Skip to content

PLATFORM_SPEC.md — Continue.dev

New to gald3r on this platform? Start at Use gald3r with Continue.dev for a short overview before the full breakdown below.

Authoring status: refresh — re-verified 2026-07-18 against the live docs.continue.dev deep-dive pages listed above, replacing the 2026-06-13 create scaffold. Root cause of the "23 vs 38/39 platform-roster disconnect": Continue's commands were being written to .continue/commands/, a directory Continue's Prompt-file engine never reads — so every gald3r command silently failed to load while every other roster bookkeeping surface reported Continue as "installed". See section 5 and 10 below for the corrected path.

Open-source AI coding extension for VS Code and JetBrains. Context providers, slash commands via .prompt-style Markdown files ("Prompts"), MCP integration, custom models, and native custom sub-agents.

Relationship: Open-source VS Code / JetBrains extension.

Highly configurable. MCP + context providers. Good match for gald3r CRASH primitives, with two real gaps: no native hooks, and no confirmed native Agent Skills support (see sections 4 and 6).

1. Folder Hierarchy

<project-root>/
└── .continue/
    ├── config.yaml                ← primary config (models, context, MCP, etc.)
    ├── rules/
    │   └── <name>.md               ← persistent instructions (rules equivalent)
    ├── prompts/
    │   └── <name>.md               ← Prompt files = custom slash commands (/name)
    ├── agents/
    │   └── <name>.md               ← custom sub-agent definitions
    ├── skills/
    │   └── <name>/SKILL.md         ← third-party-convention Agent Skills location
    │                                  (NOT confirmed as a Continue-core feature —
    │                                  see section 4)
    └── mcpServers/
        └── <name>-mcp.yaml         ← one MCP server per file (native format)
        └── mcp.json                ← Continue also auto-recognizes Claude
                                        Desktop/Cursor-style JSON MCP config files
                                        dropped directly into this folder

Config file: .continue/config.yaml (project-scope). No documented global ~/.continue/config.yaml auto-discovery was re-confirmed this scan — do not assume one without a fresh crawl if that distinction becomes load-bearing.

2. CRASH Primitive Support

Primitive Support Notes
Hooks ❌ No lifecycle-hook/event-bus documentation found in this scan.
Rules ✅ .continue/rules/*.md, loaded automatically, lexicographic order. NOT AGENTS.md — see section 3.
Skills ⚠️ Third-party marketplaces (LobeHub, mdskills.ai) document a .continue/skills/ / ~/.continue/skills/ convention, but Continue's own tracker has an OPEN "Are there any plans to support skills in Continue?" issue (#9216) as of this scan — native first-party support is unconfirmed.
Commands ✅ Prompt files (.continue/prompts/*.md, invokable: true frontmatter) — invoked /name.
MCP ✅ .continue/mcpServers/<name>-mcp.yaml, OR a plain JSON file (e.g. mcp.json) dropped in the same folder, OR inline mcpServers: in config.yaml.
Agents ✅ .continue/agents/*.md — Markdown + YAML frontmatter, one file per custom agent.

3. Instruction Files (Rules Equivalent)

Continue does NOT natively read a root AGENTS.md. This was the single most consequential correction of this refresh — the prior scaffold assumed AGENTS.md parity with gald3r's other platforms; it does not exist.

  • .continue/rules/*.md — the real, documented, native mechanism. "You can create project-specific rules by adding a .continue/rules folder to the root of your project and adding new rule files." Each rule file is Markdown with optional YAML frontmatter (name, globs, alwaysApply, description). Rules load automatically for every Agent/Chat/Edit request and are concatenated into the system message. Rule files load in lexicographic order (number-prefix to control ordering).
  • AGENTS.md root-file support is a still-open, unimplemented feature request: continuedev/continue#6716 ("Add Support for Agent Rules Standard via Root AGENTS.md File"), opened 2025-07-19. As of this scan, docs.continue.dev/customize/deep-dives/rules makes zero mention of AGENTS.md.
  • gald3r therefore ships its rule content as .continue/rules/*.md (see layout_map.yaml's continue.rules entry) and does not attempt a root AGENTS.md/root_instructions projection for this platform (see layout_map.yaml's root_instructions module comment — continue is now in the deliberately-excluded list, alongside claude/cursor/copilot/aider/etc.).

4. Skills

⚠️ Unconfirmed as a first-party Continue feature.

  • Third-party skill marketplaces (LobeHub's terminalskills-skills-continue-dev, mdskills.ai) document skills living at .continue/skills/ (project) or ~/.continue/skills/ (global), using the standard SKILL.md + YAML frontmatter (name/description) shape gald3r already ships for Claude/Cursor/ZCode/etc.
  • Continue's own GitHub tracker has an open request, continuedev/continue#9216 ("Are there any plans to support skills in Continue?"), suggesting native first-party Skills discovery may not yet be shipped as of this scan.
  • gald3r keeps the .continue/skills/ dir-per-skill layout mapping (matches the marketplace-documented convention and costs nothing if unused) but the matrix cell stays ⚠️ (partial/unconfirmed) rather than ✅ until a live install test proves discovery.

5. Commands (Slash Commands / Prompt Files)

✅ Native, via Prompt files.

  • Directory: .continue/prompts/*.md (project scope). This is the corrected path — the prior scaffold (and gald3r's shipped layout_map.yaml entry, until pointed at .continue/commands/, which Continue never reads.
  • Format: Markdown with YAML frontmatter (name, description, invokable: true). Setting invokable: true makes the file available as a / slash command in Chat, Plan, and Agent mode. Files can also use full prompt templating (referencing files, URLs, highlighted code).
  • Invocation: /<name> (the frontmatter name, not necessarily the filename).
  • Prompts can also be declared inline in config.yaml via a prompts: block (e.g. prompts: - uses: supabase/create-functions) — a distribution/reuse mechanism, not the primary per-project authoring path gald3r targets.

6. Hooks

❌ Not supported.

No lifecycle-hook / event-bus documentation was found across customize/overview, customize/deep-dives/configuration, or the other deep-dive pages crawled this scan. Do not fabricate a hook config surface for Continue.

7. MCP

✅ Native, two supported shapes.

  • Per-server YAML files (the documented native format): create .continue/mcpServers/ at the workspace root, then one <name>-mcp.yaml file per server: ```yaml name: Playwright mcpServer version: 0.0.1 schema: v1 mcpServers:
    • name: Browser search command: npx args: ["@playwright/mcp@latest"] ```
  • Plain JSON config auto-recognition: "You can also place JSON config files (from Claude Desktop, Cursor, etc.) directly in .continue/mcpServers/ and Continue will automatically recognize them, such as .continue/mcpServers/mcp.json." This is the shape gald3r's generic minimal-JSON MCP registration (generate.py's _emit_mcp) already produces — corrected layout_map.yaml target: .continue/mcpServers/mcp.json (was .continue/.mcp.json, a path Continue never reads).
  • Inline in config.yaml: mcpServers: can also be declared directly inside the main config.yaml (a Block-composition alternative to the per-file form above) — not gald3r's install target, documented here for completeness.
  • MCP is usable only in Agent mode.

8. Agents

✅ Native custom sub-agents.

  • Directory: .continue/agents/*.md (project scope, version-controlled and shared with the team).
  • Format: Markdown file with YAML frontmatter configuring the agent, plus a Markdown body that becomes the agent's system-prompt/instructions. The filename becomes the agent's name.

9. Instruction Files — see section 3 (folded in; Continue has no separate

"instruction file" concept distinct from Rules).

10. gald3r Installation

Install Path (corrected)

<project-root>/
└── .continue/
    ├── rules/<name>.md            ← g-rl-* rules
    ├── prompts/<name>.md          ← g-* commands (CORRECTED from commands/)
    ├── agents/<name>.md           ← g-agnt-* agents
    ├── skills/<name>/SKILL.md     ← g-skl-* skills (dir-per-skill, unconfirmed native discovery)
    └── mcpServers/mcp.json        ← gald3r MCP registration (CORRECTED from .continue/.mcp.json)

gald3r platform install continue --into <target-dir>

No root AGENTS.md/root-instructions file is written for Continue (see section 3).

Verification Checklist

  • [x] Rules file location confirmed: .continue/rules/*.md (2026-07-18 docs crawl)
  • [x] Commands (Prompt files) location confirmed: .continue/prompts/*.md (2026-07-18 docs crawl)
  • [x] Agents location confirmed: .continue/agents/*.md (2026-07-18 docs crawl)
  • [x] MCP config syntax verified: .continue/mcpServers/<name>-mcp.yaml or auto-recognized JSON (2026-07-18 docs crawl)
  • [x] Root AGENTS.md NOT natively read — confirmed absent, do not project one (2026-07-18)
  • [ ] Skills discovery path — third-party-documented only, not first-party-confirmed; re-check on next scan
  • [ ] Full live gald3r platform install continue + open-in-Continue smoke test

11. Official Docs

Scaffold created 2026-06-13. Refreshed 2026-07-18 against the live docs listed above — see section headers for per-capability citations. Re-run @g-platform-scan-docs continue on the next crawl_max_age_days (14) cycle.

Reference page — canonical source: g-skl-platform-continue/PLATFORM_SPEC.md in the engine repo, generated by scripts/generate_platform_support_docs.py. Pages marked curated are hand-assessed and not yet live-verified.