Adapter Research

Adapter Targets

Summary

MyACE ships 12 target adapters (backend/app/adapters/), each translating the Canonical IR into a real, currently-supported file format for its target framework. Every adapter's output has been verified directly against that target's current live documentation (most recently in a full pass in Aug 2026) — see each entry below for its confirmed format, doc citation, and any open items.

To add a new adapter, see extending.md for the checklist (including registering the new target in ProfileCompileRequest's Literal). See Future Plans below for known candidates that aren't built yet.

---

Fully built and verified adapters

1. Windsurf (Codeium)

backend/app/adapters/windsurf.py

Property	Value
Config format	Markdown with YAML frontmatter (.windsurf/rules/*.md)
Legacy format	.windsurfrules (plain text, numbered rules)
Global path	~/.codeium/windsurf/global_rules.md
Project path	.windsurf/rules/ (modern) or .windsurfrules (legacy)
Rule triggers	always_on, manual, model_decision, glob
Docs	https://docs.windsurf.com/

Translates Canonical IR rule/skill artifacts into .windsurf/rules/*.md files with YAML frontmatter, using the always_on/manual/model_decision trigger modes.

Known limitations: Windsurf was acquired by Cognition AI and rebranded to Devin Desktop — docs.windsurf.com now redirects to docs.devin.ai, and .windsurf/ is documented as a legacy fallback behind the now-preferred .devin/ directory. The adapter still targets the legacy path only, doesn't implement the glob trigger mode, and doesn't enforce the real per-file character limits (12,000 for workspace rule files, 6,000 for the global rules file). See Future Plans.

---

2. Continue.dev

backend/app/adapters/continue_dev.py

Property	Value
Config format	YAML (config.yaml — config.json is deprecated), Markdown rules (.continue/rules/*.md)
Global path	~/.continue/config.yaml
Project path	.continue/rules/, .continue/prompts/
Models	Multiple model roles (chat, edit, autocomplete)
Rules	Markdown files with YAML frontmatter (name, globs, description, alwaysApply) in .continue/rules/
Prompts	Markdown+frontmatter files in .continue/prompts/<name>.md; invokable: true makes one a slash command
Tools	MCP server integration (mcpServers in config.yaml)
Docs	https://docs.continue.dev/

Writes rules to .continue/rules/*.md and prompts to .continue/prompts/<name>.md, and can generate config.yaml model config from model_config artifacts.

Known limitations: prompts aren't yet cross-referenced from a prompts: list in config.yaml — the exact reference schema (a Continue Hub uses: block vs. a plain local file path) hasn't been confirmed with full confidence. See Future Plans.

---

3. Aider

backend/app/adapters/aider.py

Property	Value
Config format	YAML (.aider.conf.yml), Markdown (CONVENTIONS.md)
Global path	~/.aider.conf.yml
Project path	.aider.conf.yml + CONVENTIONS.md
Conventions	CONVENTIONS.md — plain Markdown coding guidelines, loaded via .aider.conf.yml's read: key
Models	Any LLM provider, configured in YAML (model: key)
Docs	https://aider.chat/docs/usage/conventions.html

Maps Canonical IR rule artifacts directly to CONVENTIONS.md and model_config artifacts to .aider.conf.yml's model: key. The simplest adapter in the set, with no known open items.

---

4. Cline

backend/app/adapters/cline.py

Property	Value
Config format	Markdown, no frontmatter
Global path	~/.cline/rules/
Project path	.clinerules (single file) or .clinerules/ (directory)
AGENTS.md fallback	Yes — Cline reads AGENTS.md when no .clinerules/ exists
Docs	https://docs.cline.bot/

Writes .clinerules/*.md files. Per docs.cline.bot/customization/cline-rules, paths is the only real recognized frontmatter field, and rules without frontmatter are always active. Since Canonical IR has no per-artifact path/glob-scoping concept to populate a real paths value with, the adapter emits no frontmatter at all rather than guessed-at fields — the correct default for this project's content anyway.

---

5. Goose

backend/app/adapters/goose.py

Property	Value
Config format	Plain text/Markdown, single file
Global path	~/.config/goose/.goosehints
Project path	.goosehints (merged root-to-leaf with nested/global hints)
Docs	https://goose-docs.ai/docs/guides/context-engineering/using-goosehints/

Extremely simple format — one file, no frontmatter, no per-artifact-type structure. Every Canonical IR artifact type except model_config (which Goose configures via ~/.config/goose/config.yaml on the host rather than a repo file) folds into a headed section of a single .goosehints.

---

6. Amazon Q Developer

backend/app/adapters/amazon_q.py

Property	Value
Config format	Plain Markdown, one file per rule
Project path	{project-root}/.amazonq/rules/*.md
Docs	https://docs.aws.amazon.com/amazonq/latest/qdeveloper-ug/context-project-rules.html

Writes plain Markdown files (no frontmatter) into a fixed directory, auto-loaded as chat context. Simple, clearly documented format backed by AWS's official docs with no ambiguity.

Known limitations: Amazon Q CLI has since added native custom agents as JSON files at .amazonq/cli-agents/{name}.json (tools/permission/ resources fields) and MCP config at .amazonq/mcp.json, with rules now loadable as "agent resources" referenced from that JSON. The adapter's current agent → .amazonq/rules/agent-{name}.md fallback still works as a plain rules file; emitting the more idiomatic native JSON format is an open enhancement. See Future Plans.

---

7. Claude Code

backend/app/adapters/claude_code.py — MyACE's own primary target.

Property	Value
Rules format	Markdown sections merged into a root CLAUDE.md — auto-loaded at session start
Subagents format	Markdown with required YAML frontmatter (name, description) under .claude/agents/*.md — identity comes from the name field, not the filename
Skills format	Markdown with YAML frontmatter (name, description) under .claude/skills/<name>/SKILL.md — loaded on demand, not injected into context up front
Commands format	Markdown with frontmatter (description, optionally argument-hint/allowed-tools/model) under .claude/commands/*.md — legacy but still supported; commands were merged into skills, so .claude/skills/ is the current recommended home for both
Model config	No repo-committed file — via /model, --model, the ANTHROPIC_MODEL env var, or the model field in .claude/settings.json
Docs	sub-agents, slash-commands, model-config

Agents and skills get real name/description frontmatter; skills are written to on-demand .claude/skills/<name>/SKILL.md rather than being inlined into the always-loaded CLAUDE.md, which keeps sessions token-efficient. Workflows go to the still-supported legacy .claude/commands/<name>.md. There is no real model_config target in Claude Code, so that artifact type isn't compiled for this adapter.

---

8. Cursor

backend/app/adapters/cursor.py

Property	Value
Config format	Markdown with YAML frontmatter (.mdc) under .cursor/rules/*.mdc — the only current mechanism
Frontmatter fields	description, globs, alwaysApply — together these determine whether/when a rule loads: alwaysApply: true = Always; a description with alwaysApply: false = Agent Requested (the agent decides whether to pull it in); globs = Auto Attached (loads when a matching file is open)
Docs	cursor.com/docs/rules

Every artifact type is written to a named file, .cursor/rules/<name>.mdc, with real description/globs/alwaysApply frontmatter — alwaysApply: true for rule-type artifacts, false (Agent Requested mode) for skill/agent/workflow-type. The legacy .cursorrules format no longer appears anywhere in current Cursor docs and is not written.

---

9. Codex CLI

backend/app/adapters/codex_cli.py

Property	Value
Rules format	AGENTS.md at the project root
Skills format	Markdown with YAML frontmatter (name, description) under .agents/skills/<name>/SKILL.md — scanned from cwd up to the repo root
Subagents format	TOML files under .codex/agents/<name>.toml (project) or ~/.codex/agents/ (personal) — required fields name, description, developer_instructions
Workflows	No such concept exists in Codex CLI — skills + subagents + MCP are the only customization primitives
Model/provider config	.codex/config.toml — a top-level model = "..." string plus one [model_providers.<id>] table per provider (name, base_url, env_key, …); there is no [models] table
Docs	agents-md, build-skills, subagents, config-reference

Writes AGENTS.md, skills to .agents/skills/<name>/SKILL.md, and subagents as real TOML at .codex/agents/<name>.toml (via a small in-adapter string escaper, since no toml library is imported anywhere in this codebase). Workflows are skipped entirely, since Codex CLI has no equivalent concept. config.toml follows the real schema, picking the first model_config artifact as the active model.

Known limitation: project-scoped .codex/config.toml cannot override provider/auth/profile-selection config — that has to live in ~/.codex/config.toml, so some fields in a compiled project-local file may not take effect regardless of what MyACE generates.

---

10. OpenCode

backend/app/adapters/opencode.py — MyACE's own reference target.

Property	Value
Rules format	AGENTS.md at the project root
Skills format	Markdown with YAML frontmatter (name, description, optional license/compatibility/metadata) under .opencode/skills/<name>/SKILL.md
Agents format	Markdown with frontmatter (description, optional mode/model) under .opencode/agents/*.md
Commands format	Markdown with frontmatter (description) under .opencode/commands/*.md
Model/MCP config	Single merged root opencode.json — provider.<name>.models and mcp keys
Docs	opencode.ai/docs (skills, agents, commands, config)

This is, not coincidentally, the format the starter collections in collections/ are natively authored in — the scanner's parsers and this adapter's translate() are meant to round-trip cleanly against each other. metadata is a genuinely free-form field in OpenCode's own skill schema, which this adapter uses to stash version/priority/tags.

---

11. GitHub Copilot

backend/app/adapters/copilot_cli.py — target name copilot-cli.

Property	Value
Config format	Markdown
Project path	.github/copilot-instructions.md (repo-wide), .github/instructions/*.instructions.md (path-scoped, frontmatter: applyTo glob, excludeAgent)
Scope	One repo-wide file plus one path-scoped file per skill/agent/workflow/model-config artifact
Docs	https://docs.github.com/en/copilot/

Writes the single project-wide .github/copilot-instructions.md file (simple Markdown, no frontmatter) plus path-scoped .github/instructions/*.instructions.md files for every other artifact type. Copilot CLI (the terminal product this adapter targets) reads the same file set as the IDE extension, by design — confirmed via docs.github.com/.../copilot-cli/customize-copilot/add-custom-instructions.

---

12. pi.dev (Pi Coding Agent)

backend/app/adapters/pi_dev.py — target names pi-dev, pi.

Property	Value
Config format	AGENTS.md/CLAUDE.md (Markdown), skills + prompt templates (Markdown + YAML frontmatter), settings.json (JSON)
Project path	AGENTS.md; .pi/skills/<name>/SKILL.md; .pi/prompts/<name>.md; .pi/settings.json
Scope	rule/agent → AGENTS.md sections; skill → per-skill file; workflow → per-prompt-template file; model_config (model:* only) → merged settings.json
Docs	https://pi.dev/docs/latest/usage, /skills, /prompt-templates, /settings

Pi loads AGENTS.md (or CLAUDE.md) hierarchically from the cwd upward, same mechanism as OpenCode/Goose, so rule artifacts fold into one root AGENTS.md. Pi has a real skills system (.pi/skills/<name>/SKILL.md) with the same name/description/license/compatibility/metadata frontmatter shape OpenCode's skill format uses — confirmed directly against /docs/latest/skills. It also has a real slash-command mechanism ("prompt templates" at .pi/prompts/<name>.md, frontmatter description), so workflow artifacts map there rather than folding into AGENTS.md. Pi has no documented subagent/agent concept (no "agents" page anywhere in the docs nav), so agent artifacts fold into AGENTS.md as a labeled section — the same fallback OpenCode/Goose use for artifact types their target has no dedicated file for.

Known limitations: model_config artifacts named model:<name> merge into .pi/settings.json's defaultProvider/defaultModel/enabledModels (confirmed against /docs/latest/settings), but mcp:<name> artifacts are skipped — pi.dev/docs/latest/mcp 404s, and the only MCP config shape found (a ~/.pi/agent/mcp.json / .pi/mcp.json file with a mcpServers object) comes from third-party sources, not pi.dev's own docs. Emitting that file once the format is confirmed against primary docs is an open enhancement. See Future Plans.

---

Future Plans

MyACE doesn't yet ship adapters for every AI coding tool. If you use a framework not listed above, please open a PR, issue, or discussion — the adapter-adding checklist walks through what's involved, and a new adapter is usually a self-contained addition: one file implementing BaseAdapter, registered in backend/app/adapters/__init__.py, plus a Literal update in ProfileCompileRequest.

Known unbuilt-but-viable candidates:

• Zed AI — plain text rule files under .zed/rules/; the format is comparatively immature and worth re-checking against current docs before building. Docs: zed.dev
• CodeGPT — uses the same .cursorrules-compatible format as Cursor, so it likely doesn't need a dedicated adapter — but this hasn't been re-verified against CodeGPT's current docs. Docs: codegpt.co

Open items on existing adapters — also good starting points for a contribution:

• Windsurf rebranded to Devin Desktop; windsurf.py still targets the legacy .windsurf/rules/ path rather than the now-preferred .devin/rules/ — needs a naming decision before switching the primary target.
• Continue.dev prompts aren't yet cross-referenced from config.yaml's prompts: list — the exact reference schema needs confirming.
• Amazon Q Developer could emit the newer native .amazonq/cli-agents/*.json agent format instead of falling back to a plain rules file.
• pi.dev doesn't yet emit an MCP server config file (mcp:* artifacts are skipped) — the format needs confirming against primary pi.dev docs first (/docs/latest/mcp currently 404s); see pi_dev.py's docstring.

Retired or rejected candidates — researched and deliberately not built, so they don't get re-proposed:

• Roo Code — evaluated but not built: its VS Code extension was shut down and its GitHub repo archived on 2026-05-15.
• Sourcegraph Cody — was built (cody.py) and later retired. A documentation audit found its target format (.sourcegraph/*.rule.md) was never a real Cody capability, and separately that Cody Free/Pro were discontinued on 2025-07-23, leaving only Cody Enterprise ($59/user/month). The adapter was deleted and fully deregistered rather than left as dead code — see the cody.py entry in docs/plans/starter-collections-improvements.md for the full audit.

---

Key Insight: AGENTS.md as a Universal Format

Several targets (pi.dev, Cline) natively support AGENTS.md as a configuration format. Since MyACE already compiles to AGENTS.md via the OpenCode and Codex CLI adapters, these targets may already work partially without a dedicated adapter. Adapter work for an AGENTS.md-native target mostly comes down to:

• Target-specific file layouts (directory structures)
• Format-specific features (YAML frontmatter, trigger modes, conditional rules)
• Settings/config files (JSON/YAML configs beyond AGENTS.md)