golem15/msd-core
3
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity
Files
5d804dd2877a506735e4959341e0efbf33658809
msd-core/docs/reference/host-integration-capability-matrix.md
Tom Boucher 585a8b7f1b fix(#3747): correct antigravity matrix evidence and pin the CLI-only skills install path (#4274) 
* test(#3747): fail-first regression — matrix must not cite configHome skills path for antigravity

* fix(#3747): correct disproven antigravity stateIO evidence; pin CLI-only probe branch install path

* fix(#3747): scope doc evidence claim to skills discovery per adversarial review

* chore(#3747): add changeset

* chore(#3747): backfill PR number in changeset

---------

Co-authored-by: sim <sim@local>
2026-09-04 11:48:38 -04:00

139 KiB
Raw Blame History

Host Integration Capability Matrix

This document is the maintainer-facing source of truth for the hostIntegration block in every capabilities/<cli>/capability.json runtime descriptor. Every per-CLI axis value is either:

  • documented — backed by a cited authoritative source and evidence quote, or
  • undocumented — the explicit fail-closed sentinel used when the CLI's public documentation does not state a value for that axis. undocumented validates in the registry but never propagates into effective axes: negotiation degrades closed to the safe default.

Values are generated from per-CLI documentation research (Context7 + official docs). They are consumed verbatim by gen:capability-registry and validated by capability-validator.cjs.

Axes legend

Axis Meaning
embeddingMode Whether the CLI exposes an in-process programmatic API (imperative) or integrates purely through configuration files (declarative).
commandSurface How slash commands are registered: slash-file (markdown), slash-toml (TOML), slash-programmatic (code API), palette, prose-only.
modelMode Whether extensions can programmatically request or supply a model (active) or select only by config (passive).
hookBus Who owns the hook lifecycle: host (the CLI fires hooks), engine (VS Code/Electron extension host), none.
stateIO Filesystem access model: filesystem (full local FS), sandboxed-storage, session-log-append.
transport Integration transport: mcp (Model Context Protocol), native-extension.
runtime Plugin/extension execution runtime: node, bun, python, go, rust, electron, sandboxed-web, other.
effortSurface How reasoning effort reaches this host: argv (deliverable as an argument on the host's own invocation), none (the host exposes no reasoning-effort mechanism), or undocumented. Added by #2481; there is deliberately no config-file member — the only host that ever had one (Gemini CLI's thinkingConfig) was removed as a sunset runtime, and naming a member with no host would be a guess.

dispatch sub-axes

Sub-axis Meaning
namedDispatch Whether agents can be invoked by name (true/false/undocumented).
nested Whether subagents can themselves spawn subagents (true/false/undocumented).
maxDepth Maximum nesting depth (integer; -1 = unbounded; undocumented).
background Whether subagents can run asynchronously in the background (true/false/undocumented).
subagentToolkit Tool surface available to subagents: full, read-only, built-in-only, or undocumented. built-in-only means the host ships a fixed set of built-in subagent types whose tool surfaces differ from one another, so no single full/read-only value describes them; it degrades closed to read-only in the negotiated axes.
backgroundDispatch Whether a BACKGROUND-dispatched sub-agent can itself spawn further named sub-agents — the #853 discriminator (true/false/undocumented).

Interface points

Point Meaning
command Slash-command routing and invocation capability.
dispatch Subagent/multi-agent dispatch capability.
model Programmatic model selection capability.
hooks Lifecycle hook registration capability.
state Filesystem/state I/O capability.
artifact Artifact delivery (skills, commands) surface capability.

Trigger precedence (#2871 Phase 2 — adjacent to, not part of, hostIntegration)

runtime.triggerPrecedence (an ordered list of trigger-bearing kind names, highest priority first) is declared as a sibling of hostIntegration in capability.json's runtime body, not inside it — it is not researched per-CLI documentation the way the axes above are, so it carries no per-host Source/Evidence row. Only commands and skills are members of the vocabulary; agents/kimi-agents are excluded because they are not trigger-bearing (a /gsd-<name> a user types) — an agent is invoked through named/subagent_type dispatch, the separate dispatch interface point above, never through the command interface point. Every shipped runtime descriptor declares the same value, ["skills", "commands"] (skills wins a same-scope collision), matching capability-validator.cjs's DEFAULT_TRIGGER_PRECEDENCE — the axis is required-with-default (absence resolves to that default) so a third-party descriptor authored before this phase keeps validating unchanged. runtime-artifact-layout.cts's resolveTriggerSurface reads it to decide the winner among same-trigger candidates once scope rank (Install Scope Module) has already been applied. See CONTEXT.md's Runtime Artifact Layout Module entry and .gsd/phase/feat-2871-trigger-resolution/40-design.md.

claude

Axis Value Source Evidence
embeddingMode imperative https://code.claude.com/docs/en/agent-sdk/overview "The Agent SDK offers hooks to execute custom code at critical points within the agent's lifecycle. These callback functions enable developer"
commandSurface slash-file https://code.claude.com/docs/en/agent-sdk/slash-commands "Each custom command is a markdown file where the filename (without the .md extension) becomes the command name. The file content defines w"
modelMode passive https://code.claude.com/docs/en/agent-sdk/typescript "setModel(model?: string): Changes the model (only available in streaming input mode) ... model overrides the default model for this subagent"
hookBus host https://code.claude.com/docs/en/agent-sdk/python "HookEvent = Literal['PreToolUse', 'PostToolUse', 'PostToolUseFailure', 'UserPromptSubmit', 'Stop', 'SubagentStop', 'PreCompact', 'Notificati"
stateIO filesystem https://code.claude.com/docs/en/sandboxing "The sandboxed Bash tool restricts file system access, granting read and write access to the current working directory and session temp direc"
transport mcp https://code.claude.com/docs/en/mcp "Project-Scoped MCP Server Configuration in .mcp.json ... This JSON structure illustrates the format for a project-scoped MCP server configur"
runtime node https://code.claude.com/docs/en/agent-sdk/typescript "import { query } from "@anthropic-ai/claude-agent-sdk"; ... pathToClaudeCodeExecutable (string) - Specifies the path to the Claude Code CLI"
effortSurface argv https://code.claude.com/docs/en/cli-reference ; claude --help --effort <level> is a documented flag on the host's own invocation, so GSD renders the resolved universal effort straight onto the argv it spawns (#2481).
dispatch.namedDispatch true https://code.claude.com/docs/en/agent-sdk/subagents "agents: { 'code-reviewer': AgentDefinition({ description: 'Expert code reviewer.', ... }) } ... subagent_type: block.inp"
dispatch.nested true https://code.claude.com/docs/en/sub-agents "As of Claude Code v2.1.172, a subagent can spawn its own subagents, allowing delegated tasks to split into parallel subt"
dispatch.maxDepth 5 https://code.claude.com/docs/en/sub-agents "foreground subagents can spawn at any depth, blocking their parent until completion. Background subagents are limited to"
dispatch.background true https://code.claude.com/docs/en/sub-agents "Subagents can run in the foreground, blocking the main conversation and passing permission prompts to you, or in the bac"
dispatch.subagentToolkit full https://code.claude.com/docs/en/sub-agents "If all tools remain selected, the subagent inherits all tools available to the main conversation."
dispatch.backgroundDispatch false https://code.claude.com/docs/en/sub-agents "Background subagents are limited to a depth of five and cannot spawn further, "
dispatch.isolation harness-worktree https://code.claude.com/docs/en/sub-agents ; Claude Code Agent tool (Agent(isolation="worktree")) The Claude Code Agent tool accepts an isolation="worktree" harness primitive — the host's own harness creates + binds a git worktree per executor; GSD passes the flag and calls no git itself (#2584)
dispatch.maxConcurrency 20 https://code.claude.com/docs/en/sub-agents The subagent documentation states a default concurrent-subagent limit of 20, configurable via the CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS environment variable (#3673).

Sources consulted:

Cross-scope trigger shadowing and spec-root reachability (#2873, epic #2866 Phase 4, resolving #2218). GSD's claude artifactLayout installs global=[skills] and local=[commands, agents] (see "Trigger precedence" above). Claude Code's own documented precedence — personal overrides project, and a same-named skill overrides a same-named command — means both rules point the same direction when a user installs both scopes: the global skill always wins the /gsd-<name> trigger, and the project-local .claude/gsd-core/ spec tree the local command correctly points at becomes unreachable through that trigger, silently. GSD Core now (a) detects this at install time and from /gsd-health (diagnostic W028) and prints which scope wins — an advisory only, exit code unchanged; and (b) for claude at global scope only, resolves the winning skill's own workflow-spec @-include at runtime rather than pre-expanding it: the skill body carries an explicit imperative instruction that resolves .claude/gsd-core/workflows/<name>.md relative to the working directory first, falling back to ~/.claude/gsd-core/workflows/<name>.md when no local tree exists. This is instruction-following, not the guaranteed inclusion a real @-include provides — it costs the agent one file read — and is deliberately confined to this one reference, in this one runtime, at this one scope: the skill's references//templates/ includes and the local scope's own emission are unchanged. See Interpret install-shadow warnings and Install on your runtime — Claude Code.

codex

Note: ADR-1239's host matrix lists Codex as prose-only; current OpenAI Codex dev docs document slash-commands, so commandSurface is slash-file here (docs are the source of truth).

Axis Value Source Evidence
embeddingMode declarative https://developers.openai.com/codex/plugins/build "No in-process programmatic API exists. Plugins integrate through: External command execution (hooks), MCP server processes, Configuration fi"
commandSurface slash-file https://github.com/openai/codex/blob/main/codex-rs/core-skills/src/loader.rs "const SKILLS_FILENAME: &str = "SKILL.md"; ... Each skill is a folder with a SKILL.md file containing YAML frontmatter with name and descript"
modelMode passive https://github.com/openai/codex/blob/main/codex-rs/config/src/config_toml.rs "pub model_provider: Option ... model is selected by config field; no programmatic model request API"
hookBus host https://github.com/openai/codex/blob/main/codex/codex-rs/hooks/src/lib.rs "pub const HOOK_EVENT_NAMES: [&str; 10] = ["PreToolUse", "PermissionRequest", "PostToolUse", "PreCompact", "PostCompact", "SessionStart", "Us"
stateIO filesystem https://developers.openai.com/codex/concepts/sandboxing "workspace-write: The default mode allowing Codex to read files, edit within the workspace, and run routine local commands inside that bounda"
transport mcp https://github.com/openai/codex/blob/main/codex-rs/config/src/config_toml.rs "pub mcp_servers: HashMap<String, McpServerConfig> ... Definition for MCP servers that Codex can reach out to for tool calls."
runtime node https://github.com/openai/codex/blob/main/codex-cli/package.json ""engines": {"node": ">=16"} ... The npm-distributed CLI wrapper is a Node.js script (#!/usr/bin/env node)"
effortSurface argv https://github.com/openai/codex/blob/main/codex-rs/exec/src/cli.rs ; https://developers.openai.com/codex/config-reference model_reasoning_effort is a config.toml key and not a dedicated CLI flag, so the generic -c <key>=<value> override is the only argv route — which is still argv, hence argv rather than a config-file member (#2481).
dispatch.namedDispatch true https://github.com/openai/codex/blob/main/codex-rs/core/src/tools/handlers/multi_agents_spec.rs ""agent_type".to_string(), JsonSchema::string(Some(agent_type_description.to_string())) ... apply_role_to_config(&mut con"
dispatch.nested true https://developers.openai.com/codex/multi-agent "agents.max_depth defaults to 1, which allows a direct child agent to spawn but prevents deeper nesting."
dispatch.maxDepth 1 https://developers.openai.com/codex/config-reference "agents.max_depth: Maximum nesting depth allowed for spawned agent threads (root sessions start at depth 0; default: 1)"
dispatch.background true https://github.com/openai/codex/blob/main/codex-rs/core/src/tools/handlers/multi_agents_spec.rs "spawn_agent returns the spawned agent id immediately; a separate wait_agent tool polls for final status."
dispatch.subagentToolkit full https://developers.openai.com/codex/multi-agent "Subagents inherit the sandbox policy and tool surface from the parent session."
dispatch.backgroundDispatch true https://github.com/openai/codex/blob/main/codex-rs/core/templates/collab/experimental_prompt.md "Sub-agents have access to the same set of tools as you do so you must tell them if they are allowed to spawn sub-agents themselves or not." The config (codex-rs/config/src/config_toml.rs) exposes an
dispatch.isolation orchestrator-worktree https://learn.chatgpt.com/docs/environments/git-worktrees ; https://github.com/openai/codex/blob/main/codex-rs/utils/cli/src/shared_options.rs "Worktrees are available only in Codex in the ChatGPT desktop app." (no native CLI worktree) + codex exec --cd <dir> sets an explicit working root, so GSD creates+manages the worktree and points the executor at it (#2584)
dispatch.maxConcurrency undocumented https://developers.openai.com/codex/config-reference agents.max_concurrent_threads_per_session is a documented config key, but the reference states no numeric default — "Codex chooses the default" — so there is no single value to cite (#3673).

GSD integration status — Phase D dogfood complete (#2088, ADR-1239). Codex installs through the declarative embedding adapter (createDeclarativeAdapter → installRuntimeArtifacts); the hardcoded runtime === 'codex'/isCodex projection is folded into descriptor-driven runtime.hostBehaviors, and install/uninstall output is byte-parity-gated at the time (tests/fixtures/golden-install-parity/codex.json; superseded by the differential attribution check, #2724). Three capability upgrades land, each with a test driving the user-reachable surface:

  • Skill root — global skills install to the canonical $HOME/.agents/skills (Codex core-skills loader.rs user-scope root), not the deprecated $CODEX_HOME/skills fallback; local skills install to <project>/.codex/skills. The global path is declared via the global skills-kind home: ".agents" override, while the local kind intentionally has no home override. Pre-move global installs are migrated (stale ~/.codex/skills/gsd-* cleaned on both install and uninstall); local installs do not remove $HOME/.agents/skills because those skills may be intentionally global.
  • Hook events — GSD registers all documented hooks.json lifecycle events beyond SessionStart: SubagentStart, Stop, PostToolUse (#772), plus the six added in #2088 — PreToolUse, PermissionRequest, PreCompact, PostCompact, SubagentStop, UserPromptSubmit — all routed through gsd-context-monitor.js. (The descriptor extendedHookEvents field reflects the schema-valid cross-runtime subset SubagentStop/Stop/PreCompact; Codex's full event set is codex-hooks-json-native, registered directly in hooks.json.)
  • Dispatch tuning — [agents] max_depth = 1 is written explicitly into the managed config.toml block, pinning the dispatch.maxDepth: 1 axis instead of relying on codex-cli's implicit default. Because maxDepth === 1, degradationFor flattens GSD-hosted wave dispatch to single-level even though dispatch.nested/background/backgroundDispatch are all true. The block is a bare [agents] AgentsToml scalar table; it does not carry per-role [agents.gsd-*] sub-tables — those pointed config_file back at the standalone agents/gsd-*.toml files Codex already auto-discovers, so emitting them was a duplicate role registration (Codex logged "Ignoring malformed agent role definition: duplicate agent role name" once per agent) removed in #2406. validateCodexConfigSchema permits a known-scalar-only [agents] while still rejecting [[agents]] and unknown-key forms.

Sources consulted:

Documentation gaps:

  • dispatch.maxDepth is configurable (Option with no documented upper bound); the documented default is 1 but the actual enforced maximum is not stated.
  • dispatch.subagentToolkit: docs say subagents 'inherit the tool surface' but do not enumerate whether any tools are excluded.
  • runtime: the Node.js entry point is a thin launcher shim; the actual agent execution runtime is a compiled Rust binary — axis classification is ambiguous.

opencode

Axis Value Source Evidence
embeddingMode imperative https://opencode.ai/docs/plugins "Plugins are JavaScript/TypeScript modules that export plugin functions; they register hooks via `import type { Plugin } from '@opencode-ai/p'"
commandSurface slash-file https://opencode.ai/docs/commands ""Create markdown files in the commands/ directory to define custom commands." and "The markdown file name becomes the command name."
modelMode active /anomalyco/opencode (Context7) — packages/plugin/src/v2/promise/README.md "ctx.aisdk.sdk(async (event) => { ... event.sdk = mod.createXai(event.options) }) and `ctx.aisdk.language((event) => { ... event.language ="
hookBus host https://opencode.ai/docs/plugins "Host fires events including: tool.execute.before, tool.execute.after, session.created, session.compacted, session.deleted"
stateIO filesystem https://opencode.ai/docs/plugins "Plugin context includes directory (working directory path), worktree (git worktree path), and $ ("Bun's shell API")"
transport mcp https://opencode.ai/docs/mcp-servers ""OpenCode supports both local and remote servers." and "Once added, MCP tools are automatically available to the LLM""
runtime bun https://opencode.ai/docs/plugins ""$": Bun's shell API for executing commands" (plugin context property); "OpenCode runs bun install at startup""
effortSurface argv https://opencode.ai/docs/cli ; opencode run --help --variant is accepted on opencode run, so the resolved effort is deliverable as an invocation argument (#2481).
dispatch.namedDispatch true https://opencode.ai/docs/agents ""Subagents can be invoked: Automatically by primary agents for specialized tasks based on their descriptions. Manually b"
dispatch.nested undocumented no authoritative doc — searched: https://opencode.ai/docs/agents —
dispatch.maxDepth undocumented no authoritative doc — searched: https://opencode.ai/docs/agents —
dispatch.background false https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/effect/runtime-flags.ts ; https://github.com/anomalyco/opencode/issues/29638 "experimentalBackgroundSubagents: enabledByExperimental(\"OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS\") — enabledByExperimental falls back to the experimental flag, and bool() defaults to false, so the Task tool's background parameter is hidden from the model unless the operator opts in by env var. #29638 (OPEN) confirms the session loop still tasks.pop()s one subtask at a time. (#2598 — corrects #2087, whose "v1.17 default-on in all modes" reading does not hold against current dev)"
dispatch.subagentToolkit full https://opencode.ai/docs/agents "The 'general' subagent "Has full tool access (except todo), so it can make file changes when needed.""
dispatch.backgroundDispatch false https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/effect/runtime-flags.ts ; https://github.com/anomalyco/opencode/issues/29638 ; https://github.com/anomalyco/opencode/issues/14195 "Concurrent dispatch requires the opt-in OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS flag (default false), so it cannot be relied on. #14195: "the session loop does tasks.pop() to grab a single subtask, awaits it, then continues the loop — so even 3 simultaneous Task calls run sequentially." Declaring true would overstate the capability, against the fail-closed posture negotiation is built for. (#2598 — corrects #2087)"
dispatch.isolation orchestrator-worktree https://opencode.ai/docs/cli ; opencode.ai/docs/plugins ; opencode issues #14195/#29638/#5887 "opencode run --dir <path> sets an explicit working root at the process level" — native subagent dispatch is synchronous-only, so GSD creates + manages the worktree and process-spawns the executor into it via --dir (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.nested
  • dispatch.maxDepth

cursor

Axis Value Source Evidence
embeddingMode imperative https://cursor.com/docs/sdk/typescript "local.customTools where you define tool functions that execute 'in your process, so it can reach anything your code can'; Agent.create()"
commandSurface slash-file https://cursor.com/docs/enterprise/llm-safety-and-controls "Commands are reusable prompts invoked via slash commands (e.g., /test), while workflows enable multi-step processes"
modelMode passive https://cursor.com/docs/sdk/python "The model used for a run can be overridden by passing a ModelSelection object in SendOptions to agent.send()."
hookBus host https://cursor.com/docs/hooks "Agent hooks: sessionStart, sessionEnd, preToolUse, postToolUse, subagentStart, subagentStop, beforeShellExecution, afterShellExecution"
stateIO filesystem https://cursor.com/docs/reference/sandbox "Local agents run with sandbox options disabled by default."
transport mcp https://cursor.com/docs/mcp "The Model Context Protocol (MCP) allows Cursor to connect to external tools and data sources."
runtime node https://cursor.com/docs/sdk/typescript "The SDK runs on Node.js. It requires Node.js 22.13 or later and is described as a Node-first package."
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://cursor.com/docs/subagents "Invoke specific subagents using slash commands in your prompt. This allows for direct control over which agent performs"
dispatch.nested true https://cursor.com/docs/sdk/typescript "The top-level agent and its direct subagents can launch subagents, but a subagent launched by another subagent can't lau"
dispatch.maxDepth 2 https://cursor.com/docs/sdk/typescript "The top-level agent and its direct subagents can launch subagents, but a subagent launched by another subagent can't lau"
dispatch.background true https://cursor.com/docs/subagents "Background, which returns immediately while the subagent works independently, best for long-running tasks or parallel wo"
dispatch.subagentToolkit full https://cursor.com/docs/subagents "Subagents can utilize MCP tools, inheriting all tools available to their parent agent, including those from configured s"
dispatch.backgroundDispatch true https://cursor.com/docs/subagents (FAQ: Can subagents launch other subagents?) and https://cursor.com/docs/sdk/typescript (Subagents > Nested subagents) FAQ: "As of Cursor 2.5, subagents have the capability to launch child subagents, enabling the creation of a hierarchical structure for coordinated tasks. This nested launching functionality requires T
dispatch.isolation harness-worktree https://cursor.com/docs/cli/reference/parameters ; cursor.com/docs/cli/using ; cursor.com/docs/cli/changelog "-w, --worktree [name] — cursor-agent creates/binds a git worktree per agent (~/.cursor/worktrees/…); native parallel-agent dispatch" (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

GSD integration status — Phase D dogfood complete (#2089, ADR-1239). Cursor installs through the imperative embedding adapter (createImperativeAdapter → installRuntimeArtifacts); the hardcoded runtime === 'cursor' / isCursor projection is folded into descriptor-driven runtime.hostBehaviors, and install/uninstall output is byte-parity-gated at the time (tests/fixtures/golden-install-parity/cursor.json; superseded by the differential attribution check, #2724). Two capability upgrades land, each with a test driving the user-reachable surface:

  • Expanded hook-bus coverage — GSD registers all 6 managed lifecycle events in hooks.json beyond the original sessionStart/postToolUse: preToolUse, stop, subagentStart, subagentStop (AC4a, cite https://cursor.com/docs/hooks). The hook-bus binding is descriptor-driven via src/host-integration-adapters/imperative-hook-bus.cts (reads hostBehaviors.managedHookEvents), not a hardcoded event pair.
  • Named/background nested subagent dispatch — dispatch.background/backgroundDispatch/nested are all true with maxDepth: 2; shouldFlattenDispatch(cursor) returns false so GSD's wave-based execution drives Cursor's native background + depth-2 nested subagent dispatch instead of flattening to inline sequential calls (AC4b, cite https://cursor.com/docs/subagents + https://cursor.com/docs/sdk/typescript).

cline

Axis Value Source Evidence
embeddingMode imperative /cline/cline (Context7) — https://github.com/cline/cline/blob/main/docs/sdk/plugins.mdx "Implement the AgentPlugin interface to register tools, hooks, and configuration. The setup function is used for registering capabilities."
commandSurface slash-file /cline/cline (Context7) — https://github.com/cline/cline/blob/main/cline/apps/vscode/src/test/slash-commands.test.ts "workflow markdown files (with .md, .markdown, or .txt extensions) are invoked as slash commands using their filename."
modelMode active /cline/cline (Context7) — https://github.com/cline/cline/blob/main/sdk/packages/llms/README.md "The Runtime API, accessible via createLlmsRuntime(...), allows for the creation of a registry that manages configured providers and their de"
hookBus host /cline/cline (Context7) — https://github.com/cline/cline/blob/main/sdk/README.md "Package agent capabilities as extensions (plugins) that can register tools, observe lifecycle events, and modify agent behavior."
stateIO filesystem /cline/cline (Context7) — https://github.com/cline/cline/blob/main/cline/sdk/packages/shared/src/storage/paths.ts "resolveClineDir() returns ~/.cline; resolveDocumentsExtensionPath('Workflows') returns ~/Documents/Cline/Workflows."
transport mcp /cline/cline (Context7) — https://github.com/cline/cline/blob/main/docs/mcp/mcp-overview.mdx "MCP (Model Context Protocol) enables Cline to interact with external tools and data sources"
runtime node /cline/cline (Context7) — https://github.com/cline/cline/blob/main/sdk/examples/plugins/typescript-lsp/README.md "Installs a portable subagent plugin ... cp examples/plugins/agents-squad/index.ts ~/.cline/plugins/portable-subagents.ts."
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true /cline/cline (Context7) — https://github.com/cline/cline/blob/main/sdk/examples/plugins/agents-squad/README.md "parent → start_subagent(preset: "phantom", task: "Map the auth module") → phantom: save_handoff(...)"
dispatch.nested false /cline/cline (Context7) — https://github.com/cline/cline/blob/main/docs/features/subagents.mdx "subagents are restricted from editing files, using the browser, accessing MCP servers, or creating nested subagents."
dispatch.maxDepth 1 /cline/cline (Context7) — https://github.com/cline/cline/blob/main/docs/features/subagents.mdx "They are explicitly prohibited from ... spawning other subagents."
dispatch.background true /cline/cline (Context7) — https://github.com/cline/cline/blob/main/docs/features/subagents.mdx "Commands executed by subagents run in the background and are strictly limited to read-only operations"
dispatch.subagentToolkit read-only /cline/cline (Context7) — https://github.com/cline/cline/blob/main/docs/features/subagents.mdx "Subagents are equipped with tools for read-only operations, including reading file contents (read_file), listing directo"
dispatch.backgroundDispatch false https://docs.cline.bot/features/subagents (mirrored at https://github.com/cline/cline/blob/main/docs/features/subagents.mdx) "They cannot edit files, use the browser, or spawn nested subagents" — and from the GitHub source: "subagents are restricted from editing files, using the browser, accessing MCP servers, or creating n
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

GSD integration status — Phase D dogfood complete (#2090, ADR-1239). Cline installs through the imperative embedding adapter (createImperativeAdapter → installRuntimeArtifacts); the hardcoded runtime === 'cline' / isCline projection is folded into descriptor-driven runtime.hostBehaviors, and install/uninstall output is byte-parity-gated at the time (tests/fixtures/golden-install-parity/cline.json; superseded by the differential attribution check, #2724). Two capability upgrades land, each with a test driving the user-reachable surface:

  • AgentPlugin.hooks.beforeTool planning guard — the .clinerules/hooks/PreToolUse file-convention hook (#787) is re-implemented as a real Cline SDK AgentPlugin registered through the negotiated hookBus: host interface point. Guard semantics are preserved exactly (fail-open, cancels write-class calls targeting .planning/); the SDK maps the file hook's {cancel, errorMessage} to {skip, reason}. The binding lives in src/host-integration-adapters/cline-sdk-binding.cts (cite https://github.com/cline/cline/blob/main/docs/sdk/plugins.mdx).
  • createAgentModel per-subagent model overrides — DefaultGateway.createAgentModel({providerId, modelId}) is wired so GSD's model_overrides / model_profile_overrides resolution (already used for OpenCode/Codex passive hosts) applies to cline subagents (modelMode: active), instead of leaving model selection untouched (cite https://github.com/cline/cline/blob/main/docs/sdk/reference/gateway.mdx).
  • Dispatch stays degraded/flat (deliberate) — unlike cursor's dispatch upgrade, cline's dispatch is maxDepth: 1, nested: false, subagentToolkit: 'read-only', backgroundDispatch: false. shouldFlattenDispatch(cline) returns true and degradationFor('dispatch', cline) returns {level:'degraded', fallback:'flat dispatch — waves run inline'}. This is NOT upgraded: cline's own docs restrict subagents to a single level with a read-only toolkit and no nested spawning, so claiming full dispatch would misrepresent the host and violate the fail-closed negotiation contract (cite https://github.com/cline/cline/blob/main/docs/features/subagents.mdx).

hermes

Axis Value Source Evidence
embeddingMode imperative https://hermes-agent.nousresearch.com/docs/guides/build-a-hermes-plugin "ctx.register_tool() puts your tool in the registry — the model sees it immediately"
commandSurface slash-programmatic https://hermes-agent.nousresearch.com/docs/guides/build-a-hermes-plugin "ctx.register_command('mystatus', handler=_handle_status, description='Show plugin status') — The command appears in autocomplete, /help output"
modelMode active https://hermes-agent.nousresearch.com/docs/guides/build-a-hermes-plugin "register_provider(ProviderProfile(name=..., aliases=(...), display_name=..., env_vars=(...), base_url=..., auth_type=..., default_aux_model="
hookBus host https://hermes-agent.nousresearch.com/docs/user-guide/features/hooks "Hermes owns and manages the entire hook infrastructure. At runtime, HookRegistry.discover_and_load() scans ~/.hermes/hooks/"
stateIO filesystem https://hermes-agent.nousresearch.com/docs/user-guide/configuration "The agent has the same filesystem access as your user account."
transport mcp https://hermes-agent.nousresearch.com/docs/user-guide/features/mcp "MCP support ships with the standard install — no extra step needed."
runtime python Context7 /nousresearch/hermes-agent "The plugin and agent runtime is Python (confirmed by register(ctx) in init.py, importlib.import_module, run_agent.py, tools/registry.py)"
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch false https://hermes-agent.nousresearch.com/docs/user-guide/features/delegation "The documentation contains no mention of named agents. Subagents are identified only by role ('leaf' or 'orchestrator')"
dispatch.nested true /nousresearch/hermes-agent (Context7) — configuration.md "max_spawn_depth: 1 — Delegation tree depth cap (1-3, clamped). 1 = flat (default): parent spawns leaves that cannot dele"
dispatch.maxDepth 1 /nousresearch/hermes-agent (Context7) — configuration.md "max_spawn_depth: 1 # Delegation tree depth cap (1-3, clamped). 1 = flat (default): parent spawns leaves that cannot dele"
dispatch.background true https://github.com/NousResearch/hermes-agent/releases/tag/v2026.6.19 "delegate_task(background=true) dispatches a subagent that runs in the background and returns a handle immediately"
dispatch.subagentToolkit read-only https://hermes-agent.nousresearch.com/docs/guides/delegation-patterns "Nested delegation is opt-in; by default, leaf subagents cannot call delegate_task, clarify, memory, send_message, or exe"
dispatch.backgroundDispatch false https://github.com/nousresearch/hermes-agent/blob/main/website/docs/user-guide/features/delegation.md (via Context7 query of /nousresearch/hermes-agent) "Nested delegation is an opt-in feature, requiring role="orchestrator" for children and an increased max_spawn_depth from its default of 1. It can also be globally disabled with orchestrator_enabled
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

EoS migration status (#2091): Migrated onto the imperative adapter. All runtime === 'hermes' branches in bin/install.js folded into descriptor-driven runtime.hostBehaviors. New extensionEvents: "hermes" dialect registered (13 real plugin hook events, replacing the borrowed hookEvents: "claude" 6-event surface). Cite: https://github.com/nousresearch/hermes-agent/blob/main/website/docs/user-guide/features/hooks.md

Documentation gaps:

  • runtime — Hermes plugins and agent core run in Python, but this was confirmed by code inspection rather than explicit docs statement.
  • dispatch.namedDispatch — docs explicitly confirm no named-agent dispatch in delegate_task; Kanban has named profiles but that is a separate board system not a dispatch mechanism.

antigravity

Axis Value Source Evidence
embeddingMode declarative https://github.com/alphaperseii3000/google-antigravity-docs/blob/master/google-antigravity-docs.md "Skills require a SKILL.md file; Workflows are saved as markdown files; Rules are manually defined constraints — all configuration-file-based"
commandSurface slash-file https://github.com/alphaperseii3000/google-antigravity-docs/blob/master/google-antigravity-docs.md "Workflows are saved as markdown files, providing a repeatable method for executing key processes. They can be invoked in the Agent using a s"
modelMode passive https://dev.to/arindam_1729/antigravity-cli-a-hands-on-guide-to-googles-terminal-coding-agent-5bc7 "Selection occurs via -m flag or /model command inside the TUI. No programmatic model request API is documented for extensions/skills"
hookBus host https://www.aibuilderclub.com/blog/antigravity-cli-guide "The CLI fires hooks, not the engine. These are JSON lifecycle interceptors (before tool call, after file edit, on session start)."
stateIO filesystem live agy 1.1.17 install probe — https://github.com/open-gsd/gsd-core/issues/3747 "Global skills discovery scans ~/.gemini/config/skills/: a probe skill placed there is visible to agy, while an identical probe skill under the configHome is not. An earlier third-party blog claim about configHome skills placement was disproved by that probe; GSD installs global skills/agents under ~/.gemini/config (#3738)"
transport mcp https://dev.to/arindam_1729/antigravity-cli-a-hands-on-guide-to-googles-terminal-coding-agent-5bc7 "Both local (stdio) and remote (HTTP) Model Context Protocol servers are supported"
runtime go https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/ "Built in Go, Antigravity CLI is snappier and more responsive."
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch undocumented no authoritative doc — searched: https://www.aibuilderclub.com/blog/antigravity-cli-guide, https://antigravity.google/docs/agents —
dispatch.nested undocumented no authoritative doc — searched: https://antigravity.google/docs/agents —
dispatch.maxDepth undocumented no authoritative doc — searched: https://antigravity.google/docs/agents —
dispatch.background true https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/ "Antigravity CLI orchestrates multiple agents for complex tasks in the background"
dispatch.subagentToolkit full https://antigravity.google/docs/cli/features "Capabilities: Subagents have full access to tools such as code search, file editing, terminal commands, and web searches to complete their assigned tasks." (#2096 EoS migration — the page is JS-rendered/blank on a static fetch; confirmed via headless-browser render)
dispatch.backgroundDispatch undocumented no authoritative doc — Multiple sources consulted: antigravity.google/docs/cli-subagents (returned blank/JS-rendered), antigravity.google/docs/agent (blank), github.com/google-antigravity/antigravity-cli README, Context7 /google-antigravity/antigravity-cli All documentation consulted describes a two-level orchestrator→subagent architecture. Background subagents run asynchronously while the main agent continues accepting prompts. The DataCamp tutorial st
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.namedDispatch — docs describe dynamic plain-English goal dispatch where agent names subagents at runtime; no pre-registered named sub-agent API documented.
  • dispatch.nested — no documentation found on whether subagents can themselves spawn further subagents.
  • dispatch.maxDepth — no documented depth limit or explicit unbounded statement found.

EoS migration status (#2096): Migrated onto the declarative adapter. All runtime === 'antigravity' / isAntigravity / canonical === 'antigravity' branches folded into descriptor-driven runtime.hostBehaviors + runtime.hostIntegration: getConfigDirFromHome (bin/install.js) now branches on configHome.kind === 'dot-home-nested' instead of a hardcoded runtime literal; projectLocalHookPrefix (src/shell-command-projection.cts) reads hostBehaviors.hookPathStyle ('raw' → bare dirName, no $CLAUDE_PROJECT_DIR anchor); applyAgentPathRewrites (src/runtime-artifact-conversion.cts) reads hostBehaviors.noPathRewrite to skip the ~/.claude/ → pathPrefix rewrites; and getProjectInstructionFile (src/runtime-name-policy.cts) reads hostBehaviors.projectInstructionFile ("GEMINI.md" — Antigravity CLI's contextFileName, successor to the sunset Gemini CLI per #1928) instead of a hardcoded canonical === 'antigravity' check. The dead isAntigravity branches these functions previously carried are removed. dispatch.subagentToolkit flipped undocumented → full per the citation above (antigravity.google/docs/cli/features); dispatch.namedDispatch/nested/maxDepth/backgroundDispatch stay undocumented — no authoritative source states named/nested/depth-bounded dispatch or a run_in_background-style call-time param, so negotiateHostCapabilities degrades all four closed to their most-restrictive value (false/0), and shouldFlattenDispatch still forces antigravity's dispatch to flatten (inline) despite dispatch.background: true, because backgroundDispatch itself never reaches true. Two upgrades land: UPGRADE 1 — permission-writer (configureAntigravityPermissions, runtime.permissionWriter: "antigravity") writes Antigravity's native {"permissions":{"allow":[...]}} schema (antigravity.google/docs/cli/permissions) into the same settings.json GSD's own hook registration writes, granting GSD's own read_file/command rules non-destructively. UPGRADE 2 — MCP companion config (configureAntigravityMcpConfig) writes a standalone mcp_config.json (antigravity.google/docs/cli/gcli-migration) registering the gsd MCP server, non-destructively preserving any other mcpServers entries. Both upgrades are covered by tests/antigravity-upgrades.test.cjs; the axis/negotiation/source-grep coverage above is in tests/declarative-reference-antigravity.test.cjs.

augment

Axis Value Source Evidence
embeddingMode declarative https://docs.augmentcode.com/cli/plugins "Plugins can provide several types of components, including Custom Commands defined in Markdown files within the commands/ directory... Hoo"
commandSurface slash-file https://docs.augmentcode.com/cli/plugins "Slash commands are Markdown files in the commands/ directory. The filename becomes the command name"
modelMode passive https://docs.augmentcode.com/cli/subagents "
hookBus host https://docs.augmentcode.com/cli/hooks "Hook event types: PreToolUse (before a tool executes), PostToolUse (immediately after a tool completes), Stop (when the agent stops respondi"
stateIO filesystem https://github.com/augmentcode/auggie "Node.js 22+ required. Hook configurations use ${AUGMENT_PLUGIN_ROOT}"
transport mcp https://docs.augmentcode.com/cli/plugins "Auggie supports a plugin system that allows you to extend its functionality with... MCP server integrations."
runtime node https://github.com/augmentcode/auggie "Node.js 22+ required"
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://docs.augmentcode.com/cli/subagents "
dispatch.nested undocumented no authoritative doc — searched: https://docs.augmentcode.com/cli/subagents —
dispatch.maxDepth undocumented no authoritative doc — searched: https://docs.augmentcode.com/cli/subagents —
dispatch.background true https://docs.augmentcode.com/cli/subagents "Subagents run in parallel with other subagents... will show a summary of their current progress in the main thread."
dispatch.subagentToolkit full https://docs.augmentcode.com/cli/subagents "If neither [tools nor disabled_tools] is specified, the subagent has access to all tools (default behavior)."
dispatch.backgroundDispatch undocumented no authoritative doc — https://docs.augmentcode.com/cosmos/automations The Augment Code (Cosmos) docs describe workers as 'sub-agents launched mid-session by a manager Expert using the worker-launch command. Each worker is its own session with its own messages and permis
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.nested
  • dispatch.maxDepth

EoS migration status (#2097): Folded onto descriptor-driven dispatch. Augment already installed through the declarative adapter (nested-skill artifact layout, settings-json hook surface, Claude hook event dialect), but carried two remaining runtime-literal branches in src/runtime-artifact-conversion.cts: the 4 ~/.augment/$HOME/.augment dot-dir rewrites in _applyRuntimeRewrites's case 'augment': block are now built from getDirName('augment') (dirName-derived, byte-identical) instead of a hardcoded .augment literal, and the applyRuntimeContentRewritesForCommandsInPlace command-body conversion dispatch now reads runtime.hostBehaviors.commandBodyConverter ("convertClaudeToAugmentMarkdown") instead of a hardcoded runtime === 'augment' branch. Two dead-code sites were also removed from bin/install.js: the orphaned claudeToAugmentTools map (superseded by the single-sourced converters per ADR-1508 / #1675) and the unreachable else if (isAugment) { content = convertClaudeAgentToAugmentAgent(content); } inline agent-conversion branch (augment has been on the descriptor-agents path since _DESCRIPTOR_AGENTS_RUNTIMES was introduced, making that if/else if arm dead). UPGRADE 3 — MCP companion config (mergeGsdMcpServerIntoSettings) registers the gsd MCP server directly inside the same settings.json mcpServers block GSD's own hook registration already writes (Augment hosts MCP in settings.json, unlike Antigravity's standalone mcp_config.json) — non-destructively preserving any other user-configured mcpServers entries; uninstall removes only the GSD-owned gsd entry. settings.json is golden-excluded (HOOK_CONFIG_FILES), so this upgrade produces no golden fixture change. Source-grep guard + fail-closed negotiation coverage is in tests/declarative-reference-augment.test.cjs; the dispatch/hook-bus/MCP upgrade coverage is in tests/augment-upgrades.test.cjs.

qwen

Axis Value Source Evidence
embeddingMode imperative https://qwenlm.github.io/qwen-code-docs/en/developers/channel-plugins "Your entry point exports a ChannelPlugin object... this.registerCommand('mycommand', async (envelope, args) => { ... }); ... plugins load at startup as extensions."
commandSurface slash-file https://qwenlm.github.io/qwen-code-docs/en/users/extension/introduction "Extensions can provide custom commands by placing Markdown files in a commands/ subdirectory"
modelMode passive https://qwenlm.github.io/qwen-code-docs/en/developers/channel-plugins "The documentation does not expose a direct API for plugins to invoke the LLM or model directly."
hookBus host https://qwenlm.github.io/qwen-code-docs/en/users/features/hooks "Qwen Code provides 14 distinct hook events: PreToolUse, PostToolUse, PostToolUseFailure, UserPromptSubmit, SessionStart, SessionEnd, Stop"
stateIO filesystem https://qwenlm.github.io/qwen-code-docs/en/developers/channel-plugins "Runtime Environment: Node.js only. The architecture uses standard Node.js APIs: import, async/await, file I/O (writeFileSync), OS utilities"
transport mcp https://qwenlm.github.io/qwen-code-docs/en/developers/tools/mcp-server "Qwen Code integrates with MCP servers through a sophisticated discovery and execution system"
runtime node https://qwenlm.github.io/qwen-code-docs/en/developers/channel-plugins "Language: Node.js (TypeScript/JavaScript). Execution model: In-process — plugins load at startup as extensions."
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://qwenlm.github.io/qwen-code-docs/en/users/features/sub-agents/ "Named subagents are invoked when the AI identifies tasks matching their specialization... Users can also explicitly requ"
dispatch.nested false https://qwenlm.github.io/qwen-code-docs/en/users/features/sub-agents/ "Fork children cannot create further forks. This is enforced at runtime — if a fork attempts to spawn another fork, it re"
dispatch.maxDepth 1 https://qwenlm.github.io/qwen-code-docs/en/users/features/sub-agents/ "Fork children cannot create further forks. This is enforced at runtime"
dispatch.background true https://qwenlm.github.io/qwen-code-docs/en/users/features/sub-agents/ "Runs in background, parent continues immediately... Forks run parallel to the parent; the main conversation continues im"
dispatch.subagentToolkit full https://qwenlm.github.io/qwen-code-docs/en/users/features/sub-agents/ "When omitted, the subagent inherits all available tools from the parent session."
dispatch.backgroundDispatch false https://qwenlm.github.io/qwen-code-docs/en/users/features/sub-agents/ (official Qwen Code documentation, 'Subagents' user guide page) and https://qwenlm.github.io/qwen-code-docs/en/design/fork-subagent/fork-subagent-design (Qwen Code fork-subagent design document, section '4. Recursive Fork Prevention') The official user-facing Qwen Code docs state verbatim: "Fork children cannot create further forks. If a fork attempts spawning another fork, it receives an error instructing direct task execution ins
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.nested — docs only restrict fork-type sub-agents from nesting; whether named sub-agents can themselves spawn named sub-agents is not stated.
  • dispatch.maxDepth — depth=1 is documented only for fork sub-agents; depth for named sub-agent chains is undocumented.

EoS migration status (#2092): Migrated onto the imperative adapter. All runtime === 'qwen' branches in bin/install.js, src/install-engine.cts, src/runtime-artifact-conversion.cts, and src/runtime-hooks-surface.cts folded into descriptor-driven runtime.hostBehaviors. Two upgrades land: (1) native subagent projection — a new agents artifact-layout kind projects GSD's specialist agents into ~/.qwen/agents/gsd-*.md as native Qwen subagents via convertClaudeAgentToQwenAgent, emitting Qwen's own name:/description:/tools: (YAML block list) frontmatter schema instead of Claude Code's; cite https://qwenlm.github.io/qwen-code-docs/en/users/features/sub-agents/. (2) SubagentStart hook — wired into extendedHookEvents alongside the existing SubagentStop/Stop/PreCompact events, firing the context-monitor hook symmetrically at subagent start and completion; cite https://qwenlm.github.io/qwen-code-docs/en/users/features/hooks.

codebuddy

Axis Value Source Evidence
embeddingMode declarative https://www.codebuddy.ai/docs/cli/plugins-reference "Commands are 'plain Markdown file[s]' located in commands/ by default ... a skill is a directory containing a SKILL.md ... The documentation"
commandSurface slash-file https://www.codebuddy.ai/docs/cli/plugins-reference "Commands are 'plain Markdown file[s]' located in commands/ by default ... Skills are prefixed with this (e.g., /my-first-plugin:hello)"
modelMode passive https://www.codebuddy.ai/docs/cli/sdk "The SDK is not for building plugins that run inside CodeBuddy. It's an external SDK for standalone applications"
hookBus host https://www.codebuddy.ai/docs/cli/hooks "Full support for the hook event family (27+ events), covering tool lifecycle (PreToolUse / PostToolUse / PostToolUseFailure)"
stateIO filesystem https://www.codebuddy.ai/docs/cli/settings "Storage operates in non-sandboxed mode by default ... Default: Full filesystem access governed by permission rules"
transport mcp https://www.codebuddy.ai/docs/cli/cli-reference "MCP (Model Context Protocol) is built-in as a core feature ... codebuddy mcp command to 'Configure Model Context Protocol (MCP) servers'"
runtime node https://www.codebuddy.ai/docs/cli/sdk "TypeScript/JavaScript: Node.js >= 18.20 ... npm install @tencent-ai/agent-sdk"
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://www.codebuddy.ai/docs/cli/sub-agents "Sub-agents can be invoked explicitly by name: 'Request a specific sub-agent by mentioning it in your command'"
dispatch.nested false https://www.codebuddy.ai/docs/cli/sub-agents "This prevents infinite nesting of agents (sub-agents cannot spawn other sub-agents)"
dispatch.maxDepth 1 https://www.codebuddy.ai/docs/cli/sub-agents "The architecture enforces exactly one level of nesting — only the main CodeBuddy Code instance can invoke sub-agents."
dispatch.background true https://www.codebuddy.ai/docs/cli/sub-agents "Launch a background agent using the run_in_background: true parameter ... Tasks return immediately with an ID"
dispatch.subagentToolkit full https://www.codebuddy.ai/docs/cli/sub-agents "By default, sub-agents inherit all tools when the tools field is omitted ... Sub-agents can access MCP tools from config"
dispatch.backgroundDispatch false https://www.codebuddy.ai/docs/cli/sub-agents "This prevents infinite nesting of agents (sub-agents cannot spawn other sub-agents)" — the restriction is stated as universal in the Sub-Agents documentation page. The daemon/background docs (https:/
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

EoS migration status (#2098): Migrated onto the declarative adapter (dogfooded in tests/declarative-reference-codebuddy.test.cjs). The two remaining isCodebuddy branches in bin/install.js — a duplicate commands/ slash-command output report, and a dead legacy agent-converter dispatch arm (unreachable since codebuddy is in _DESCRIPTOR_AGENTS_RUNTIMES) — were folded onto the generic runtime.hostBehaviors.reportCommandsDir and removed outright; isCodebuddy no longer appears as a live read anywhere in bin/install.js, src/runtime-artifact-conversion.cts, src/shell-command-projection.cts, or src/runtime-name-policy.cts. Cursor also used that report flag historically, but retired its parallel command surface in #2644 because Cursor skills already appear in the slash menu. Two upgrades land: (1) extended hook events — codebuddy's extendedHookEvents was previously [] (none wired); this PR wires all four — SubagentStop/Stop/PreCompact/SubagentStart — into extendedHookEvents (mirrors qwen/kimi), so an install now registers all four as hooks in settings.json alongside the pre-existing base session/tool events (SessionStart/PreToolUse/PostToolUse); cite https://www.codebuddy.ai/docs/cli/hooks. (2) dispatch.background — the descriptor already declared true, exceeding the declarative-cli profile baseline of false; the negotiation contract (negotiateHostCapabilities) now surfaces that value with no downgrade warning, documenting the legitimate deviation. Note: the CodeBuddy CLI has no background-dispatch frontmatter field on sub-agents (agentMode/enabledAutoRun are IDE-only per https://www.codebuddy.ai/docs/cli/sub-agents) — background dispatch remains a caller-side invocation parameter (run_in_background: true), not a field GSD's agent artifacts emit.

copilot

Axis Value Source Evidence
embeddingMode declarative https://docs.github.com/en/copilot/concepts/agents/copilot-cli/comparing-cli-features "Declarative elements include custom instructions, skills, custom agents, and plugin configurations—all defined through configuration files"
commandSurface slash-file https://docs.github.com/en/copilot/concepts/agents/copilot-cli/comparing-cli-features "Skills: Markdown files with instructions for specific contexts. Users can invoke via slash commands (e.g., /Markdown-Checker check README.md)"
modelMode passive https://github.com/github/copilot-sdk/blob/main/docs/auth/byok.md "Model selection via config: model: 'gpt-4.1', provider: { type: 'openai', ... }."
hookBus host https://docs.github.com/en/copilot/reference/hooks-reference "Hooks allow you to extend and customize the behavior of GitHub Copilot agents by executing custom shell commands at key points during agent"
stateIO filesystem https://docs.github.com/en/copilot/how-tos/copilot-cli/customize-copilot/add-mcp-servers "Configuration file Location: ~/.copilot/mcp-config.json. Hook config files stored in .github/hooks/*.json"
transport mcp https://docs.github.com/en/copilot/how-tos/copilot-cli/customize-copilot/add-mcp-servers "Copilot CLI comes with the GitHub MCP server already configured. STDIO is the standard transport."
runtime undocumented no authoritative doc — searched: https://github.com/github/copilot-cli/blob/main/README.md, https://github.com/github/copilot-sdk/blob/main/nodejs/README.md —
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://github.com/github/copilot-sdk/blob/main/docs/features/custom-agents.md "A custom agent is a named agent configuration that includes its own prompt and tool set. A sub-agent is a custom agent i"
dispatch.nested false https://awesome-copilot.github.com/learning-hub/agents-and-subagents/ "By default, subagents do not keep spawning additional subagents."
dispatch.maxDepth 1 https://awesome-copilot.github.com/learning-hub/agents-and-subagents/ "Depth counts how many agents are nested within one another. When the depth limit is reached, the innermost agent cannot"
dispatch.background true https://docs.github.com/en/copilot/how-tos/copilot-cli/speed-up-task-completion "Allow Copilot to use subagents and work autonomously to implement the plan without any further input."
dispatch.subagentToolkit full https://docs.github.com/en/copilot/how-tos/copilot-cli/customize-copilot/create-custom-agents-for-cli "By default, custom agents have access to all tools. If you restrict an agent's access, a tools specification is added"
dispatch.backgroundDispatch false https://code.visualstudio.com/docs/copilot/agents/subagents "By default, subagents cannot spawn further subagents. This prevents infinite recursion when agents accidentally call themselves in a loop." The setting chat.subagents.allowInvocationsFromSubagents
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • runtime — docs describe the CLI binary and the SDK (Node.js/Go/Python/Rust) but do not state what runtime the CLI host itself or its plugin/extension loader executes in.
  • dispatch.nested exact authoritative source is awesome-copilot.github.com (community docs) not docs.github.com.

EoS migration status (#2099): Migrated onto the declarative adapter (dogfooded in tests/declarative-reference-copilot.test.cjs). The residual isCopilot branches were folded onto descriptor-driven runtime.hostBehaviors: the .agent.md destination-suffix rename in src/install-engine.cts now reads hostBehaviors.agentFileExtension; bin/install.js's two uninstall side-effect branches (repo-root AGENTS.md cleanup, copilot-instructions.md/hook cleanup) now gate on resolveInstallPlan(runtime).installSurface === 'copilot-instructions' (unique to copilot, so byte-identical); and the two skipSharedHooksInstall checks now read hostBehaviors.skipSharedHooksInstall:true (copilot's golden has only hooks/gsd-session.json, no shared gsd-*.js scripts). A dead legacy agent-converter dispatch arm in the inline agent-copy loop — unreachable since copilot is a member of _DESCRIPTOR_AGENTS_RUNTIMES — was removed outright; isCopilot no longer appears as a live read anywhere in bin/install.js or src/install-engine.cts. Two upgrades land: (1) multi-event hook bus — buildCopilotHookConfig() previously emitted only sessionStart; this PR wires four additional events — preToolUse/postToolUse/userPromptSubmitted/sessionEnd — each a static, deterministic advisory command (no node-runner invocation), so an install's hooks/gsd-session.json now registers all five events. (2) dispatch.background — the descriptor already declared true, exceeding the declarative-cli profile baseline of false; the negotiation contract (negotiateHostCapabilities) surfaces that value with no downgrade warning, documenting the legitimate deviation. Note: Copilot's .agent.md frontmatter has no background-dispatch field (fields are description/infer/mcp-servers/model/name/tools) — background dispatch remains a negotiated-contract-only axis, not a field GSD's agent artifacts emit. MCP companion tooling is out of scope for this migration (AC4 names only the two upgrades above).

kilo

Axis Value Source Evidence
embeddingMode imperative https://kilo.ai/docs/automate/extending/plugins "Plugins extend Kilo by hooking into events and adding functionality. They can: add custom tools the model can call (like read, write, bash)"
commandSurface slash-file https://kilo.ai/docs/customize/workflows "Workflows, also known as slash commands, allow users to automate repetitive tasks by defining step-by-step instructions"
modelMode active https://kilo.ai/docs/automate/extending/plugins "provider — dynamically supply model catalogs. auth — register OAuth or API-key flows for model providers. chat.params — Mutate temperature"
hookBus host https://kilo.ai/docs/automate/extending/plugins "event — fires for every internal bus event. Session: session.created, session.updated, session.idle, session.error, session.deleted"
stateIO filesystem https://kilo.ai/docs/contributing/architecture "Local execution and hosted execution are separate boundaries. Local runtime instances are Directory-keyed runtime context"
transport mcp https://kilo.ai/docs/automate/mcp/what-is-mcp "Kilo Code implements the Model Context Protocol to connect to both local and remote MCP servers"
runtime bun https://kilo.ai/docs/automate/extending/plugins "npm plugins are installed automatically at startup using Bun. Plugin context includes $ (Bun shell). Plugins are TypeScript or JavaScript mo"
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://kilo.ai/docs/customize/custom-subagents "Configured subagents can be invoked automatically by primary agents (like the Orchestrator) using the Task tool"
dispatch.nested true https://github.com/Kilo-Org/kilocode/issues/7055 "A subagent can still call the task tool if its merged permissions contain an explicit task rule, which enables nested su"
dispatch.maxDepth -1 https://github.com/Kilo-Org/kilocode/issues/8637 "there is no maximum nesting depth and the system relies entirely on permission gating"
dispatch.background true https://kilo.ai/docs/code-with-ai/agents/orchestrator-mode "Agents are also capable of launching multiple subagent sessions concurrently to facilitate parallel processing."
dispatch.subagentToolkit undocumented no authoritative doc — searched: https://kilo.ai/docs/customize/custom-subagents —
dispatch.backgroundDispatch false https://kilo.ai/docs/automate/tools/new-task "Importantly, subagents cannot spawn further subagents; only primary agents can use the new_task tool."
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.subagentToolkit — docs describe per-subagent configurable permissions (allow/ask/deny) but do not document a single default toolkit level (full vs read-only) for subagents that lack explicit permission overrides.

EoS migration status (#2093): Migrated onto the imperative adapter. All runtime === 'kilo' / isKilo logic branches in bin/install.js, src/install-engine.cts, src/runtime-artifact-conversion.cts, and src/runtime-artifact-layout.cts folded into descriptor-driven runtime.hostBehaviors (finishPermissionWriter, skipSharedHooksInstall, and the skills converter registry are now resolved off the descriptor, not a hardcoded Kilo check). Four upgrades land: (1) native hook-bus plugin — .kilo/plugins/gsd-core.js (byte-identical to .opencode/plugins/gsd-core.js, cite: Kilo is an OpenCode fork sharing the same plugin/extension event bus) bridges GSD's hook scripts onto Kilo's plugin event bus; extensionEvents: "kilo" reuses OPENCODE_EXTENSION_EVENTS verbatim. (2) active-model routing — convertClaudeToKiloFrontmatter now emits a model: field from the resolved model_overrides/model_profile_overrides.kilo.<tier> value instead of always stripping it (mirrors the OpenCode upgrade; #2256). (3) MCP companion documented — docs/how-to/connect-gsd-mcp-server.md covers Kilo's mcp-keyed config (not mcpServers) and its {type:"local", command, timeout} entry shape. (4) named subagent dispatch — GSD's specialist agents install as <configDir>/agents/gsd-*.md with mode: subagent frontmatter (the slug Kilo's Task tool dispatches by) and a permission: block. dispatch.subagentToolkit stays undocumented — no authoritative Kilo doc states a default subagent toolkit level — so degradationFor('dispatch', …) returns 'degraded', not 'full', by design (fail-closed negotiation, not a regression).

windsurf

Axis Value Source Evidence
embeddingMode declarative https://docs.devin.ai/desktop/cascade/cascade "Cascade operates through configuration files rather than code plugins: .codeiumignore for file filtering, Memories and Rules for customizing"
commandSurface slash-file https://docs.devin.ai/desktop/cascade/workflows "Workflows are authored as markdown files (.md extension) … triggered through slash commands using the format /[workflow-name]."
modelMode passive https://docs.devin.ai/desktop/models.md "Models are selectable via configuration/UI only (SWE-1.5, SWE-1.6, Adaptive, Arena tiers, Claude, GPT)."
hookBus host https://docs.devin.ai/desktop/cascade/hooks.md "Cascade supports twelve hook events covering critical workflow points … Pre-hooks (can block actions): pre_read_code, pre_write_code, pre_run_command, …" (quote elided beyond the pre-hook enumeration — see #2100 CASCADE FACTS reference)
stateIO filesystem https://docs.devin.ai/desktop/cascade/cascade "Cascade can create and modify codebases directly … File access can be restricted through .codeiumignore files"
transport mcp https://docs.devin.ai/desktop/cascade/mcp "Cascade now natively integrates with MCP, allowing you to bring your own selection of MCP servers for Cascade to use."
runtime undocumented no authoritative doc — searched: https://docs.devin.ai/windsurf/plugins/getting-started.md, /llmstxt/windsurf_llms-full_txt (Context7) —
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch undocumented no authoritative doc — searched: https://docs.devin.ai/cli/subagents.md, https://docs.devin.ai/desktop/agent-command-center.md —
dispatch.nested undocumented no authoritative doc — searched: https://docs.devin.ai/cli/subagents.md —
dispatch.maxDepth undocumented no authoritative doc — searched: https://docs.devin.ai/cli/subagents.md —
dispatch.background undocumented no authoritative doc — searched: https://docs.devin.ai/desktop/acp.md, https://docs.devin.ai/cli/subagents.md —
dispatch.subagentToolkit undocumented no authoritative doc — searched: https://docs.devin.ai/cli/subagents.md —
dispatch.backgroundDispatch undocumented no authoritative doc — https://docs.devin.ai/desktop/cascade/cascade and https://docs.devin.ai/desktop/devin-local (official Windsurf/Devin docs, via docs.windsurf.com redirects) The Windsurf/Cascade docs describe a background planning agent only in these terms: "In the background, a specialized planning agent continuously refines the long-term plan while your selected model f
dispatch.isolation none shipped descriptor (dispatch.backgroundDispatch: undocumented) no documented background/concurrent-dispatch primitive — isolation is moot; same-wave plans run inline (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.namedDispatch — Cascade docs do not document a user-facing named sub-agent dispatch system.
  • dispatch.nested — no documentation for nested sub-agent support in Windsurf Cascade.
  • dispatch.maxDepth — no documented depth limit for Cascade sub-agents.
  • dispatch.background — Cascade has an internal background planning agent but no documented user-facing background sub-agent dispatch.
  • dispatch.subagentToolkit — no documentation for toolkit restrictions on Cascade sub-agents.
  • runtime — Windsurf IDE is Electron-based but no programmatic plugin runtime is documented to developers.

EoS migration status (#2100 Stage 2 — HOOK-BRIDGE): hooksSurface moved from "none" to "windsurf-hooks-json". GSD now wires two of Cascade's documented pre-hooks with BLOCKING semantics via .windsurf/hooks.json (local) / ~/.codeium/windsurf/hooks.json (global): pre_write_code (write-path guard — blocks a write resolving to a different git root than cwd, or into a .git/ internals directory) and pre_run_command (a conservative destructive-command deny-list — whole-disk/home rm -rf, force-push to a protected branch). Cascade blocks via exit code 2 (+ a stderr reason string) — a materially different protocol from Cursor's stdout-JSON {block, reason} hooks.json form, even though the surrounding install/reconcile infra (writeWindsurfHooksJson/removeWindsurfHooksJson in src/runtime-hooks-surface.cts) mirrors writeCursorHooksJson/removeCursorHooksJson's shape. Cascade has no context-injection channel (no additional_context-style advisory response channel), so the 4 advisory hook events GSD registers on Cursor (sessionStart, postToolUse, stop, subagentStart/subagentStop) have no Windsurf/Cascade counterpart and are deliberately not ported — only the 2 events with a genuine blocking analog are wired. installSurface stays profile-marker-only (unchanged); the hook bus is wired from inside that branch, gated on hooksSurface === 'windsurf-hooks-json' rather than a hardcoded runtime check.

trae

Axis Value Source Evidence
embeddingMode imperative https://traeide.com/docs/how-to-manage-extensions-in-trae-ide "Trae IDE is a VSCode fork; 'If an extension isn't available in Trae's store, you can install it from VS Code's marketplace' — inherits VSCode in-process extension model"
commandSurface slash-file https://docs.trae.ai/ide/skills "Skills stored as SKILL.md files in '.trae/skills/{skill_name}/' directory; 'Trae allows you to manually trigger skills if needed'"
modelMode passive https://docs.trae.ai/ide/models "Model selection via UI: 'click on the current model name to open the model list'; no programmatic model/LLM request API documented for plugins"
hookBus engine https://news.ycombinator.com/item?id=44703164 "Trae is 'ByteDance's VSCode fork' built on Electron/Monaco; inherits VSCode extension host lifecycle (activate/deactivate hooks, event subsc"
stateIO filesystem https://traeide.com/news/6 "Rules at '.trae/project_rules.md', skills at '.trae/skills/', MCP config at '.trae/mcp.json'; 'codebase files always remain on your local de"
transport mcp https://docs.trae.ai/ide/model-context-protocol "Page title from official docs: 'In TRAE IDE, MCP servers support three transport types' — MCP is built-in"
runtime node https://news.ycombinator.com/item?id=44703164 "Trae is a VSCode fork built on Electron; 'Electron is designed to create desktop applications… a backend using the Node.js runtime'"
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://docs.trae.ai/ide/agent "Agents in Trae 'can be called individually, or automatically called by SOLO Agent at the corresponding stage'"
dispatch.nested undocumented no authoritative doc — searched: https://docs.trae.ai/ide/solo-mode, https://docs.trae.ai/ide/agent —
dispatch.maxDepth undocumented no authoritative doc — searched: https://docs.trae.ai/ide/solo-mode —
dispatch.background true https://news.aibase.com/news/22829 "SOLO 'supports multi-tasking, allowing you to work on multiple development tasks simultaneously'; 'run multiple agents i"
dispatch.subagentToolkit undocumented no authoritative doc — searched: https://docs.trae.ai/ide/agent —
dispatch.backgroundDispatch undocumented no authoritative doc — https://docs.trae.ai/ide/agent; https://github.com/bytedance/trae-agent/blob/main/docs/roadmap.md Trae's official documentation (docs.trae.ai) and the trae-agent GitHub roadmap do not document background/async agent dispatch or whether a background-spawned agent can itself spawn further sub-agents
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.nested — docs describe two-tier orchestration (SOLO → named agents) but do not state whether a spawned sub-agent can itself spawn further sub-agents.
  • dispatch.maxDepth — no integer depth limit documented beyond one orchestrator level.
  • dispatch.subagentToolkit — docs say agents can be configured with 'callable MCP services and other capabilities' but do not state whether sub-agents receive a full vs. restricted tool set.

EoS migration status (#2094): Migrated onto the imperative adapter — partially. Two runtime === 'trae' string-equality branches folded into descriptor-driven runtime.hostBehaviors: skipSharedHooksInstall:true gates the shared-hooks install (Trae has no hook surface: hooksSurface: "none"), and the case 'trae' global-config-dir path-rewrite's self-alias regex is now built off the descriptor's dirName rather than a hardcoded ~/.trae/ literal (byte-identical output). Skills dispatch was already descriptor-driven before this migration (artifactLayout.skills.converter: "convertClaudeCommandToTraeSkill", resolved by converter name, not a runtime check). Still runtime-keyed (not folded by #2094, matching the same posture as cursor/windsurf/cline, pending a future cross-runtime content-dispatch consolidation): RUNTIME_CONTENT_DISPATCH.trae in bin/install.js — its md/js bodies are regex-callback rewrites that cannot be reduced to a byte-identical descriptor map; and the case 'trae': switch arm itself in src/runtime-artifact-conversion.cts — the arm's structure (not just its self-alias regex) is boilerplate shared verbatim across 7 runtimes (codex, cline, cursor, windsurf, augment, trae, codebuddy) and remains a runtime-keyed switch. trae also remains in RUNTIME_FLAG_IDS (and isTrae remains in bin/install.js, gating only the agents-converter dispatch) pending the cross-runtime agents-converter dispatch migration — agents conversion is out of scope for #2094. One upgrade lands: SOLO stage/trigger metadata — every emitted SKILL.md now carries a stage: workflow frontmatter line (runtime.hostBehaviors.soloStageMetadata), so Trae's SOLO Agent can recognize GSD skills as workflow-stage skills for auto-invocation instead of requiring manual triggering; cite https://docs.trae.ai/ide/agent ("Agents in Trae can be called individually, or automatically called by SOLO Agent at the corresponding stage"). The field is a single fixed, best-effort/inferred GSD-side value — Trae's thin SPA docs don't publish a formal stage-metadata schema. The four undocumented dispatch sub-axes (nested, maxDepth, subagentToolkit, backgroundDispatch) keep dispatch flattened (shouldFlattenDispatch fails closed to inline) — fail-closed negotiation, not a regression.

kimi

EoS migration status (#2095): Two upgrades landed. Upgrade 1 — native hook bus: hooksSurface moved from "none" to "kimi-hooks-toml" (extendedHookEvents: ["SubagentStop", "Stop", "PreCompact", "SubagentStart"], hookEvents: "claude" — Kimi's 13 lifecycle events include exact-name equivalents for every Claude-dialect event GSD wires). GSD's hook scripts (session-state, phase-boundary, graphify, context monitor, the prompt/read/workflow/worktree guards, commit validation) are now registered as [[hooks]] entries in Kimi's own config.toml (default ~/.kimi/config.toml, overridable via Kimi's own KIMI_SHARE_DIR env var — a directory deliberately separate from the ~/.config/agents Agent-Skills root GSD installs into, since Kimi's docs confirm the skills search path is independent of KIMI_SHARE_DIR). GSD-owned entries are wrapped in # GSD Hooks BEGIN/END marker comments (writeKimiHooksToml / stripKimiHooksTomlBlock in src/runtime-hooks-surface.cts) so a reinstall replaces only GSD's own block, and installSurface deliberately stays "profile-marker-only" — the config.toml write is independent of the artifact-install surface. hooks/ and hooks/lib/ now install for kimi (the three && !isKimi install-guard exclusions were removed) — but SELF-CONTAINED under kimi's own native hook root (~/.kimi/, alongside config.toml), never under the ~/.config/agents Agent-Skills root: kimi declares hostBehaviors.skipSharedHooksInstall:true like Cline/Kilo/Cursor/Trae, so the shared install path never writes hooks/package.json there, and a dedicated call installs the same bundle into resolveKimiHooksTomlDir() instead, with buildHookCommand pointed at that root so the generated [[hooks]] command paths resolve. Upgrade 2 — background dispatch: hostIntegration.dispatch.backgroundDispatch flipped false → true (Kimi's Agent tool takes a call-time run_in_background param — same evidence as dispatch.background below), which flips shouldFlattenDispatch to false for kimi (may background, joining codex/cursor/opencode) — a negotiation-only axis with no install-output effect, confirmed via golden parity. Exercising the actual run_in_background call end-to-end is Kimi's own runtime behavior and is out of the installer's test scope; the installer's deliverable stops at the kimi_cli.tools.agent:Agent tool grant on the root agent YAML (buildKimiAgentArtifacts, only emitted when a subagent is present) plus the negotiated backgroundDispatch axis above — both covered by tests/kimi-upgrades.test.cjs. MCP transport deferred: kimi's transport: mcp axis (declared below) is descriptor-only, like every other runtime's — no runtime has installer-driven MCP registration (GSD's installer never invokes kimi mcp add); users register the GSD MCP companion server with Kimi CLI manually.

Axis Value Source Evidence
embeddingMode imperative https://context7.com/moonshotai/kimi-cli/llms.txt "from kimi_cli.app import KimiCLI, enable_logging ... instance = await KimiCLI.create(session, agent_file=myagent) ... class Ls(CallableTool2)"
commandSurface slash-file https://github.com/moonshotai/kimi-cli/blob/main/docs/en/customization/skills.md "/skill:code-style ... /flow:code-review — Skills are SKILL.md markdown files with YAML frontmatter that become /skill: and /flow:"
modelMode passive https://github.com/moonshotai/kimi-cli/blob/main/docs/en/configuration/providers.md "Use the /model command to switch between available models and thinking modes ... --model option overrides the default model"
hookBus host https://moonshotai.github.io/kimi-cli/en/customization/hooks.html "Core: Add hooks system (Beta) — configure [[hooks]] in config.toml to run custom shell commands at 13 lifecycle events including `PreToo"
stateIO filesystem https://github.com/MoonshotAI/kimi-cli "Kimi Code CLI is an AI agent that runs in the terminal ... capable of reading and editing code, executing shell commands, searching files"
transport mcp https://github.com/moonshotai/kimi-cli/blob/main/docs/en/reference/kimi-mcp.md "kimi mcp add ... --transport stdio
runtime python https://context7.com/moonshotai/kimi-cli/llms.txt "from kimi_cli.app import KimiCLI ... from kosong.tooling import CallableTool2 — CLI core is Python"
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://moonshotai.github.io/kimi-cli/en/customization/agents.html "subagents:\n coder:\n path: ./coder-sub.yaml\n description: "Handle coding tasks"\n reviewer:\n path: ./reviewer-sub.yaml"
dispatch.nested false https://moonshotai.github.io/kimi-cli/en/customization/agents.html "All subagent types are prohibited from nesting the Agent tool (subagents cannot create their own subagents). Only root"
dispatch.maxDepth 1 https://moonshotai.github.io/kimi-cli/en/customization/agents.html "All subagent types are prohibited from nesting the Agent tool (subagents cannot create their own subagents). Only root"
dispatch.background true https://moonshotai.github.io/kimi-cli/en/customization/agents.html "Subagents support foreground and background modes. The run_in_background parameter allows tasks to execute asynchronou"
dispatch.subagentToolkit undocumented no authoritative doc — searched: https://moonshotai.github.io/kimi-cli/en/customization/agents.html —
dispatch.backgroundDispatch true (#2095 Upgrade 2; was false) https://moonshotai.github.io/kimi-cli/en/customization/agents.html "Subagents support foreground and background modes. The run_in_background parameter allows tasks to execute asynchronously" (same evidence as dispatch.background above — the root agent's Agent tool call itself takes the run_in_background param)
dispatch.isolation orchestrator-worktree https://github.com/moonshotai/kimi-cli/blob/main/docs/en/faq.md ; /docs/en/customization/agents.md "--work-dir flag sets an explicit working directory"; concurrent "explore" subagents documented — GSD creates + manages the worktree and points the executor at it via --work-dir (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.subagentToolkit — docs show three built-in subagent types each with different tool subsets (coder=full, explore=read-only, plan=no shell/write); no single 'full' or 'read-only' value covers all types; maintainer should clarify the intended classification.
  • runtime — CLI core is Python; a Rust Wire implementation also exists; docs do not state a canonical plugin extension runtime.

kimi-code

kimi-code is a different product from kimi above — every axis below is sourced independently. kimi is Moonshot's Python kimi-cli (from kimi_cli.app import KimiCLI, ~/.kimi/config.toml, --work-dir); kimi-code is the TypeScript/Node Kimi Code CLI (~/.kimi-code/config.toml, $KIMI_CODE_HOME, process-cwd, AgentSwarm). None of the values here are inherited from the kimi section. See docs/migration/kimi-to-kimi-code.md for the user-facing split.

GSD hook destination (#2755): although kimi and kimi-code share hooksSurface: "kimi-hooks-toml", they do not share a hooks root. GSD writes its [[hooks]] block, hook bundle and CommonJS marker into ~/.kimi-code/config.toml for kimi-code (overridable via KIMI_CODE_HOME) and into ~/.kimi/config.toml for kimi (overridable via KIMI_SHARE_DIR); each product's env var is scoped to that product and does not redirect the other. resolveKimiHooksTomlDir({ runtime }) in src/runtime-homes.cts is the single seam that makes this choice, for both the install and uninstall paths. Until #2755 that function was unparameterized and every kimi-code install wrote into Kimi CLI's root, leaving Kimi Code with no hooks — this row is documented explicitly because its absence is what let that go unnoticed.

Axis Value Source Evidence
embeddingMode declarative https://github.com/moonshotai/kimi-code/blob/main/docs/en/customization/plugins.md "Plugins package reusable Kimi Code CLI capabilities into installable units — they can add Agent Skills, automatically load a specified Skill at session start, and declare MCP servers to provide real tool capabilities." A plugin is a kimi.plugin.json manifest plus markdown Skills; real tool capability arrives via external MCP processes. No in-process programmatic extension API is documented — the same shape as codex above.
commandSurface slash-file https://github.com/moonshotai/kimi-code/blob/main/docs/en/reference/slash-commands.md "External skills are automatically registered as slash commands using the skill: namespace prefix. Users can invoke these by typing /skill:<name> followed by optional text, which is then appended to the skill prompt." Skills are SKILL.md markdown files with YAML frontmatter (name, description, type, whenToUse, arguments).
modelMode passive https://github.com/moonshotai/kimi-code/blob/main/apps/kimi-code/src/cli/commands.ts ; /docs/en/guides/getting-started.md "-m, --model <model> — LLM model alias to use for this invocation. Defaults to default_model in config.toml." Model selection is config alias + --model flag + the interactive /model command; no API lets an extension programmatically supply a model.
hookBus host https://github.com/moonshotai/kimi-code/blob/main/docs/en/customization/hooks.md "All hook rules are written in the [[hooks]] array in ~/.kimi-code/config.toml, where each entry is one rule." The CLI fires the lifecycle events (UserPromptSubmit, PreToolUse, PostToolUse, PostToolUseFailure, PermissionRequest, PermissionResult, SessionStart, SessionEnd, SubagentStart, SubagentStop, PreCompact, PostCompact, Stop, Notification, Interrupt); each rule is event + optional matcher + command + optional timeout (1–600s, default 30).
stateIO filesystem https://github.com/MoonshotAI/kimi-code "Kimi Code CLI is an AI coding agent that runs in your terminal — it can read and edit code, run shell commands, search files, fetch web pages, and choose the next step based on the feedback it receives." Full local filesystem; no sandboxed-storage tier is documented.
transport mcp https://github.com/MoonshotAI/kimi-code ; /docs/en/customization/plugins.md "AI-native MCP configuration. Add, edit, and authenticate Model Context Protocol servers conversationally with /mcp-config, without hand-editing JSON." Plugin manifests additionally carry an mcpServers field.
runtime node https://github.com/MoonshotAI/kimi-code "Requirements: Node.js ≥ 24.15.0, pnpm 10.33.0." The CLI is a TypeScript pnpm monorepo (apps/kimi-code, packages/agent-core-v2); hook commands are shell, and MCP servers are external processes.
effortSurface (not declared) https://github.com/moonshotai/kimi-code/blob/main/apps/kimi-code/src/tui/commands/registry.ts Kimi Code documents /effort (alias /thinking) to change reasoning effort, but only as an interactive slash command — there is no --effort-style flag on its invocation, and -m, --model is the only model/effort-adjacent argv. The axis vocabulary is argv
dispatch.namedDispatch false https://github.com/moonshotai/kimi-code/blob/main/docs/en/customization/agents.md ; capabilities/kimi-code/capability.json (artifactLayout — skills only) "The system includes three built-in sub-agents: 'coder' for general software engineering tasks like file modification, 'explore' for read-only codebase navigation and summarization, and 'plan' for architecture design without file or shell access." GSD's kimi-code artifact layout installs Agent Skills only (no agents kind), so no named GSD subagent is ever registered with the host and every GSD role resolves to one of the three built-ins (resolveDispatchType, src/host-integration.cts). This is a GSD-integration-scoped false, not a claim that the host lacks named agents — see Documentation gaps.
dispatch.nested true https://github.com/moonshotai/kimi-code/blob/main/docs/en/customization/agents.md The coder sub-agent "can dispatch its own nested sub-agents when a task decomposes naturally." (Agent files also expose a subagents delegation allowlist, where subagents: [] is what prevents further delegation — nesting is the default.)
dispatch.maxDepth undocumented searched: https://github.com/moonshotai/kimi-code/blob/main/docs/en/customization/agents.md Nesting is documented, but no maximum nesting depth is stated anywhere in the agents or tools reference.
dispatch.background true https://github.com/moonshotai/kimi-code/blob/main/docs/en/reference/tools.md "Agent delegates a subtask to a sub-Agent. Supports foreground execution (waiting for completion) or background execution (returning a task ID). … run_in_background (boolean) - Optional - Defaults to false."
dispatch.subagentToolkit built-in-only https://github.com/moonshotai/kimi-code/blob/main/docs/en/customization/agents.md The three built-ins carry deliberately different tool surfaces — coder "Shares most of the main Agent's toolset; can run shell commands, maintain todo lists, enter Plan mode, and invoke Agent Skills"; explore "Performs read-only operations only and does not modify any files"; plan "Even shell commands are not available". No single full/read-only value covers the set, so the built-in-only member applies (degrades closed to read-only when negotiated).
dispatch.backgroundDispatch true https://github.com/moonshotai/kimi-code/blob/main/docs/en/reference/tools.md ; /docs/en/customization/agents.md The coder built-in "can dispatch its own nested sub-agents" and the Agent tool's run_in_background is a call-time parameter available to it, so a background-dispatched sub-agent may itself dispatch further. AgentSwarm additionally fans out concurrently: "By default the tool ramps up concurrency without an upper limit (5 subagents start immediately, then 1 more every 700 ms); set KIMI_CODE_AGENT_SWARM_MAX_CONCURRENCY to a positive integer to cap how many subagents run at the same time during that ramp, or leave it unset for no cap."
dispatch.isolation orchestrator-worktree https://github.com/moonshotai/kimi-code/blob/main/docs/en/reference/tools.md ; ADR-1239 §Codex-binding amendment Neither Agent nor AgentSwarm exposes a working-directory/cwd parameter — sub-agents inherit the CLI's process cwd — and the CLI has no --work-dir-style flag (unlike kimi). GSD therefore creates, validates and merges the worktree itself and spawns the executor with that cwd (#2584).
dispatch.maxConcurrency undocumented https://github.com/moonshotai/kimi-code/blob/main/docs/en/reference/tools.md AgentSwarm concurrency is configurable ("set KIMI_CODE_AGENT_SWARM_MAX_CONCURRENCY to a positive integer to cap how many subagents run at the same time … or leave it unset for no cap") but the CLI states no bound as its OWN default — "By default the tool ramps up concurrency without an upper limit" — so there is no single default integer to cite, only a user-configurable env var (#3673).

Negotiation note. Because dispatch.namedDispatch is false, negotiateHostCapabilities caps nested, maxDepth, background and backgroundDispatch to false/0 in the effective axes for structural consistency (src/host-integration.cts). The declared values above are still the host-capability record the matrix exists to hold; they are what a future named-dispatch upgrade would negotiate against.

Sources consulted:

Documentation gaps:

  • dispatch.namedDispatch — host capability vs. GSD surface. Kimi Code does support user-authored named agents: "Beyond the three built-in sub-agents, you can define your own agents as Markdown files", discovered across five scopes (--agent-file > .kimi-code/agents/, .agents/agents/ > extra dirs > $KIMI_CODE_HOME/agents/ > built-in). The axis is false because GSD installs no agent files for this host, so its reachable dispatch surface is the three built-ins — flipping the axis without also shipping agent artifacts would reintroduce the dispatch failure recorded in docs/migration/kimi-to-kimi-code.md ("Every workflow that called a named GSD subagent … failed at dispatch"). Shipping GSD agent files to kimi-code is an unbuilt capability upgrade, not a gap in the host's documentation.
  • effortSurface — a real mechanism the vocabulary cannot name. Kimi Code documents /effort (alias /thinking) for changing reasoning effort, but only as an interactive slash command; -m, --model is the only model-adjacent argv on its invocation. The axis vocabulary is argv | none, and neither is accurate here — none denies a mechanism the host has, argv claims one it does not expose. Kimi Code's descriptor was created a day after #2481 added the axis and has carried no value since. Resolving this needs a vocabulary decision (an interactive-only member, mirroring the deliberate absence of a config-file member), which is a negotiation change rather than a documentation one; until then the absent value degrades closed exactly as the sentinel does.
  • dispatch.maxDepth — nesting is documented but no depth bound is published, so the axis carries the undocumented sentinel rather than a guessed integer.
  • runtime — the published artifact is a single bundled binary; Node.js ≥ 24.15.0 is stated as a development requirement. The sources are a TypeScript/Node monorepo, so node is the best-supported classification, but the docs do not name a canonical plugin-execution runtime (the same ambiguity noted for codex above).

zcode

ZCode (Z.ai) is a desktop Agentic Development Environment for the GLM-5.2 model, distributed as an Electron app. It exposes a Claude-Code-shaped extensibility surface (per-user ~/.zcode/skills/<name>/SKILL.md, slash commands, named subagents, native MCP, and a plugin system). All values below are sourced verbatim from the official ZCode docs.

Axis Value Source Evidence
embeddingMode declarative https://zcode.z.ai/en/docs/plugin "A single plugin can bundle several capabilities. ZCode detects which components a plugin includes from its directory layout" — plugins/skills/commands/agents are config/markdown files; no in-process programmatic extension API is documented.
commandSurface slash-file https://zcode.z.ai/en/docs/commands "Custom commands are stored as .md files under ~/.zcode/commands ... invoke the command with /command-name"
modelMode passive https://zcode.z.ai/en/docs/configuration Models are connected by provider config (Z.ai/BigModel/OpenAI-compat/Anthropic-compat base URLs + API keys in Model Settings); no programmatic model request API is documented.
hookBus host https://zcode.z.ai/en/docs/plugin A plugin's bundled components include a "Hook — Automation hooks triggered on specific events" — the host fires the events a plugin subscribes to.
stateIO filesystem https://zcode.z.ai/en/docs/skill "User-level skills for ZCode Agent: ~/.zcode/skills/<skill-name>/SKILL.md" — full local filesystem (desktop app).
transport mcp https://zcode.z.ai/en/docs/mcp-services "MCP (Model Context Protocol) connects external capabilities ... type as stdio (SSE and HTTP remote servers are also supported)" — native MCP.
runtime electron https://zcode.z.ai/en/docs/install (download path cdn-zcode.z.ai/zcode/electron/releases/3.2.5/ZCode-3.2.5-mac-arm64.dmg) ZCode is shipped as an Electron desktop application; the release artifact lives under the electron/releases path.
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://zcode.z.ai/en/docs/subagents "you can let the Agent pick the subagent automatically, or reference it with @ in the chat box" — subagents are invoked by name via the Agent tool.
dispatch.nested undocumented searched: https://zcode.z.ai/en/docs/subagents The docs do not state whether a subagent can itself spawn further subagents.
dispatch.maxDepth undocumented searched: https://zcode.z.ai/en/docs/subagents No maximum nesting depth is documented.
dispatch.background false https://zcode.z.ai/en/docs/subagents "Foreground execution. Subagents run in the foreground ... Background execution is not enabled yet."
dispatch.subagentToolkit full https://zcode.z.ai/en/docs/subagents "general-purpose is the default built-in subagent ... It has access to all tools"; custom subagents default to "All permissions by default" (inherits every tool).
dispatch.backgroundDispatch false https://zcode.z.ai/en/docs/subagents "Background execution is not enabled yet" — background dispatch is therefore impossible.
dispatch.isolation none https://zcode.z.ai/en/docs/subagents (shipped descriptor: dispatch.backgroundDispatch: false) "Background execution is not enabled yet" — no concurrent fan-out primitive, so same-wave plans run inline/sequentially (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.nested / dispatch.maxDepth — ZCode's subagent docs do not state whether subagents can spawn further subagents or any depth bound.
  • configHome — skills/commands/agents homes are documented (~/.zcode/skills, ~/.zcode/commands, ~/.zcode/agents); the exact settings filename under ~/.zcode (where MCP server config is stored) is not fully documented at time of writing.
  • Maintenance note — ZCode is a young, fast-moving app (observed at v3.2.x); these axes may need revision as its on-disk config layout stabilizes. Because ZCode also natively imports skills/MCP from ~/.claude, installing GSD to BOTH claude and zcode can surface duplicated skills inside ZCode; this overlap is expected and documented.

EoS migration status (#2101, ADR-1239): ZCode's install is fully dogfooded through the declarative adapter — its shared-hooks exclusion (previously a hardcoded !isZcode branch in bin/install.js) is now folded onto hostBehaviors.skipSharedHooksInstall, byte-parity with the prior install (ZCode's golden install tree has zero hook files). The two capability upgrades anticipated for ZCode both remain blocked on undocumented on-disk formats — hookBus and transport above stay documented-but-unimplemented pending ZCode publishing those formats, and implementing a guessed format risks a false-green descriptor, so neither upgrade is wired:

  • Hook automation (the plugin Hook component, hookBus: host above) — https://zcode.z.ai/en/docs/plugin documents the capability only at a high level ("Automation hooks triggered on specific events"; components are "detected from directory layout, shown as badges"). No config file format, on-disk location, event-name vocabulary, or payload schema is published, so GSD cannot faithfully wire hook events into a plugin bundle. BLOCKED (undocumented on-disk hook-config format).
  • MCP registration (transport: mcp above) — https://zcode.z.ai/en/docs/mcp-services confirms servers are "stored in the .zcode configuration file of the chosen scope" and accepts both a bare {"server-name":{...}} map and an {"mcpServers":{...}} wrapper shape, but does not document the exact settings filename/path or full schema (the docs describe the UI flow, not the on-disk contract) — this is the same gap already noted under configHome above. BLOCKED (undocumented settings-filename/schema gap).

pi

pi (pi.dev) is a bun-runtime Programmatic-CLI: it exposes an in-process TypeScript ExtensionAPI (registerCommand/registerTool/registerProvider/pi.on) rather than a settings-file or slash-markdown surface. GSD ships a single native-extension file (pi/gsd.cjs) installed to ~/.pi/agent/extensions/gsd.js (global) or .pi/extensions/gsd.js (local) — the programmatic-CLI peer of the OpenCode/Kilo native-plugin binding. Sourcing note: the citations below are the pi.dev documentation pages named in ADR-1239 Stage 1 (#2102) as the source for each axis; this environment did not have live doc-fetch access at authoring time, so the Evidence column below is a paraphrase of pi's documented extension model rather than a verbatim excerpt — a maintainer with Context7/web access should verify the exact wording before treating this section as fully cited (flagged in the #2102 PR). Partially discharged (#2470, 2026-07-20): pi's extension-loader contract specifically has now been read at source — packages/coding-agent/src/core/extensions/loader.ts in earendil-works/pi — confirming discoverExtensionsInDir() keeps only names passing isExtensionFile() (.ts/.js, everything else skipped silently), that accepted files load via jiti (CommonJS and ESM alike), and that explicit settings.json paths bypass the filter. The remaining axes below are still paraphrase.

Axis Value Source Evidence
embeddingMode imperative https://pi.dev/docs/latest/extensions pi extensions are loaded in-process (via jiti) and call an ExtensionAPI object directly (registerCommand/registerTool/registerProvider/pi.on) — an in-process programmatic API, not a config-file-only integration.
commandSurface slash-programmatic https://pi.dev/docs/latest/extensions Commands are registered by calling registerCommand(name, definition) from extension code, not by dropping a markdown/TOML file — the command surface is code, not a file format.
modelMode active https://pi.dev/docs/latest/extensions The ExtensionAPI exposes registerProvider, letting an extension supply/select model providers programmatically rather than only reading a static config value.
hookBus host https://pi.dev/docs/latest/extensions pi.on(event, handler) subscribes an extension to host-fired lifecycle events (e.g. tool_call) — the pi host owns and fires the event bus; extensions only subscribe.
stateIO session-log-append https://pi.dev/docs/latest/session-format pi persists conversation/tool-call state as an append-only session log/transcript format rather than exposing unrestricted local filesystem access to extensions.
transport native-extension https://pi.dev/docs/latest/extensions Integration is a single loaded extension file (~/.pi/agent/extensions/<file>.cjs), not an MCP server process — the peer mechanism to OpenCode's native plugins/*.js adapter.
runtime bun https://pi.dev pi is distributed and executed as a bun-runtime CLI (its extensions are loaded via jiti under bun, not Node.js or Python).
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch undocumented no authoritative doc — searched: https://pi.dev/docs/latest/extensions The ExtensionAPI documents registerCommand/registerTool/registerProvider/pi.on; it does not document a named-subagent-invocation primitive.
dispatch.nested undocumented no authoritative doc — searched: https://pi.dev/docs/latest/extensions No documented subagent-of-subagent nesting capability.
dispatch.maxDepth 0 no authoritative doc — searched: https://pi.dev/docs/latest/extensions No named-dispatch primitive is documented at all (see dispatch.namedDispatch), so there is no nesting depth to bound; 0 records "no dispatch levels beyond the root extension," not a measured limit.
dispatch.background false no authoritative doc — searched: https://pi.dev/docs/latest/extensions No documented background/async subagent-execution primitive.
dispatch.subagentToolkit undocumented no authoritative doc — searched: https://pi.dev/docs/latest/extensions pi has no named-dispatch primitive (see dispatch.namedDispatch), so there is no subagent tool-surface to classify as full/read-only.
dispatch.backgroundDispatch false no authoritative doc — searched: https://pi.dev/docs/latest/extensions Same gap as dispatch.background — no background-dispatch primitive is documented, so a background-dispatched agent spawning further named sub-agents is not possible.
dispatch.isolation none shipped descriptor (dispatch.background: false, dispatch.backgroundDispatch: false) pi has no named-dispatch/background-dispatch primitive documented — cannot fan out concurrently, so isolation is moot (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.namedDispatch / dispatch.nested / dispatch.subagentToolkit — pi's ExtensionAPI (registerCommand/registerTool/registerProvider/pi.on) does not document a named-subagent-dispatch primitive at all, unlike Claude Code/Codex/OpenCode-style "Agent tool" surfaces; all three axes stay undocumented and negotiation fails closed (no named dispatch, dispatch flattened).
  • dispatch.maxDepth / dispatch.background / dispatch.backgroundDispatch — recorded as 0/false/false (not undocumented) because the absence of any dispatch primitive is itself the documented ceiling, matching shouldFlattenDispatch's fail-closed default.
  • This section's Evidence-column wording was authored without live Context7/web-fetch access (see the sourcing note above the table) — verify against the cited pi.dev pages before relying on it for a future capability upgrade.

EoS migration status (#2102 Stage 1, ADR-1239): pi lands as a NET-NEW installable runtime — pure additive descriptor + installer wiring, no prior runtime === 'pi' branches existed to fold. artifactLayout is declared empty (global: [], local: []) — pi has no skills/commands/agents layout, and installs as PLUGIN-ONLY: hostBehaviors.pluginOnlyInstall: true explicitly skips bin/install.js's generic flat-commands-and-agents fallback (the legacy path Claude Code's LOCAL layout also uses), which would otherwise write inert commands/gsd-<cmd>.md + agents/gsd-<name>.md reference files no part of pi ever reads. pi's /gsd command and gsd_invoke tool are registered programmatically by the native extension (pi/gsd.cjs → extensions/gsd.js, mirroring OpenCode/Kilo's nativePlugin shape) — pi has no host-read markdown surface at all (unlike Claude/OpenCode/Kilo, which scan a commands//command/ directory), so a declarative artifact surface would be dead weight, not merely unused. dispatch.subagentToolkit: "undocumented" and dispatch.backgroundDispatch: false are both required by the capability validator's dispatch schema and reflect that pi has no documented named-dispatch primitive at all. (Stage 1 originally also set hostBehaviors.skipSharedHooksInstall:true, reasoning the staged hooks/*.js bundle would be dead weight for pi the way it genuinely is for Kilo/ZCode — corrected in Stage 2 below: pi's native extension DOES spawn them, so they are live, not dead, and the flag was removed.)

EoS migration status (#2102 Stage 2, ADR-1239): Stage 1's "in-process gsd-core command-routing hub" framing was aspirational and is corrected here — no fully-populated hub factory exists anywhere in gsd-core (every createHub() caller in the tree builds a single-family hub for its own narrow purpose), so /gsd and gsd_invoke instead dispatch via SUBPROCESS REUSE: dispatchGsdCommand (src/shell-command-projection.cts) spawns gsd-core/bin/gsd-tools.cjs <family> [subcommand] ... --cwd <dir> --raw --json-errors bounded and non-throwing, mirroring the precedent already established for the OpenCode/Kilo hook bridge (.opencode/plugins/gsd-core.js's "Architecture: SUBPROCESS REUSE" header). The companion MCP server's gsd_invoke_command tool dispatches through the SAME shared helper (it had the identical createHub()-with-no-args bug). /gsd's command handler is handler(args, ctx) (pi's real ExtensionAPI shape — a raw args string, not execute(ctx)); gsd_invoke's tool handler is the real 5-arg execute(toolCallId, params, signal, onUpdate, ctx). The event surface (EXTENSION_EVENT_SURFACES.pi, src/host-integration.cts) now declares the full ~30-event pi ExtensionAPI vocabulary (was a placeholder ['tool_call']), and pi/gsd.cjs binds session_start (→ gsd-ensure-canonical-path.js), before_agent_start (→ gsd-workflow-guard.js, a forward-compatible no-op today since that hook's triggers are tool-scoped), session_before_compact (→ gsd-context-monitor.js), and tool_call, each as a bounded fail-open spawnSync subprocess (mirroring .opencode/plugins/gsd-core.js's runHook). modelMode: active is realized via pi.on('before_provider_request', ...), which resolves a tier through the model-catalog's now-populated runtimeTierDefaults.pi entries (bare anthropic ids — claude-opus-4-8/claude-sonnet-5/claude-haiku-4-5, matching the claude runtime's own ids since pi talks the anthropic API) and returns a modified payload, or undefined (fail-open, pi's model left untouched) when resolution comes back null — not registerProvider, which would register a new model provider rather than steering pi's existing built-in anthropic models.

Adversarial-review correction (#2102 Stage 2, post-review): the event bridges above and the /gsd tokenizer's hooks/lib/git-cmd.js require were DEAD in a real install — Stage 1's hostBehaviors.skipSharedHooksInstall:true meant pi shipped NO hooks/ directory at all, so runHook('gsd-ensure-canonical-path.js', ...) etc. always hit the "hook file absent → silent no-op" branch, and the tokenizer always fell back to plain whitespace-splitting. The tests masked this because they run against the dev tree, where hooks/ genuinely exists. Fix: capabilities/pi/capability.json no longer sets skipSharedHooksInstall — pi is architecturally identical to OpenCode here (hooksSurface: "none" + a native extension that spawns the staged hooks), not to Kilo/ZCode (hooksSurface: "none" with NO plugin surface, where the same hooks genuinely are dead weight). pi now installs hooks/ + hooks/lib/ (27 entries: the same INSTALLED_HOOK_FILES set OpenCode gets) alongside extensions/gsd.js, verified end-to-end via a real node bin/install.js --pi --global/--local — resolveEngineRoot's walk-up from the installed extension's own directory finds ENGINE_ROOT/hooks/{gsd-ensure-canonical-path.js,gsd-workflow-guard.js,gsd-context-monitor.js,lib/git-cmd.js}, and each bridge/runHook call exits 0 against the real installed files. hooksSurface: "none" + configFormat: "none" + writesSharedSettings: false are unaffected — no settings/hooks.json/config.toml is written for pi; the extension spawns hooks by absolute path, not via a config-file hook bus. tests/fixtures/golden-install-parity/pi.json grew from 292 → 320 entries (the 28 new hooks//hooks/lib/ files); commands/, agents/, skills/ remain absent (pluginOnlyInstall is untouched — it only gates the declarative-markdown surfaces, not hooks). tests/install-minimal-hooks.test.cjs's #1821 suite moved pi from the Kilo/ZCode (no-hooks) group into the OpenCode (ships-hooks) group accordingly.

Superseded in part (#3023, 2026-08-07): the bundle's directory NAME became runtime-descriptor-driven (hostBehaviors.sharedHooksDirName, default hooks); pi sets it to gsd-hooks because pi reserves hooks/ as its deprecated extension directory and checkDeprecatedExtensionDirs() in packages/coding-agent/src/migrations.ts warns on bare directory existence (no emptiness check, unlike its tools/ sibling — source read 2026-08-07); the set of staged files and every other negotiated axis is UNCHANGED (hooksSurface: "none", configFormat: "none", writesSharedSettings: false, pluginOnlyInstall all untouched — only the directory name moved); pi/gsd.cjs now resolves the bundle by probing gsd-hooks then hooks so dev checkouts and half-upgraded trees still work.

vscode

VS Code is the IDE-profile reference host: a Marketplace/VSIX-distributed extension, NOT file-projected onto a config directory — it has no runtime.localConfigDir in the usual sense (configHome.kind: "none", localConfigDir: null) and no CLI install surface at all (installSurface: "none"; it is never installed by bin/install.js — no --vscode flag, no allRuntimes membership; see capabilities/vscode/capability.json). The extension IS the host. Sourcing note: the citations below are the VS Code extension API documentation pages named in ADR-1239 (#2103) as the source for each axis; this environment did not have live Context7/ web-fetch access at authoring time, so the Evidence column is a paraphrase of VS Code's documented extension model rather than a verbatim excerpt — a maintainer with Context7/web access should verify the exact wording before treating this section as fully cited (same caveat already flagged for the pi section above).

Axis Value Source Evidence
embeddingMode imperative https://code.visualstudio.com/api/references/vscode-api The extension is loaded in-process by the extension host and calls the vscode namespace API directly (vscode.commands.registerCommand, vscode.chat.createChatParticipant, vscode.lm.registerTool) — an in-process programmatic API, not a config-file-only integration.
commandSurface palette https://code.visualstudio.com/api/extension-guides/command Commands are contributed via contributes.commands in package.json and registered with vscode.commands.registerCommand, surfaced through the Command Palette (and the Chat view via the chat participant) — not a markdown/TOML slash-command file format.
modelMode active https://code.visualstudio.com/api/extension-guides/ai/language-model The vscode.lm namespace lets an extension actively select a model (vscode.lm.selectChatModels) and send requests to it programmatically, rather than only reading a static config value.
hookBus engine https://code.visualstudio.com/api/references/activation-events VS Code has no cross-extension lifecycle-hook bus that GSD subscribes to; the extension host (the "engine" here, per this axis's own host/engine/none vocabulary) owns activation events, and GSD's own hook lifecycle runs fully in-process/engine-owned inside the extension.
stateIO sandboxed-storage https://code.visualstudio.com/api/references/vscode-api#Memento context.globalState/context.workspaceState (both Memento) are the extension's persistent storage surface — sandboxed key/value storage scoped to the extension, not unrestricted local filesystem access.
transport mcp https://code.visualstudio.com/api/extension-guides/ai/mcp VS Code 1.99 added native MCP client support; on the Web (webworker) entry, full GSD command dispatch is available through VS Code's native MCP client connecting to the GSD companion MCP server (gsd-mcp-server), not an in-process Node dispatch (which the web entry cannot run at all).
runtime sandboxed-web https://code.visualstudio.com/api/extension-guides/web-extensions The browser entry point (vscode/browser.js) runs in a webworker context with no Node core modules — the Web Extension execution model VS Code documents for extensions that must run in vscode.dev/github.dev.
effortSurface undocumented no authoritative doc — per-host reasoning-effort survey, #2481 (09b535ac0) This host's documentation states no reasoning-effort setting, so the axis carries the fail-closed sentinel rather than inheriting a profile baseline. No host-specific URL is cited because the finding is an ABSENCE: #2481 surveyed all hosts for a reasoning-effort mechanism and found one only for claude/opencode/codex.
dispatch.namedDispatch true https://code.visualstudio.com/docs/copilot/chat/chat-agent-mode#_agent-mode-tools Registered languageModelTools (and the chat participant) are addressable by name — the primary agent references a tool/participant by its declared name/toolReferenceName, not only positionally.
dispatch.nested true https://code.visualstudio.com/docs/copilot/copilot-chat-agents (subagents) VS Code's chat subagent model (#runSubagent) explicitly supports a subagent invoking further subagents, gated by chat.subagents.allowInvocationsFromSubagents.
dispatch.maxDepth 5 https://code.visualstudio.com/docs/copilot/copilot-chat-agents (subagents) Documented as VS Code's maximum nesting depth for #runSubagent chains — also matches this repo's existing PROFILE_BASELINES.ide.dispatch.maxDepth baseline.
dispatch.background true https://code.visualstudio.com/api/extension-guides/ai/tools Language Model Tools can be invoked as part of an asynchronous agent turn (the primary agent does not block synchronously on a single extension call).
dispatch.subagentToolkit undocumented no authoritative doc found at authoring time VS Code's subagent documentation does not state whether a subagent's tool surface is restricted to read-only tools or the full set an extension registers; recorded undocumented (fails closed to read-only in negotiation) rather than guessed.
dispatch.backgroundDispatch undocumented no authoritative doc found at authoring time Whether a background-dispatched subagent can itself spawn further NAMED subagents (the #853 discriminator) is not stated in the sources reviewed; recorded undocumented (fails closed to false) rather than guessed.
dispatch.isolation undocumented not researched / no concurrent fan-out documented for this axis no authoritative source consulted for concurrent-executor isolation on this host — fails closed to none (sequential) in negotiation (#2584)
dispatch.maxConcurrency undocumented not researched / no documented concurrent-subagent capacity for this axis no authoritative source consulted for concurrent-executor capacity on this host — fails closed to 1 in negotiation (#3673)

Sources consulted:

Documentation gaps:

  • dispatch.subagentToolkit / dispatch.backgroundDispatch — the reviewed sources document that #runSubagent exists (v1.105+, chat.subagents.allowInvocationsFromSubagents, max nesting depth 5) but do not state the subagent tool-restriction model or whether a background-dispatched subagent can itself spawn further named subagents; both stay undocumented and negotiation fails closed.
  • This section's Evidence-column wording was authored without live Context7/web-fetch access (see the sourcing note above the table) — verify against the cited pages before relying on it for a future capability upgrade, same caveat as the pi section above.

EoS migration status (#2103): vscode lands as a registry runtime (role:runtime) for validator/host-integration coverage ONLY — it is deliberately NOT a CLI-installable runtime (installSurface: "none", never in bin/install.js's allRuntimes; see the NON_INSTALLABLE_RUNTIMES carve-out in tests/runtime-flags.test.cjs). The extension surface (vscode/extension.js, vscode/browser.js, vscode/host-binding.js, vscode/package.json) is distributed via the Marketplace/VSIX, not npx --vscode — there is no docs/how-to/install-on- your-runtime.md entry for it. Dispatch is SUBPROCESS REUSE on desktop (the same shared dispatchGsdCommand in gsd-core/bin/lib/shell-command-projection.cjs the pi extension and the companion MCP server use) via vscode/extension.js's main entry (Node). The browser entry (vscode/browser.js) is a SEPARATE, independently zero-Node-API file: it does NOT require host-binding.js because that module's engine-lib dependencies (state-io.cjs, adapter-imperative.cjs → install-engine.cjs/capability-loader.cjs, model-adapter.cjs → model-resolver.cjs → config-loader.cjs/configuration.cjs) all pull in Node's fs/os/path at module-load time — requiring any of them from a webworker context would throw immediately. browser.js instead composes its own minimal surface directly against vscode.lm, and its command/tool/chat handlers surface an honest "full dispatch is unavailable on web; configure the GSD MCP server" message rather than a silent failure. The chat participant (@gsd) and Language Model Tools (a representative 3-tool set — gsd_progress, gsd_workstreams, gsd_plan_phase — matching real shipped skills that map onto a single, safe, read-only gsd-tools.cjs command) are registered on BOTH entries identically; only the dispatch behavior differs. #runSubagent wiring (registerSubagentDispatch/dispatchAsSubagent, gated on chat.subagents.allowInvocationsFromSubagents availability, fail-soft on older/Insiders-gated hosts) adds a belt-and-suspenders maxDepth: 5 ceiling independent of whatever VS Code's own chat engine enforces natively — there is no separate extension-side "subagent contribution" registration API beyond the chat participant + Language Model Tools already registered; VS Code's chat engine surfaces them to #runSubagent on its own.

Reference in New Issue View Git Blame Copy Permalink