output behavior are tested without claiming deterministic tool blocking.
• Unsupported plugin/hook/MCP capabilities are reported honestly; no implicit third-party bridge is installed.
• Update/uninstall removes only owned bindings/files and leaves user configuration, histories, secrets, git changes and foreign instructions intact.
• Composite required/forbidden instruction sharing is tested under Composite conformance: test cross-host auto-discovery and precedence, not only file collisions #651; static process absence follows Small-plugin acceptance: skills and plain hooks without an unnecessary MCP/RSC runtime or empty Workbench #652.
• Record native loading/configuration proof separately from any optional authenticated model-behavior check; no guarantee that prose is a security policy.
• English/Chinese install docs link the exact host references and distinguish content support from full native plugin support.

Non-goals

No monkeypatching Aider internals, unofficial MCP bridge deployment, new Python package manager, model/provider configuration, automated commits, or general agent orchestration. Reuse existing content, manifest and receipt machinery instead of inventing aider-plugin.json or aider plugin install.

[ScriptedAlchemy]

Owner decision (2026-09-06): Aider support is no longer required. No implementation or compatibility claim should be added.