@opengroundplan/core 0.1.1

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Groundplan contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,26 @@
+# @opengroundplan/core
+
+The Groundplan engine, and the catalog it operates on.
+
+Most consumers want [`groundplan`](https://www.npmjs.com/package/groundplan) instead.
+This package is the programmatic API, and it carries the data the CLI needs: `catalog/`,
+`templates/`, `adapters/`, and `registry/` ship inside it.
+
+```ts
+import { loadConfig, buildPlan, apply } from '@opengroundplan/core';
+```
+
+`loadConfig()` locates that data by walking up from its own module until it finds a directory
+holding all four. `GROUNDPLAN_HOME` overrides it; every other setting is environment-driven the
+same way — see `config.ts`, which is the only place a default value is defined.
+
+Two things worth knowing before writing a template:
+
+- Substitution is flat — `{{ name | filter }}` in contents, `__name__` in paths. Templates are
+  data and never execute.
+- A file that must land as `.gitignore` or `.npmrc` is stored undotted, because npm strips
+  those names from any package tarball. The engine restores the dot.
+
+Full documentation is in the [repository](https://github.com/groundplan/groundplan#readme).
+
+MIT licensed.

package/adapters/README.md

@@ -0,0 +1,155 @@
+# Adapters
+
+One directory per AI coding tool. Adapters are the layer that makes a generated repo work
+with every tool, not just the one you happened to scaffold with.
+
+## The governing rule
+
+`AGENTS.md` is the single source of truth. Every file an adapter emits is either:
+
+- a **wrapper** — imports or points at `AGENTS.md` and adds nothing but a pointer, or
+- a **delta** — genuinely tool-specific content that has no place in `AGENTS.md`, or
+- **settings** — machine-readable config, never prose.
+
+An adapter that restates canonical content is a bug. Duplication causes drift, and drifted
+agent instructions are worse than none. The rule is enforced mechanically: **an adapter with
+`readsAgentsMd: true` may not emit any output with `kind: instructions` or `mode: wrapper`.**
+
+## Capability matrix
+
+Verified against each vendor's primary documentation in July 2026. `AGENTS.md` means the tool
+discovers the file itself, with no configuration.
+
+| Tool | AGENTS.md | instructions | nested | glob rules | slash cmds | subagents | hooks | skills | MCP |
+|---|:--:|:--:|:--:|:--:|:--:|:--:|:--:|:--:|:--:|
+| `amp` | native | ✅ | ✅ | — | — | — | ✅ | ✅ | ✅ |
+| `cline` | native | ✅ | — | ✅ | — | — | — | — | — |
+| `codex` | native | ✅ | ✅ | — | — | — | — | — | ✅ |
+| `cursor` | native | ✅ | ✅ | ✅ | — | — | — | — | — |
+| `github-copilot` | native | ✅ | ✅ | ✅ | ✅ | — | — | — | ✅ |
+| `opencode` | native | ✅ | ✅ | — | — | — | — | — | ✅ |
+| `roo-code` | native | ✅ | — | ✅ | — | — | — | — | — |
+| `zed` | native | ✅ | — | — | — | — | — | — | — |
+| `aider` | config | ✅ | — | — | — | — | — | — | — |
+| `claude-code` | wrapper | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| `continue` | wrapper | ✅ | — | ✅ | — | — | — | — | — |
+| `gemini-cli` | wrapper | ✅ | ✅ | ✅ | ✅ | — | — | — | ✅ |
+| `windsurf` | wrapper | ✅ | — | ✅ | ✅ | — | — | — | — |
+
+**AGENTS.md column:**
+- `native` — reads it with no setup. Adapter emits **no** instruction file.
+- `config` — does not discover it, but can be pointed at it by config. Adapter emits config only, still zero duplicated prose. (`aider` uses `.aider.conf.yml` `read:`.)
+- `wrapper` — needs a real file. Adapter emits the thinnest possible pointer.
+
+A `—` means the tool has no such mechanism, or Groundplan could not verify one from a primary
+source. Adapters skip what their tool cannot express rather than writing dead files.
+
+## What each adapter emits
+
+| Tool | Instruction file | Rules | Commands | Other |
+|---|---|---|---|---|
+| `amp` | none | — | — | `.agents/skills/<id>/SKILL.md` |
+| `cline` | none | `.clinerules/<id>.md` | — | — |
+| `codex` | none | — | — | `.codex/config.toml` (MCP only) |
+| `cursor` | none | `.cursor/rules/<id>.mdc` | — | — |
+| `github-copilot` | none | `.github/instructions/<id>.instructions.md` | `.github/prompts/<id>.prompt.md` | — |
+| `opencode` | none | — | — | `opencode.json` |
+| `roo-code` | none | `.roo/rules/<id>.md` | — | — |
+| `zed` | none | — | — | — |
+| `aider` | none | — | — | `.aider.conf.yml` |
+| `claude-code` | `CLAUDE.md` | `.claude/rules/<id>.md` | `.claude/commands/<id>.md` | `.claude/{settings.json,skills,agents}`, `.mcp.json` |
+| `continue` | `.continue/rules/00-agents-md.md` | `.continue/rules/<id>.md` | — | — |
+| `gemini-cli` | `GEMINI.md` | — | `.gemini/commands/<id>.toml` | — |
+| `windsurf` | `.windsurf/rules/agents-md.md` | `.windsurf/rules/<id>.md` | `.windsurf/workflows/<id>.md` | — |
+
+### Deliberate omissions
+
+These files are **not** emitted, on purpose:
+
+- `.github/copilot-instructions.md` — Copilot reads root `AGENTS.md` natively; this would duplicate it.
+- `CONVENTIONS.md` — aider's `read:` points at `AGENTS.md` directly, so the copy is unnecessary.
+- `.rules` (Zed) — it sits *ahead* of `AGENTS.md` in Zed's precedence list, so emitting one would shadow the source of truth.
+- `.roorules`, `.windsurfrules`, `.clinerules` (single-file forms) — the directory form wins in each tool; emitting both creates two competing sources.
+- Any `context.fileName` override for Gemini CLI — a visible `GEMINI.md` pointer beats invisible config.
+
+## Shared templates
+
+`_shared/` holds the tool-neutral files every adapter's tool can read, because they are plain
+markdown at fixed paths:
+
+| Template | Renders to | Role |
+|---|---|---|
+| `delegation.md.tmpl` | `.agent/DELEGATION.md` | Routing table from `agent-kit`'s `delegation` field |
+| `agent/ACTIVE.md.tmpl` | `.agent/ACTIVE.md` | What is in flight right now |
+| `agent/HANDOFF.md.tmpl` | `.agent/HANDOFF.md` | State needed to resume in a *different* tool |
+| `agent/DECISIONS.md.tmpl` | `.agent/DECISIONS.md` | Append-only; why things are the way they are |
+| `agent/DONE.md.tmpl` | `.agent/DONE.md` | Append-only execution log |
+
+These are written once per repo, not once per adapter. Claude Code can start a task, Cursor
+can continue it, and Codex can finish it without sharing a vendor session.
+
+## Template variable contract
+
+The engine (`packages/core/src/tokens.ts`) does **flat substitution only** — no loops, no
+conditionals — and throws `UnknownTokenError` on any token it cannot resolve. Anything
+list-shaped must therefore arrive pre-rendered as a single string.
+
+`{{ name }}` and `{{ name | filter }}` substitute into contents; `__name__` substitutes into
+paths. Filters: `pascal`, `camel`, `kebab`, `snake`, `upperSnake`, `dot`, `title`, `lower`,
+`upper`, `json`.
+
+**Project scope** — available to every template:
+
+| Variable | Source | Shape |
+|---|---|---|
+| `solutionName` | interview | raw string, filtered at use site |
+| `agentsMdPath` | config | path to the canonical file, normally `AGENTS.md` |
+| `buildCommand` `testCommand` `lintCommand` | `agent-kit.commands` | single shell command |
+| `protectedPathsDenyJson` | `agent-kit.protectedPaths` | pre-rendered JSON array |
+| `delegationTable` | `agent-kit.delegation` | pre-rendered markdown table |
+| `today` | engine | ISO date |
+
+**Registry-item scope** — available only to `rule` / `command` / `agent` / `skill` templates,
+one render per item:
+
+| Variable | Shape |
+|---|---|
+| `itemId` | kebab-case identifier |
+| `itemName` | human title |
+| `itemDescription` | one line |
+| `itemBody` | the item's markdown body |
+| `itemGlobsCsv` | `src/**/*.ts,lib/**/*.ts` — comma-separated, for Cursor/Copilot/Cline/Windsurf/Continue |
+| `itemGlobsYamlList` | pre-indented YAML block sequence, for Claude Code's `paths:` |
+| `itemTools` | comma-separated tool allowlist |
+| `itemArgumentHint` | slash-command autocomplete hint |
+
+### Token collisions
+
+Some tools use `{{ }}` in their own file formats. Gemini CLI command TOML uses `{{args}}`,
+which the engine would try to resolve and fail on. Mark such files `raw: true` in the
+blueprint's `FileRule` so they are copied verbatim.
+
+## Adding a new adapter
+
+1. **Verify from the vendor's own docs.** Not a blog post, not memory — these paths change
+   often, and a wrong path produces a file no tool ever reads. If you cannot verify it, do
+   not ship the adapter.
+
+2. **Create `adapters/<id>/adapter.yaml`.** `<id>` is kebab-case and must equal the `id` field.
+   Cite the source URL in a comment directly above the `outputs:` block, and set
+   `readsAgentsMd` honestly — guessing `false` costs a redundant file, guessing `true` costs
+   the tool all of its instructions.
+
+3. **Decide wrapper vs delta for each output.** Ask: *would this content be wrong in
+   `AGENTS.md`?* If no, it belongs in `AGENTS.md` and the adapter emits a pointer instead.
+
+4. **Add templates under `adapters/<id>/templates/`.** Name each after the file it produces.
+   Use only the variables in the contract above; adding a new one means teaching the engine
+   to supply it. If two tools share a format exactly, put the template in `_shared/` and
+   reference it from both — but each tool still gets its own `adapter.yaml`.
+
+5. **Declare capabilities you actually emit outputs for.** A capability with no corresponding
+   output is a claim the adapter does not honour.
+
+6. **Check the matrix above stays true**, and re-run `groundplan plan` against a sample repo to
+   confirm no token resolves to empty.

package/adapters/_shared/README.md

@@ -0,0 +1,16 @@
+# _shared
+
+Templates used by more than one adapter live here.
+
+## The `.agent/` protocol moved
+
+`ACTIVE.md`, `HANDOFF.md`, `DECISIONS.md`, `DONE.md`, and `DELEGATION.md` are **not** adapter
+output. They are tool-neutral by definition — plain markdown at fixed paths that every tool
+reads and writes — so they are emitted once per repo by the agent-kit blueprint:
+
+```
+templates/agent-kits/core/files/.agent/
+```
+
+Keeping a second copy here would have made two sources of truth for the exact files whose
+purpose is to be a single shared source. Adapters reference these paths; they do not own them.

package/adapters/aider/adapter.yaml

@@ -0,0 +1,30 @@
+id: aider
+name: Aider
+# No auto-discovery of any instruction file. But `read:` can point at AGENTS.md directly,
+# so aider needs zero duplicated content — just a config line.
+readsAgentsMd: false
+# --message processes one instruction and exits; no separate headless flag exists.
+invoke:
+  binary: aider
+  args: ["--message", "{prompt}", "--no-stream"]
+  promptVia: arg
+  autoApprove: ["--yes-always"]  # `--yes` is only an argparse abbreviation, not a documented flag
+  modelFlag: ["--model", "{model}"]
+  detect: ["--version"]
+  install: "python -m pip install aider-install && aider-install"
+  verifiedOn: "2026-07-19"
+  verified: "https://aider.chat/docs/scripting.html"
+capabilities:
+  - instructions
+
+# Source: https://aider.chat/docs/usage/conventions.html
+#   Conventions load three ways: /read <file>, --read <file>, or `read:` in .aider.conf.yml.
+#   `read:` marks the file read-only and prompt-cacheable, which is why it beats --file.
+#   CONVENTIONS.md is only aider's documentation convention, not a path it auto-loads.
+outputs:
+  # Deliberately NO CONVENTIONS.md: that file would be a verbatim copy of AGENTS.md.
+  - path: .aider.conf.yml
+    kind: settings
+    mode: settings
+
+docs: https://aider.chat/docs/usage/conventions.html

package/adapters/aider/templates/.aider.conf.yml.tmpl

@@ -0,0 +1,5 @@
+# Aider has no auto-discovered instruction file, so point `read:` at the canonical one.
+# Read-only means aider will not edit it, and it stays in the prompt cache across turns.
+read:
+  - {{ agentsMdPath }}
+  - .agent/DELEGATION.md

package/adapters/amp/adapter.yaml

@@ -0,0 +1,32 @@
+id: amp
+name: Amp (Sourcegraph)
+# AGENTS.md is Amp's native format; AGENT.md and CLAUDE.md are legacy fallbacks.
+readsAgentsMd: true
+# Amp never asks for approval, so there is no auto-approve flag — the sandbox is the guardrail.
+invoke:
+  binary: amp
+  args: ["-x", "{prompt}", "--stream-json"]
+  promptVia: arg
+  autoApprove: []
+  modelFlag: []
+  detect: ["--version"]
+  install: "curl -fsSL https://ampcode.com/install.sh | bash"
+  verifiedOn: "2026-07-19"
+  verified: "https://ampcode.com/manual"
+capabilities:
+  - instructions
+  - nested-instructions
+  - skills
+
+# Source: https://ampcode.com/manual
+#   Instructions: AGENTS.md walked from cwd up to $HOME, plus subtrees the agent reads into.
+#   Skills: .agents/skills/ (project) — note the leading dot and plural, not .amp/.
+#   MCP: amp.mcpServers setting, or an mcp.json bundled inside a skill.
+outputs:
+  - path: .agents/skills/__skillId__/SKILL.md
+    template: SKILL.md
+    kind: skill
+    mode: delta
+    when: hasSkills
+
+docs: https://ampcode.com/manual

package/adapters/amp/templates/SKILL.md.tmpl

@@ -0,0 +1,8 @@
+---
+name: {{ itemId }}
+description: {{ itemDescription }}
+---
+
+# {{ itemName }}
+
+{{ itemBody }}

package/adapters/antigravity-cli/adapter.yaml

@@ -0,0 +1,36 @@
+id: antigravity-cli
+name: Antigravity CLI
+# Google's successor to Gemini CLI. Reads AGENTS.md no more natively than its predecessor did,
+# so it takes the same @import wrapper.
+readsAgentsMd: false
+
+# Verified against Google's own codelab and the antigravity-cli README.
+#
+# Deliberately NO --output-format here. The repo README lists one; the codelab does not
+# document it, and an unverified output flag is exactly the "approximately-right flag" that
+# produces a silent no-op (ADR 0007 §4). Add it once `agy --help` confirms the value.
+invoke:
+  binary: agy
+  args: ["-p", "{prompt}"]
+  promptVia: arg
+  autoApprove: ["--dangerously-skip-permissions"]
+  modelFlag: ["--model", "{model}"]
+  detect: ["--version"]
+  # Not on npm; @google/antigravity-cli does not exist.
+  install: "curl -fsSL https://antigravity.google/cli/install.sh | bash"
+  verifiedOn: "2026-07-19"
+  verified: "https://codelabs.developers.google.com/antigravity-cli-hands-on"
+
+capabilities:
+  - instructions
+  - nested-instructions
+
+# Source: https://antigravity.google/docs/gcli-migration
+#   Skills, hooks, subagents and extensions carry over from Gemini CLI as plugins, so the
+#   registry items that targeted gemini-cli apply here unchanged.
+outputs:
+  - path: ANTIGRAVITY.md
+    kind: instructions
+    mode: wrapper
+
+docs: https://antigravity.google/docs/gcli-migration

package/adapters/antigravity-cli/templates/ANTIGRAVITY.md.tmpl

@@ -0,0 +1,11 @@
+@{{ agentsMdPath }}
+
+# Antigravity CLI
+
+The import above is the whole project instruction set. Nothing below restates it.
+
+- Google's successor to Gemini CLI. Skills, hooks, subagents and extensions carry over as
+  plugins, so anything that worked under `gemini` should work under `agy`.
+- Cross-tool handoff state lives in `.agent/`; routing is in `.agent/DELEGATION.md`.
+- Delegating here uses `agy -p`. `--output-format` is deliberately not wired up until the
+  flag is confirmed against `agy --help` — see `adapters/antigravity-cli/adapter.yaml`.

package/adapters/claude-code/adapter.yaml

@@ -0,0 +1,65 @@
+id: claude-code
+name: Claude Code
+# Anthropic docs state plainly: "Claude Code reads CLAUDE.md, not AGENTS.md."
+readsAgentsMd: false
+# Headless delegation target. Flags verified from Anthropic's CLI reference.
+invoke:
+  binary: claude
+  args: ["-p", "{prompt}", "--output-format", "json"]
+  promptVia: arg
+  # Skips every permission prompt. Only ever passed when the caller opts in explicitly.
+  autoApprove: ["--dangerously-skip-permissions"]
+  modelFlag: ["--model", "{model}"]
+  detect: ["--version"]
+  install: "npm i -g @anthropic-ai/claude-code"
+  verifiedOn: "2026-07-19"
+  verified: "https://code.claude.com/docs/en/cli-reference"
+capabilities:
+  - instructions
+  - nested-instructions
+  - glob-scoped-rules
+  - slash-commands
+  - subagents
+  - hooks
+  - skills
+
+# Sources:
+#   https://code.claude.com/docs/en/memory        (CLAUDE.md, @import, .claude/rules/ paths:)
+#   https://code.claude.com/docs/en/skills        (.claude/skills/<id>/SKILL.md, .claude/commands/)
+#   https://code.claude.com/docs/en/sub-agents    (.claude/agents/*.md)
+#   https://code.claude.com/docs/en/settings      (.claude/settings.json)
+outputs:
+  - path: CLAUDE.md
+    kind: instructions
+    mode: wrapper
+
+  - path: .claude/settings.json
+    kind: settings
+    mode: settings
+
+  # `paths:` frontmatter loads the rule only when Claude touches a matching file.
+  - path: .claude/rules/__ruleId__.md
+    template: rule.md
+    kind: rule
+    mode: delta
+    when: hasGlobRules
+
+  - path: .claude/skills/__skillId__/SKILL.md
+    kind: skill
+    mode: delta
+    when: hasSkills
+
+  # Commands merged into skills upstream; the directory still works and stays cheaper for one-shots.
+  - path: .claude/commands/__commandId__.md
+    template: command.md
+    kind: command
+    mode: delta
+    when: hasCommands
+
+  - path: .claude/agents/__agentId__.md
+    template: agent.md
+    kind: agent
+    mode: delta
+    when: hasAgents
+
+docs: https://code.claude.com/docs/en/memory

package/adapters/claude-code/templates/CLAUDE.md.tmpl

@@ -0,0 +1,24 @@
+@{{ agentsMdPath }}
+
+# Claude Code
+
+Everything above is canonical and shared with every other agent. This section is the only
+Claude-specific part — keep it that way.
+
+## Extensions available here
+
+- Path-scoped rules live in `.claude/rules/`; they load only when you touch matching files.
+- Skills live in `.claude/skills/`; invoke with `/<skill-name>` or let Claude pick one.
+- Subagents live in `.claude/agents/`; delegate research and review so the main context stays clean.
+
+## Working agreement
+
+- Use plan mode for changes that span more than one package.
+- Protected paths in the canonical instructions are enforced by `permissions.deny` in
+  `.claude/settings.json`, not by good intentions. Do not edit that list to get unblocked.
+- Delegate rather than dump: send broad searches to a subagent and keep the summary.
+
+## Cross-tool handoff
+
+Other tools continue this work. Before ending a session with anything unfinished, update
+`.agent/ACTIVE.md` and write `.agent/HANDOFF.md`. Routing lives in `.agent/DELEGATION.md`.

package/adapters/claude-code/templates/SKILL.md.tmpl

@@ -0,0 +1,8 @@
+---
+name: {{ itemId }}
+description: {{ itemDescription }}
+---
+
+# {{ itemName }}
+
+{{ itemBody }}

package/adapters/claude-code/templates/agent.md.tmpl

@@ -0,0 +1,9 @@
+---
+name: {{ itemId }}
+description: {{ itemDescription }}
+tools: {{ itemTools }}
+---
+
+{{ itemBody }}
+
+Canonical project rules apply — read `{{ agentsMdPath }}` before acting.

package/adapters/claude-code/templates/command.md.tmpl

@@ -0,0 +1,6 @@
+---
+description: {{ itemDescription }}
+argument-hint: {{ itemArgumentHint }}
+---
+
+{{ itemBody }}

package/adapters/claude-code/templates/rule.md.tmpl

@@ -0,0 +1,8 @@
+---
+paths:
+{{ itemGlobsYamlList }}
+---
+
+# {{ itemName }}
+
+{{ itemBody }}

package/adapters/claude-code/templates/settings.json.tmpl

@@ -0,0 +1,11 @@
+{
+  "$schema": "https://json.schemastore.org/claude-code-settings.json",
+  "permissions": {
+    "allow": [
+      "Bash({{ buildCommand }})",
+      "Bash({{ testCommand }})",
+      "Bash({{ lintCommand }})"
+    ],
+    "deny": {{ protectedPathsDenyJson }}
+  }
+}

package/adapters/cline/adapter.yaml

@@ -0,0 +1,31 @@
+id: cline
+name: Cline
+# Cline auto-detects AGENTS.md alongside its own rules — no wrapper needed.
+readsAgentsMd: true
+# Trap: `-p` here means --plan, NOT --print. Headless activates on --json or piped stdin.
+invoke:
+  binary: cline
+  args: ["--json", "{prompt}"]
+  promptVia: arg
+  autoApprove: ["-y"]
+  modelFlag: ["-m", "{model}"]
+  detect: ["--version"]
+  install: "npm i -g cline"
+  verifiedOn: "2026-07-19"
+  verified: "https://github.com/cline/cline/blob/main/apps/cli/README.md"
+capabilities:
+  - instructions
+  - glob-scoped-rules
+
+# Source: https://docs.cline.bot/features/cline-rules
+#   Workspace rules: .clinerules/ directory (preferred) at project root.
+#   Also auto-detects .cursorrules, .windsurfrules, and AGENTS.md for cross-tool compatibility.
+#   Rules support YAML frontmatter with glob patterns for conditional activation.
+outputs:
+  - path: .clinerules/__ruleId__.md
+    template: rule.md
+    kind: rule
+    mode: delta
+    when: hasGlobRules
+
+docs: https://docs.cline.bot/features/cline-rules

package/adapters/cline/templates/rule.md.tmpl

@@ -0,0 +1,8 @@
+---
+description: {{ itemDescription }}
+globs: {{ itemGlobsCsv }}
+---
+
+# {{ itemName }}
+
+{{ itemBody }}

package/adapters/codex/adapter.yaml

@@ -0,0 +1,29 @@
+id: codex
+name: OpenAI Codex
+# Codex is the reference implementation of the AGENTS.md standard. No wrapper, ever.
+readsAgentsMd: true
+# `exec` is the scriptable mode; --json emits NDJSON events. --ephemeral keeps CI clean.
+invoke:
+  binary: codex
+  args: ["exec", "--json", "--ephemeral", "--sandbox", "workspace-write", "{prompt}"]
+  promptVia: arg
+  autoApprove: ["--dangerously-bypass-approvals-and-sandbox"]
+  modelFlag: ["--model", "{model}"]
+  detect: ["--version"]
+  install: "npm i -g @openai/codex"
+  verifiedOn: "2026-07-19"
+  verified: "https://learn.chatgpt.com/docs/non-interactive-mode"
+capabilities:
+  - instructions
+  - nested-instructions
+
+# Sources:
+#   https://agents.md/                                     (Codex listed as native consumer)
+#   https://learn.chatgpt.com/docs/config-file/config-basic (project .codex/config.toml, precedence)
+#
+# No outputs. Codex reads AGENTS.md natively and needs no wrapper, and its only other file is
+# .codex/config.toml, which exists to declare MCP servers — see docs/EXECUTION.md for why
+# Groundplan does not claim MCP support today.
+outputs: []
+
+docs: https://learn.chatgpt.com/docs/config-file/config-basic

package/adapters/continue/adapter.yaml

@@ -0,0 +1,36 @@
+id: continue
+name: Continue
+# Continue's rules docs describe no AGENTS.md discovery, so a pointer rule is required.
+readsAgentsMd: false
+# Trap: headless silently EXCLUDES write tools unless --auto or --allow is passed, so a run
+# can report success having changed nothing. autoApprove is effectively required for edits.
+invoke:
+  binary: cn
+  args: ["-p", "{prompt}", "--format", "json"]
+  promptVia: arg
+  autoApprove: ["--auto"]
+  modelFlag: ["--model", "{model}"]
+  detect: ["--version"]
+  install: "npm i -g @continuedev/cli"
+  verifiedOn: "2026-07-19"
+  verified: "https://docs.continue.dev/cli/headless-mode"
+capabilities:
+  - instructions
+  - glob-scoped-rules
+
+# Source: https://docs.continue.dev/customize/deep-dives/rules
+#   .continue/rules/*.md, loaded in lexicographic order (hence the 00- prefix on the pointer).
+#   Frontmatter: name (required), globs, regex, description, alwaysApply.
+#   Rules feed Agent/Chat/Edit system messages — never Autocomplete.
+outputs:
+  - path: .continue/rules/00-agents-md.md
+    kind: instructions
+    mode: wrapper
+
+  - path: .continue/rules/__ruleId__.md
+    template: rule.md
+    kind: rule
+    mode: delta
+    when: hasGlobRules
+
+docs: https://docs.continue.dev/customize/deep-dives/rules

package/adapters/continue/templates/00-agents-md.md.tmpl

@@ -0,0 +1,9 @@
+---
+name: Project instructions
+alwaysApply: true
+---
+
+Read `{{ agentsMdPath }}` at the repository root before making any change. It is the single
+source of truth for this project's commands, conventions, boundaries, and protected paths.
+
+Cross-tool handoff state lives in `.agent/`. Routing is in `.agent/DELEGATION.md`.

package/adapters/continue/templates/rule.md.tmpl

@@ -0,0 +1,8 @@
+---
+name: {{ itemName }}
+description: {{ itemDescription }}
+globs: {{ itemGlobsCsv }}
+alwaysApply: false
+---
+
+{{ itemBody }}

package/adapters/cursor/adapter.yaml

@@ -0,0 +1,34 @@
+id: cursor
+name: Cursor
+# Cursor reads root and nested AGENTS.md directly; .mdc rules are for glob-scoped deltas only.
+readsAgentsMd: true
+# The installer symlinks both `agent` and `cursor-agent`; the latter is safer across versions.
+# Print mode already has write and shell access — --force only removes command approval.
+invoke:
+  binary: cursor-agent
+  args: ["-p", "{prompt}", "--output-format", "stream-json"]
+  promptVia: arg
+  autoApprove: ["--force"]
+  modelFlag: ["--model", "{model}"]
+  detect: ["--version"]
+  install: "curl https://cursor.com/install -fsS | bash"
+  verifiedOn: "2026-07-19"
+  verified: "https://cursor.com/docs/cli/reference/parameters"
+capabilities:
+  - instructions
+  - nested-instructions
+  - glob-scoped-rules
+
+# Source: https://cursor.com/docs/context/rules
+#   .cursor/rules/*.mdc with frontmatter: description, globs, alwaysApply.
+#   Plain .md in .cursor/rules/ is IGNORED — the .mdc extension is mandatory.
+#   Rules apply to Agent/Chat only, never to Tab or Cmd+K inline edit.
+outputs:
+  # No instruction wrapper: a rule that restated AGENTS.md would double-load it.
+  - path: .cursor/rules/__ruleId__.mdc
+    template: rule.mdc
+    kind: rule
+    mode: delta
+    when: hasGlobRules
+
+docs: https://cursor.com/docs/context/rules