# Aegis Host Compatibility Matrix Snapshot

Status: `Reviewed`

## 1. Document Scope

This document records the current host compatibility snapshot of the `Aegis Method Pack`.

It answers:

- Which hosts have fresh evidence
- Which hosts only have structural goals without current fresh verification
- What is the reading order for the current compatibility verdict

It does not answer:

- Permanent support commitments for all hosts
- Complete platform-level runtime compatibility

---

## 2. Snapshot Date

The current snapshot is based on fresh evidence and current docs landed as of
`2026-08-27`.

---

## 3. Current Verdict

### 3.1 Hosts With Fresh Evidence

| Host | Current Verdict | Evidence Owner |
| --- | --- | --- |
| `Codex` | Representative smoke mainline available; activation/TDD mode can be projected through the managed global `AGENTS.md` block while native matcher behavior remains host-owned; managed-worktree binding is structurally guarded, while live Desktop task-binding closeout and naive routing smoke under Git Bash still require observation | `docs/testing.md`, `tests/skill-triggering/*`, `tests/explicit-skill-requests/*`, `tests/e2e/aegis-doctor-check.sh`, `tests/e2e/workflow-quality-check.sh`, `docs/README.codex.md`, `docs/current/AEGIS_KNOWN_LIMITATIONS.md` |
| `OpenCode` | Base suite and integration closeout passed | `docs/testing.md`, `tests/opencode/*`, `docs/README.opencode.md`; current mainline prefers the configured `method_pack_root` as the canonical Aegis body and treats the OpenCode-visible skills tree as a generated host view |

The [optional Codex reconciliation profile](../README.codex-reconciliation.md)
has a separate, opt-in installation and evidence scope. It does not change the
default Codex skill-only verdict above; its CLI version, platform, trust,
warning visibility and diagnostic limits are owned by that guide and
`tests/codex-reconciliation/README.md`.

### 3.2 Hosts Without a Current Fresh Release Verdict

| Host | Current Status | Why Not Yet |
| --- | --- | --- |
| `Claude Code` | Has install guide and plugin skeleton; no current release-level fresh smoke verdict | `docs/README.claude-code.md` established, but real host regression is still deferred |
| `CC GUI (JetBrains IDEA)` | Structural IDE plugin layer target for Claude Code / OpenAI-GPT provider paths; no current release-level fresh smoke verdict | `docs/README.cc-gui.md` established; CC GUI's OpenAI/Codex provider scanner expects direct `.agents/skills/<skill-name>/SKILL.md` skill directories regardless of the selected GPT model profile, so Aegis must expose individual skills rather than only an umbrella `~/.agents/skills/aegis` directory; live IDE plugin smoke and host adapter event rendering remain deferred |
| `CodeBuddy` | Has `.codebuddy-plugin/` skeleton and native `SKILL.md` manual install instructions; no current release-level fresh smoke verdict | `docs/README.codebuddy.md` established; evidence from CodeBuddy skills/plugin docs and this repo's `.codebuddy-plugin/`; real host regression still deferred |
| `DeepSeek-TUI` | Native `SKILL.md` discovery supports manual installation; no current release-level fresh smoke verdict | `docs/README.deepseek-tui.md` established; evidence from DeepSeek-TUI README/source discovery contract; real host regression still deferred |
| `DeepSeek Harness` | Native DSH bundle installation, filesystem skill discovery, and lifecycle router bootstrap support structural install; no current release-level fresh smoke verdict | `docs/README.deepseek-harness.md` established; the default path is the thin Aegis `dsh.bundle`, which registers the package-owned `skills/` tree and defers a compact router bootstrap, armed at native `agent/created` or legacy `agent/session-start` boundaries and injected after the session's first durable promotion signal; updater-managed direct-child exposure under `$DSH_HOME/skills` remains an explicit compatibility mode without that bootstrap; the official host remains a developer preview with compatibility-breaking changes expected, and it is separate from DeepSeek-TUI |
| `Trae` | Native `SKILL.md` discovery supports manual installation; no current release-level fresh smoke verdict | `docs/README.trae.md` established; evidence from Trae skills docs; real host regression still deferred |
| `GitHub Copilot` | Supports repository skills, instructions, and hooks; no current release-level fresh smoke verdict | `docs/README.copilot.md` established; evidence from GitHub Copilot agent skills, repository instructions, and hooks docs; Aegis exposes repository skills as `.github/skills/aegis-<skill-name>/SKILL.md`, verified as a prefixed direct-child compatibility view, but real host regression is still deferred |
| `Qoder` | Native `SKILL.md` discovery and rules surfaces support structural install; no current release-level fresh smoke verdict | `docs/README.qoder.md` established; evidence from Qoder skills and rules docs, but real host regression is still deferred |
| `Kimi Code CLI` | Native plugin bootstrap and Agent Skills discovery are structurally supported; no current release-level fresh smoke verdict | `docs/README.kimi-code.md` established; the default automatic path uses root `kimi.plugin.json` with `sessionStart.skill = using-aegis`; updater-managed direct-child exposure under `$KIMI_CODE_HOME/skills/` (`~/.kimi-code/skills/` by default) remains the explicit compatibility mode, while `~/.agents/skills/` is only a shared fallback; these exposure routes must not be enabled together |
| `Grok Build` | Native Agent Skills, `AGENTS.md`, extra `[skills] paths`, and Claude-compatible plugin discovery support structural install; no current release-level fresh smoke verdict | `docs/README.grok-build.md` established; updater-managed direct-child exposure defaults to `$GROK_HOME/skills/` (`~/.grok/skills/` by default), while config-path and Claude-plugin discovery remain alternatives that must not be enabled as duplicate Aegis owners; local `grok inspect --json` evidence confirms enumeration but not clean-install or live-trigger closeout |
| `Cursor` | Has `.cursor/INSTALL.md` install guide; no current release-level fresh smoke verdict | Structural goal established; not yet entered the current host regression slice |
| `Windsurf` | Has `.windsurf/INSTALL.md` install guide; no current release-level fresh smoke verdict | Structural goal established; not yet entered the current host regression slice |
| `Antigravity CLI` | Active closeout target; no current release-level fresh smoke verdict | Google positions Antigravity CLI as the successor terminal surface with the `agy` executable, plugin/skills management, migration import from Gemini CLI, and documented plugin discovery for skills and agents from installed plugin directories in the public `1.0.1` changelog; the public `1.0.8` changelog also adds dynamic custom-skill discovery improvements. Aegis now has a dedicated `tests/antigravity/run-tests.sh` lane, but the method-pack install / discovery contract still needs fresh local verification |
| `Antigravity IDE` | Structural target retained; no current release-level fresh smoke verdict | Google positions Antigravity IDE as an editor surface with global or workspace-specific Skills, MCPs, and JSON Hooks; this shape is not the active closeout slice and still needs fresh verification |
| `Antigravity App` | Structural target retained; no current release-level fresh smoke verdict | Google Antigravity 2.x app / project platform can carry Aegis method-pack projections, but this shape is not the active closeout slice and no live host closeout has been run |
| `Pi CLI` | Agent Skills and Pi package surface supports method-pack exposure; no current release-level fresh smoke verdict | Pi documents skill discovery from `~/.pi/agent/skills/`, `~/.agents/skills/`, `.pi/skills/`, package `skills/` directories or `pi.skills` entries, and `pi install git:github.com/GanyuanRan/Aegis` package installs; Aegis now ships a Pi extension (`pi.extensions` manifest entry) that injects the compact using-aegis hot path and a routing guard through Pi's `context`/`message_end`/`tool_call`/`tool_result` events; Aegis still needs current release live smoke before claiming host closeout |
| `OMP (Oh My Pi)` | Agent Skills `agents` provider (`~/.agents/skills/`), `alwaysApply` frontmatter, and extension loading support structural exposure; no current release-level fresh smoke verdict | `docs/README.omp.md` established; OMP is a fork of Pi that discovers `~/.agents/skills/`, honors `alwaysApply: true` (full content injected into the system prompt), auto-discovers `~/.omp/agent/extensions/`, and accepts `omp.extensions` / `pi.extensions` package manifests; Aegis marks `using-aegis` with `alwaysApply` and ships an `omp.extensions` bundle, but live OMP smoke is still deferred |
| `OpenClaw` | Native `SKILL.md` skill installer supports local skill directories; no current release-level fresh smoke verdict | `docs/README.openclaw.md` established; evidence from OpenClaw `openclaw skills install` docs that Git/local installs expect `SKILL.md` at the source root; Aegis is multi-skill, so individual skill-directory install is the structural path pending live smoke |
| `Hermes Agent` | Skills Hub / `~/.hermes/skills/` skill host surface supports method-pack exposure; no current release-level fresh smoke verdict | `docs/README.hermes-agent.md` established; evidence from Hermes Skills Hub, documented local skills path, GitHub path installs, and built-in coding delegation skills; live host regression still deferred |
| `ZCode` | Depth-1 `~/.agents/skills/<skill-name>/SKILL.md` skill discovery (like CC GUI and Windsurf) plus Claude-Code-compatible plugin marketplace reading `.claude-plugin/marketplace.json`; direct-child exposure is now updater-maintained at register/update time; reads `AGENTS.md`; no current release-level fresh smoke verdict | `docs/README.zcode.md` established; evidence from ZCode plugin and skill official docs; Aegis's existing `.claude-plugin/` skeleton works with zero code changes, but live host regression still deferred |

### 3.2a Retired Host Surfaces

| Host | Current Status | Retirement Boundary |
| --- | --- | --- |
| `Gemini CLI` | Retired; unsupported by Aegis | Google announced the consumer transition on `2026-05-19`; service stopped on `2026-06-18` for free usage, Google AI Pro / Ultra, and Gemini Code Assist for individuals, while Enterprise Standard / Enterprise, Google Cloud, and paid API-key paths remained outside that boundary. Aegis no longer ships `GEMINI.md`, `gemini-extension.json`, or a Gemini-specific tool mapping. This is an explicit Aegis product-support retirement rather than a claim that Gemini CLI no longer exists. Historical migration guidance remains in `docs/README.antigravity.md`. |

### 3.3 Hosts Requiring No Independent Adapter

| Host | Current Status | Rationale |
| --- | --- | --- |
| `Warp` | No independent adapter needed | As a terminal host, Warp runs third-party CLI agents (Claude Code / Codex / OpenCode) and does not provide its own skills system |

### 3.4 Trigger-Family Vocabulary

The current docs may refer to these trigger families as a compact compatibility
aid:

- `hook-bootstrap`
- `native-direct-skill`
- `provider-hybrid`

This vocabulary is advisory and diagnostic only. It does **not** replace
host-specific docs as the canonical owner for install roots, reload behavior,
or supported discovery shapes.

---

## 4. What This Snapshot Means

The current snapshot only states:

1. `Codex` and `OpenCode` are the two mainlines with the most fresh evidence
   - `OpenCode` now prefers the configured `method_pack_root` as its canonical Aegis source when available, then generates the host-visible skills view from that source instead of treating the host cache as a second editable owner
2. `Kimi Code CLI` defaults to the native Aegis plugin declared by
   `kimi.plugin.json`, whose `sessionStart.skill = using-aegis` provides the
   automatic router entry; updater-maintained direct-child discovery under
   `$KIMI_CODE_HOME/skills/` (`~/.kimi-code/skills/` by default) remains an
   explicit compatibility mode, and `~/.agents/skills/` remains a shared
   fallback; exactly one route may be active, and fresh live routing evidence
   is still pending
3. `CC GUI (JetBrains IDEA)` can expose Aegis to its OpenAI/GPT provider path through direct `~/.agents/skills/<skill-name>/SKILL.md` directories, but the selected GPT model profile does not by itself change the skill discovery shape, and live IDE plugin behavior, reload behavior, and `Tool: exec_command` rendering remain fresh-smoke pending
4. `CodeBuddy` can install Aegis via `.codebuddy-plugin/` or native `SKILL.md` discovery, but local CLI live smoke has not yet formed valid evidence
5. `DeepSeek-TUI` can manually install Aegis skills via native `SKILL.md` discovery, but `/skill install github:GanyuanRan/Aegis` is not the current canonical path
6. `Trae` can manually install Aegis skills via native `SKILL.md` discovery; `.agents/skills/` is an optional Trae capability, not Aegis's canonical Trae path
7. `GitHub Copilot` can expose Aegis through `.github/skills/aegis-<skill-name>/SKILL.md`, `.github/copilot-instructions.md`, `.github/hooks/*.json`, and `AGENTS.md`; the prefixed repository skill view is a generated compatibility exposure over the canonical `skills/<skill-name>/SKILL.md` tree, and support remains structural until a fresh Copilot install and agent smoke proves it for the current release
8. `Qoder` can expose Aegis through `~/.qoder/skills/`, `.qoder/skills/`, `.qoder/rules/`, and `AGENTS.md`, but support remains structural until a fresh Qoder install smoke proves it for the current release
9. `Cursor` and `Windsurf` have structured install guides but have not yet entered release-level fresh smoke
10. `Antigravity CLI` is now the active Google-host closeout target, but still lacks release-level fresh smoke evidence
11. `Antigravity IDE` and `Antigravity App` remain structural target surfaces rather than the active closeout slice
12. `Gemini CLI` support is retired; this does not upgrade Antigravity CLI, IDE,
    or App beyond their separately evidenced current status
13. `Pi CLI` can expose Aegis through `pi install git:github.com/GanyuanRan/Aegis`, Pi package `skills/` / `pi.skills` discovery, `~/.pi/agent/skills/`, `~/.agents/skills/`, or `.pi/skills/`, but support remains structural until a fresh Pi install smoke proves it for the current release
14. `OpenClaw` can expose Aegis through individual local skill-directory installs, but `git:owner/repo` should not be written as the canonical whole-repo Aegis installer because OpenClaw expects `SKILL.md` at the source root
15. `Hermes Agent` can expose Aegis through `~/.hermes/skills/` or documented GitHub path installs, but support remains structural until a fresh Hermes install smoke proves it for the current release
16. `ZCode` can expose Aegis through updater-maintained depth-1 `~/.agents/skills/<skill-name>/SKILL.md` direct-child skill discovery (like CC GUI and Windsurf) and through its native `.claude-plugin/marketplace.json` reading (Claude Code plugin format), so the existing Claude Code plugin skeleton works with zero code changes; support remains structural until a fresh ZCode install smoke proves it for the current release
17. `Warp`, as a terminal host, does not itself need an independent adapter
18. The current method-pack still retains the cross-host installation goal
19. "Support all plugin hosts" remains a product goal, not equivalent to "all hosts have current fresh closeout"
20. Some hosts may require a compatibility exposure shape that differs from a
    mainline host's canonical install shape; when that happens, the method-pack
    `skills/` tree remains the canonical source and the extra exposure should be
    treated as a generated compatibility view rather than a second editable
    owner
21. The shared local `~/.config/aegis/config.toml` may declare one canonical
    `method_pack_root`; when it does, new host registrations should prefer that
    root and treat host-specific discovery roots, plugin caches, or copied skill
    trees as generated or host-managed views into the same Aegis body
22. `Grok Build` can expose Aegis through updater-maintained direct-child
    skills under `$GROK_HOME/skills/` (`~/.grok/skills/` by default), an
    explicit `[skills] paths` entry, or Claude-compatible plugin discovery;
    these are alternative exposure routes, not three simultaneous owners, and
    support remains structural until clean-install and live-trigger smoke is
    fresh for the current release
23. `DeepSeek Harness` defaults to the thin Aegis `dsh.bundle`, installed and
    updated by the host profile plugin manager; the bundle mounts the
    package-owned `skills/` tree through Harness's native filesystem provider
    and, in `auto` mode, defers a compact `using-aegis` routing bootstrap that
    is armed at native `startup`, `resume`, `clear`, and `compact` session
    starts and injected once after the session's first durable promotion
    signal (`tool/call` or `assistant/message`), so the first model request of
    every gated epoch stays free of injected context. This is advisory
    model-facing context, not a hard tool guard or runtime authority.
    The official installation source is this Git repository, whose package
    identity is `aegis`. The Cordis row ID `aegis-method-pack` does not declare
    an npm distribution. The public npm package of that name is published by
    an external account; Aegis does not maintain an official npm registry
    channel. Marketplace repository metadata must not override the selected
    Git source or its pinned revision. Keep exactly one Aegis exposure per
    profile and verify the actual installed source and version.
    Profile management is host-owned: the CLI manages Web and Headless,
    while DSH `0.2.0-rc.2` reserves `desktop` for the Electron application.
    Desktop installation, update, removal, and inspection use the application's
    plugin manager; CLI commands must not be projected onto that profile.
    A reporter-confirmed Desktop install and visible plugin do not establish
    the installed source/version or native skill loading.
    Package admission retains the three optional DSH host peers with a
    `>=0.1.0-rc.6` floor and no speculative upper version bound. DSH evaluates
    these ranges with prereleases included. Admission of a newer version is
    not evidence of API compatibility or live routing; verified host versions
    and remaining gaps belong in `AEGIS_KNOWN_LIMITATIONS.md`.
    Updater-managed direct-child skills under
    `$DSH_HOME/skills` (`~/.dsh/skills` by default), project `.dsh/skills`,
    shared `.agents/skills`, and custom skill directories remain explicit
    compatibility alternatives without bundle bootstrap; exactly one Aegis
    exposure should be active, and structural catalog/bootstrap discovery is
    not a current live-routing verdict
24. ChatGPT desktop app managed-worktree guidance now requires the task
    workspace, default command `cwd`, Git worktree root, and intended `HEAD` to
    agree before writes. External `git worktree add` plus command-level
    `workdir` is not accepted as task binding. This is structural method-pack
    coverage; the UI/task-binding matrix remains environment-bound evidence and
    is tracked in `AEGIS_KNOWN_LIMITATIONS.md`.

---

## 5. Evidence Sources

When reading the current host verdict, follow this order:

1. `docs/testing.md`
2. `docs/README.claude-code.md`
3. `docs/README.codex.md`
4. `docs/README.cc-gui.md`
5. `docs/README.opencode.md`
6. `docs/README.codebuddy.md`
7. `docs/README.deepseek-tui.md`
8. `docs/README.trae.md`
9. `docs/README.copilot.md`
10. `docs/README.qoder.md`
11. `docs/README.kimi-code.md`
12. `docs/README.antigravity.md`
13. `tests/antigravity/*`
14. `docs/README.pi.md`
15. `docs/README.openclaw.md`
16. `docs/README.hermes-agent.md`
17. `.windsurf/INSTALL.md`
18. `.cursor/INSTALL.md`
19. `docs/README.zcode.md`
20. `docs/README.grok-build.md`
21. `docs/README.deepseek-harness.md`

---

## 6. Compatibility Boundary

The current snapshot only covers:

- Method-pack installation and distribution
- skill discovery / representative triggering
- Project workspace support when the installed method-pack root remains
  available and can be verified by
  `cd <aegis-method-pack-root> && python scripts/aegis-doctor.py --write-config --json`
- Host-scoped explicit update registration through
  `python scripts/aegis-update.py status --json` and
  `python scripts/aegis-update.py update --host <host> --json`
- Plugin loading / priority / distribution sync

The current snapshot does not cover:

- Runtime core integration
- Host adapter event normalization
- Complete live production workflow orchestration

Skill discovery and project workspace support are related but distinct. A
skills-only copy can make Aegis workflows visible to a host while failing to
prove complete project workspace support. A complete install should preserve
the method-pack root or equivalent configured support path so workspace
bootstrap, indexing, work records, and structural checks remain available.
Complete-install verification should write local helper paths and read back
`"ok": true`, `"workspaceSupport": "available"`, and `"configStatus":
"configured"` from `aegis-doctor.py`.

Do not run `aegis-doctor.py` from the target project directory. The script
belongs to the installed Aegis method-pack root; target projects are passed
separately to workspace helper commands with `--root <target-project-root>`.

---

## 7. Architecture Review

Three misjudgments must be avoided when reading this snapshot:

1. Mistaking the current pass of `Codex + OpenCode` for "all hosts have been formally closed out"
2. Mistaking current smoke / integration verdicts for full-platform readiness
3. Mistaking host compatibility work for elevating method-pack authority
