CrewRig

CLI Support Matrix

Reference of every CLI-specific integration point in CrewRig. Update this file in the same PR as any change that adds, removes, or modifies a CLI-specific feature — drift here is a parity bug.

Supported CLIs

CLI Config root Entry point file Plugin install dir Notes
Claude Code .claude/ CLAUDE.md (re-exports AGENTS.md) CLI-managed marketplace Plugin-based loading. Components must be declared in a marketplace and installed via the CLI.
Gemini CLI .gemini/ .gemini/GEMINI.md (extension form) / GEMINI.md ~/.gemini/extensions/ Extension-based loading. Settings live in ~/.gemini/settings.json.
Copilot CLI Split: .github/copilot/ (settings + hooks) + .github/skills/ (skills) + .github/agents/ (agents) .github/copilot-instructions.md (re-exports AGENTS.md) In-place via .github/ Open Agent Skills standard. No single config root by design — Copilot reads each sibling path independently. @copilot mentions on issues / PRs trigger the coding agent natively.
Antigravity CLI ~/.gemini/antigravity-cli/ .agents/ANTIGRAVITY.md ~/.gemini/config/plugins/ Plugin-based loading via agy plugin install. Workspace entry point is .agents/ANTIGRAVITY.md; user-level config lives under ~/.gemini/antigravity-cli/. (spec 0052)

Feature matrix

One row per integration point. ✅ = present, ❌ = absent, note when relevant.

# Feature / integration point Source of truth Claude Code Gemini CLI Copilot CLI Antigravity CLI
1 Workspace config root directory repo layout ✅ .claude/ ✅ .gemini/ ✅ Split: .github/copilot/ + .github/skills/ + .github/agents/ ✅ .agents/ (spec 0052 / PR #439)
2 Top-level agent-context entry point repo layout ✅ CLAUDE.md ✅ GEMINI.md (in extensions; absent at repo root) ✅ .github/copilot-instructions.md ✅ .agents/ANTIGRAVITY.md (spec 0052)
3 Skill definitions directory (built output) — the build is tier-agnostic (globs artifacts/*/); output is routed by tier (ADR-0011, spec 0019): core → committed project tree (paths shown); every non-core tier (library, community, org) → gitignored staging dist/<tier>/.<cli>/skills/<name>/SKILL.md. The init-expertise and init-team guided skills (spec 0021) are core artifact sources and build to all four CLI skill surfaces through this path — but they are not the only guided bootstrap components at the core tier: init-personal-profile and init-soul (spec 0191) are also core bootstrap sources reaching all four CLI surfaces, through the command kind (row 5) rather than skill, since only the command kind gives Gemini CLI a native slash-command output. Relative Markdown link ascent depths are adjusted symmetrically across all targets (issue #872). scripts/build-components.sh build_skills + output_root_for_tier ✅ core: .claude/skills/<name>/SKILL.md · non-core: dist/<tier>/.claude/skills/<name>/SKILL.md ✅ core: .gemini/skills/<name>/SKILL.md · non-core: dist/<tier>/.gemini/skills/<name>/SKILL.md ✅ core: .github/skills/<name>/SKILL.md · non-core: dist/<tier>/.github/skills/<name>/SKILL.md ✅ scripts/build-components.sh --target antigravity (spec 0053 / PR #441); core: .agents/skills/<name>/SKILL.md · non-core: dist/<tier>/.agents/skills/<name>/SKILL.md
4 Agent definitions directory (built output) — same tier routing as row 3 (core → project tree; non-core → dist/<tier>/) scripts/build-components.sh build_agents + output_root_for_tier ✅ .claude/agents/<name>.md (flat file; spec 0201) — the Claude Code subagents page documents .claude/agents/ and ~/.claude/agents/ as the two scopes and a single flat file per agent, with no nested per-agent directory form documented anywhere on it; only the page's separately-documented recursive-scan behavior made the retired nested form work incidentally. Unchanged by spec 0201, and recorded here (requirements 24, 26): the Antigravity cell of this row stays nested — its shipped customization documentation names no agent customization type at all, and requirement 3 of specs/0053-antigravity-build-pipeline.md normatively pins the nested Antigravity path; and the Claude Code plugin packaging of extension agents (row 5c) is unchanged — deciding whether to flatten that packaged copy is a named follow-up, not settled here. ✅ .gemini/agents/<name>.md (flat file) ✅ .github/agents/<name>.md (flat file) [GAP-confirmation] — repo-level layout not in the public reference ✅ .agents/agents/<name>/AGENT.md (directory; spec 0053). Corrected in spec 0119: this cell previously read .antigravity/agents/<name>.md (flat file), which contradicted both build_agents and row 1's .agents/ root. The spec 0119 collision key derives from these paths, so the contradiction would have made that guard unauditable against its own reference table. Since spec 0201 flattened the Claude cell of this row, this Antigravity path no longer mirrors it — both remain per-agent directories on their own terms, but the parallelism this cell once described no longer holds.
4b Agent provenance carrier — upstream-tier feedback resolves to canonical regardless of a fork's feedback_repo (spec 0030); the carrier mechanism varies by CLI constraint (frontmatter YAML for Claude, Copilot, and Antigravity; body HTML comment for Gemini), but the resolved value shifts symmetrically scripts/build-components.sh build_agents (inject_provenance vs gemini_provenance_comment) ✅ YAML metadata.provenance block inside frontmatter (Claude accepts arbitrary frontmatter keys) ✅ HTML comment <!-- crewrig-provenance: version="…" canonical="…" feedback="…" --> as first line of body — Gemini CLI 0.42.0 rejected this repository's own type: / metadata: keys (issue #54); a live re-probe of Gemini CLI 0.46.0 (issue #1116, spec 0198 R22) confirmed that narrower rejection still holds while model: / temperature: / max_turns: are accepted (bundled subagents.md), so metadata.provenance still travels as a body comment (build intentionally bypasses inject_provenance — see the NOTE: no $source arg comment in build_agents's Gemini branch) ⚠️ YAML metadata.provenance block inside frontmatter (parallel with Claude — [GAP-confirmation] — public Copilot reference does not document tolerated frontmatter keys; verify and downgrade to HTML comment if rejected) ✅ YAML metadata.provenance block inside frontmatter (parallel with Claude and Copilot; spec 0053 R5)
4c Agent tool mapping / capabilities — maps tool requirements/capabilities from source frontmatter to output agent definitions scripts/build-components.sh build_agents ❌ N/A (Claude Code does not use agent-level allowed-tools) ❌ N/A (Gemini CLI has no agent-level tool requirements) ❌ N/A (Copilot CLI has no agent-level tool requirements) ✅ Mapped: maps Bash from claude.allowed-tools to enable_write_tools: true and supports explicit antigravity configuration block (enable_write_tools, enable_mcp_tools, enable_subagent_tools) (spec 0137)
5 Slash-command directory (built output) — init-personal-profile and init-soul (spec 0191) are core command sources compiled through this path, reaching all four CLI surfaces shown here. On Gemini CLI specifically, they are reachable only as this native slash command (spec 0191.delta-01 R18): the .agents/skills/<name>/SKILL.md output this same source also compiles (row 3) is a Gemini-visible workspace skill by the .agents/skills/ alias (row 3's mechanism), which would otherwise let Gemini's SkillCommandLoader mint a second, competing /<name> reachable through the activate_skill tool — row 7's .gemini/settings.json deliberately disables that path for exactly these two names, so there is no model-side activation route to either interview on Gemini CLI, by design, not by omission. scripts/build-components.sh build_commands ✅ Compiled into .claude/skills/<name>/SKILL.md (commands are wrapped as user-invocable skills) ✅ .gemini/commands/<name>.toml (native TOML) — the sole Gemini-reachable form for these two names ✅ Compiled into .github/skills/<name>/SKILL.md (no first-class slash-command file format — [GAP]) ✅ Compiled into .agents/skills/<name>/SKILL.md (same skill-wrapper pattern as Claude and Copilot; spec 0053). Invocation caveat, measured during the issue #1077 probe: Antigravity CLI resolves a scripted or headless invocation (agy -p/-i) against its own last-active project state, not the shell's cwd — on a machine that already has other Antigravity projects, agy -i "/init-personal-profile" run from a fresh crewrig clone resolved against an unrelated, previously-active project's config/PROFILE.md instead. Passing --new-project anchors the session to the current directory and was required for every probe run in that logbook; a first clone with no prior Antigravity project history is unaffected, but the flag is safe and recommended unconditionally.
5b Extension command render (spec 0042) — an extension's command pivot source commands/<name>.md is rendered per-CLI by how that CLI loads the extension (row 13), NOT by build_components (spec 0024 delta-01 forbids it from walking extensions/). Both paths share scripts/lib/render-command.sh. Carrier is format-specific (R5): the command→TOML path carries provenance as a TOML # crewrig-provenance: … comment line (a bare HTML comment between TOML keys is a TOMLDecodeError); the skill/agent→Markdown paths keep the HTML-comment carrier of row 4b. scripts/build-extension.sh (Gemini, build-tree .toml), scripts/build-claude-plugin.sh (Claude, ephemeral plugin skill), scripts/lib/render-command.sh (shared) ✅ Ephemeral: build-claude-plugin.sh renders the pivot commands/<name>.md body into dist-claude-plugin/<ext>/skills/<name>/SKILL.md at install time (it no longer reads the Gemini .toml). The extension.json generic commands.convertToSkills flag (the retired components.commands.enabled wrapper is gone — spec 0183 R1; this was already stale before this change, since the generic top-level commands section had already replaced it) means "render pivot .md → Claude skill", not "convert .toml → skill" (rename deferred to spec 0044). ✅ Ephemeral, build tree: scripts/build-extension.sh --target gemini renders commands/<name>.md → build/extensions/<ext>/commands/<name>.toml, because Gemini loads an extension installed from that build tree with no install-time render hook of its own (row 13, spec 0173 delta-01). No .toml may be committed in the source tree at all; bash scripts/build-extension.sh --check fails the build the moment one is. ✅ Ephemeral: build-copilot-plugin.sh renders pivot commands/<name>.md body into dist-copilot-plugin/<ext>/skills/<name>/SKILL.md when convertToSkills is true (same pattern as Claude and Antigravity; spec 0065) ✅ Ephemeral: build-antigravity-extension.sh renders pivot commands/<name>.md body into dist-antigravity-plugin/<ext>/skills/<name>/SKILL.md when convertToSkills is true (same pattern as Claude and Copilot; spec 0063). Correction (issue #1007): this cell previously read "❌ No extension command pivot — [GAP] by design" and named a phantom scripts/build-antigravity-plugin.sh (no such file exists in the tree; the real builder is build-antigravity-extension.sh, row 13). That claim directly contradicted row 13's own description of the same builder and is corrected here rather than left standing; the matching GAP-inventory entry below is removed in the same change.
5c Extension skill/agent provenance carrier (spec 0043) — an extension's skills/<name>/SKILL.md / agents/<name>/AGENT.md is consumed in place with no render seam: install-extension.sh does ln -s / cp -rf of the whole extension dir, link-extensions.sh calls it, and build-claude-plugin.sh does cp -r of the skill dir — so all CLIs read the same source bytes. Agents are the one exception to whole-directory copying: since an agents/<name>/ dir may hold sibling pivot sources for other CLIs (e.g. PROMPT.md for Gemini), build-claude-plugin.sh copies only the files matched by the fixed default glob (agents/*/AGENT.md) — the per-extension override key claude.agents is retired (spec 0183 R7): every committed manifest already resolved to this same default, so removing the override changes nothing observable — file by file — cp -r of the whole agent dir was leaking those siblings into the Claude plugin as bogus extra agents (issue #600). The copied AGENT.md bytes themselves are unchanged, so the "same source bytes" and provenance-carrier reasoning below still hold for the file that is copied. Unlike row 4b (artifacts/ agents keep metadata.provenance in frontmatter in the source and the build strips it → comment for Gemini only), an extension skill/agent has nowhere to strip: there is no .md/.toml filename split (row 5b is commands, not skills) to host a Gemini-safe sibling. So the provenance is authored directly as the HTML-comment carrier on the first body line — Gemini-safe as written. canonical/feedback are literal repo URLs (NOT ${CANONICAL_REPO}): a consuming project's config cannot resolve a third-party extension's origin (spec 0043 R2). R5 enforcement note: check-feedback-routing.sh (spec 0030) reads metadata.provenance from frontmatter via yq, so it cannot see the comment carrier and permanently skips extension skills — R5 (feedback==canonical) for extension skills/agents lives solely in check-extension-provenance.sh; there is no defense-in-depth overlap with the 0030 guard. scripts/check-extension-provenance.sh (presence + R1/R5 routing guard, parses the comment carrier, NOT yq); carrier producer gemini_provenance_comment (row 4b) ✅ HTML comment <!-- crewrig-provenance: version="…" canonical="…" feedback="…" --> as first body line of the in-place SKILL.md (Claude tolerates the inert body comment; greeter has no prior frontmatter provenance, so this is not a downgrade) ✅ Same source bytes — the first-body-line HTML comment is Gemini-safe as authored (no non-name/description frontmatter key) ✅ Same in-place source bytes — build-copilot-plugin.sh copies skills/<name>/SKILL.md verbatim; the first-body-line HTML-comment carrier (spec 0043) is preserved unchanged (spec 0065) ❌ No extension skill/agent loader — [GAP] by design. Antigravity CLI uses agy plugin install to load plugins; no in-repo extension skill/agent loader path exists. Evidence: scripts/install-antigravity-extension.sh builds a plugin package and calls agy plugin install — the plugin is consumed by the CLI as a whole package, not via a per-skill directory link
5d Extension context render (spec 0181, issue #1007) — symmetric with row 5b: an extension's agent-facing context source CONTEXT.md (named by the generic context.source key) is rendered per-CLI by how that CLI loads the extension, through the shared render-variable resolver scripts/lib/render-context.sh, which also reads its per-target constants (display name, commandRef/skillRef templates, contextOutput location) from scripts/lib/extension-targets.json — the same descriptor shared with rows 5b's sibling hook/MCP translators. No per-CLI context key remains admissible in extension.json (gemini.contextFileName, claude.contextFileName, antigravity.contextFileName all retired); the built filename is the render's own knowledge on every target. scripts/lib/render-context.sh (shared resolver), scripts/build-extension.sh (Gemini), scripts/build-claude-plugin.sh, scripts/build-copilot-plugin.sh, scripts/build-antigravity-extension.sh ✅ Rendered into the built plugin root: dist-claude-plugin/<ext>/CLAUDE.md ✅ Rendered into the build tree: build/extensions/<ext>/GEMINI.md, sibling to gemini-extension.json, which still names it via contextFileName (render-supplied, not authored) ✅ Rendered as a user-invocable skill — the same surface row 5b's command render already uses for this target — dist-copilot-plugin/<ext>/skills/<ext>-context/SKILL.md. Pinned by live evidence (tests/extension-context-delivery-evidence.md): the CLI's plugin surface carries no context/instructions concept of its own (copilot plugin --help), confirmed rather than assumed before this delivery was declared mapped ✅ Rendered as a plugin rule file — dist-antigravity-plugin/<ext>/rules/AGENTS.md, pinned by the same live-evidence probe (a throwaway plugin's rules/AGENTS.md was loaded as active session context by a real agy -p call, and a skill in the same plugin was listed and invoked by its bare name — no vendor-documentation-only claim)
6 Settings file in config/ config/ ✅ config/claude/settings.json.template ✅ config/gemini/settings.json ✅ config/copilot/settings.json.template ❌ No config/antigravity/ settings template — [GAP]. Antigravity CLI settings are written to ~/.gemini/antigravity-cli/settings.json directly by the CLI and the setup script; no repo template equivalent exists. Evidence: scripts/setup-antigravity-interactive.sh (PR #443) does not write a config/antigravity/ file
7 Active workspace settings file .claude/, .gemini/, .github/copilot/ ✅ .claude/settings.json ✅ .gemini/settings.json (spec 0191.delta-01) — single purpose: skills.disabled names exactly init-personal-profile and init-soul, so Gemini CLI's SkillCommandLoader does not mint a competing skill-derived /<name> slash command for either, leaving the native .gemini/commands/<name>.toml workspace command the sole claimant of the bare name (row 5). Declares no other key and disables no other skill (R15). Distinct from ~/.gemini/settings.json, which the CLI still loads at user scope and which this repo's own config/gemini/settings.json template deploys there — that template carries no skills key and is untouched by this file. ✅ .github/copilot/settings.json ❌ No in-repo workspace settings file — [GAP]. Antigravity CLI reads settings from ~/.gemini/antigravity-cli/settings.json only; no .agents/settings.json analog is defined. Evidence: scripts/setup-antigravity-interactive.sh (PR #443) writes no workspace settings file to the project tree
7b User-level layered context (00–65 priority files) config/SOUL.md, config/level/, config/ORGANIZATION.md, config/PROFILE.md, config/expertise/, config/teams/, artifacts/core/rules/60-tools.md (priority 60, core), config/TOOLS.md (priority 65, overlay) ✅ ~/.claude/rules/60-tools.md (core) + ~/.claude/rules/65-org-tools.md (overlay), plus other *.md files — deployed by setup-claude-interactive.sh ✅ ~/.gemini/60_TOOLS.md (core) + ~/.gemini/65_TOOLS.md (overlay), plus other [0-6][0-9]_*.md files — deployed by setup-gemini-interactive.sh. Gemini loads only files enrolled in context.fileName (~/.gemini/settings.json ← config/gemini/settings.json), unlike Claude/Copilot which read an entire directory, so every deployed overlay MUST also be enrolled there (e.g. 65_TOOLS.md; the priority-66 org-rules overlay is enrolled likewise, see row 7e) (#576) ✅ ~/.copilot/instructions/60-tools.instructions.md (core) + ~/.copilot/instructions/65-org-tools.instructions.md (overlay), plus other *.instructions.md files — deployed by setup-copilot-interactive.sh ✅ ~/.gemini/antigravity-cli/ — priority [0-6][0-9]_*.md files deployed by setup-antigravity-interactive.sh; concatenated into ~/.gemini/config/AGENTS.md (the single system context file Antigravity CLI reads) at the end of each setup run; any superseded ~/.gemini/GEMINI.md context file is removed to prevent duplicate/stale rule loading (spec 0054 / PR #443; spec 0061 delta-01 / issue #478; spec 0061 delta-02 / issue #1082)
7c MCP server configuration file — pre-existing non-reserved mcpServers declarations are preserved across a setup run on all four CLIs (spec 0089). The names mempalace / sequentialthinking are framework-reserved (MCP_RESERVED_NAMES in scripts/lib/common.sh): framework-wins on selection (the reserved entry replaces any prior one), removed on decline (a decline that since spec 0113 delta-02 no longer exists for mempalace — setup registers it against the shared HTTP daemon by default, row 10; sequentialthinking keeps its decline), each with a non-silent warning + timestamped backup. The three overwrite-based setups (Gemini/Copilot/Antigravity) get this by folding the captured pre-run config back over the framework write via the shared merge_preexisting_mcp_servers helper; Claude already preserves declarations through its idempotent claude mcp add and is the R5 reference behavior. Org-declared servers are folded as a third precedence tier on top of this merge — see row 7h. config/, scripts/lib/common.sh (merge_preexisting_mcp_servers) Spec 0113 — the mempalace entry switches transport. It is registered against the shared HTTP daemon instead of a per-session stdio command, through one helper emitting four native shapes: Claude claude mcp add --transport http … --header, Gemini and Copilot {type:"http",url,headers}, Antigravity {serverUrl,headers} with no transport key. Built rather than reused: the spec-0091 appliers skip every name in MCP_RESERVED_NAMES by design and mempalace is reserved, so they refuse the one server this needs. The switch is a machine-wide transaction (task mempalace:switch-http) — readable/writable floor first, capture, apply, restore on failure, and report which assistant is in which arrangement when a restore itself fails. A scoped single-CLI run stays legitimate and sits outside that obligation. The arrangement classifier reports a configuration that does not parse as JSON as unknown (neither arrangement), not none — a non-parseable file is residue, not a convergeable no-registration (spec 0165 R3). ✅ User-managed via claude mcp add / ~/.claude/mcp.json (no repo template — managed per-user by the CLI). Reference behavior for spec 0089 R5 — the idempotent server-add already preserves operator declarations ✅ config/gemini/settings.json → ~/.gemini/settings.json (2 servers: MemPalace, SequentialThinking; no forge MCP — CLI-only forge access per spec 0090). Preserves pre-existing non-reserved mcpServers across setup (spec 0089) ✅ config/copilot/mcp-config.json.template → ~/.copilot/mcp-config.json (2 servers: MemPalace, SequentialThinking; no forge MCP — CLI-only forge access per spec 0090; deployed by setup-copilot-interactive.sh). Preserves pre-existing non-reserved mcpServers across setup (spec 0089) ✅ ~/.gemini/config/mcp_config.json mcpServers block written by setup-antigravity-interactive.sh (spec 0054 / PR #443). Preserves pre-existing non-reserved mcpServers across setup despite its empty MCP base (spec 0089)
7d MemPalace MCP backend scripts/lib/mempalace-http-wrapper.py + scripts/lib/mempalace_pin.py + scripts/doctor-mempalace.sh + ADR 0006. Launch-time version guard (spec 0108): the wrapper determines, inside the very process that goes on to serve and before it begins serving, the MemPalace version its own interpreter resolves, and refuses to serve when that version is outside the range pinned in scripts/lib/common.sh — Phase A range-checks importlib.metadata.version("mempalace") ahead of import chromadb, Phase B independently range-checks mempalace.__version__ after the handoff import (never an equality test against Phase A's value: mempalace/version.py is a hand-maintained literal, so a .postN rebuild disagrees with both values in range). The comparator is packaging when importable and a closed stdlib whitelist grammar otherwise, never more permissive than packaging. Refusals name the version, the range, the interpreter, the comparator and the remedy on stderr only, then exit non-zero; the session's protocol channel carries nothing extra on the happy path. Operator diagnostic: bash scripts/doctor-mempalace.sh (task mempalace:doctor) reports, for every MemPalace command on PATH and for each of the four CLI registrations, the resolved path, the interpreter, the declared pin and the served version, and exits non-zero on divergence. Uniform across all four CLIs because the guard lives in the one wrapper they all invoke; a registration naming a checkout that predates the guard is reported as GUARD ABSENT, distinct from an argv routing through no wrapper at all (UNGUARDED). Residue repair (spec 0165): bash scripts/repair-mempalace-http.sh (task mempalace:repair) detects the residue repetition cannot reach — a configuration that does not parse as JSON, or whose mempalace registration matches neither the http nor the stdio shape — reports each affected assistant with its config path, backup existence and available actions, and repairs it with --restore-backup (most recent usable backup, mode preserved, 0600 on a bearer token and 0600 when the backup's mode would deny its own owner read, so the restored config is one the repair can verify and converge on) or --reset-none (removes the registration). It restores recognisability only; it never registers anyone for http and never places the bearer token in argv. Shared MCP HTTP daemon (spec 0113): the four CLIs no longer each spawn a stdio memory server; one supervised daemon holds the palace writer lease and serves all of them over --transport http, so concurrent sessions stop contending for it (the failure was MCP error -32001 for every session but the first to mutate). Its ExecStart is ~/.crewrig/mcp-daemon-launcher.sh, installed OUTSIDE the repository so a revert cannot delete the program the unit names; the launcher reads the 0600 bearer token and refuses to start when it is absent or empty — MemPalace requires a token only on a non-loopback bind and its auth gate short-circuits on an empty one, so serving without it would silently disable authentication for every client. Uniform across all four CLIs: the daemon and the launcher are one shared mechanism, and only the registration shape differs (see row 7c). #749: the wrapper's orphan-reaper thread is armed only for the stdio transport. A supervised daemon is reparented to PID 1 by construction, so the reaper read that as an abandoned session and called os._exit(0) every poll interval — the daemon killed itself roughly once a minute, silently, with exit code 0. The default port also moved from 8021 (registered as intu-ec-client, and held by a launchd system service on the first machine this ran on) to 41893, outside the ephemeral range; a taken port is now refused once with a diagnostic instead of respawned forever (#748). Half-converted writability & lockout detection (spec 0172): When the shared MCP HTTP daemon is serving and healthy, any assistant configured in stdio mode is permanently locked out of writing by the daemon's lease. scripts/status-mcp-server.sh (task mempalace:status) flags stdio registrations as stdio (LOCKED OUT by shared daemon) and exits non-zero (exit 1); scripts/doctor-mempalace.sh (task mempalace:doctor) reports the lockout as a failure; the daemon launcher warns on stderr at startup if stdio assistants remain; and 60-tools.md mandates task mempalace:status as the primary diagnostic step upon receiving error -32001 (Peer MCP writer active). Token rotation (spec 0176): scripts/switch-mempalace-http.sh --rotate (task mempalace:rotate-token) rotates the bearer token, replaces the daemon process (spec 0139), purges stale backup files holding the old token, and re-registers all present assistants with the new credential. ✅ MCP entry → bash <repo>/scripts/lib/tls-exec.sh <mempalace-python> <repo>/scripts/lib/mempalace-http-wrapper.py → HttpClient → chroma run (registered by setup-claude-interactive.sh via claude mcp add; the tls-exec.sh prefix comes from spec 0084). Guard + diagnostic per the mechanism cell — no CLI-specific code ✅ Same wrapper invoked from mcpServers.mempalace in ~/.gemini/settings.json; setup script substitutes __CREWRIG_REPO_DIR__ with the repo root at install time and rewrites .command to bash with the tls-exec.sh + interpreter prefix in .args. Guard + diagnostic per the mechanism cell — no CLI-specific code ✅ Same wrapper invoked from mcpServers.mempalace in ~/.copilot/mcp-config.json; setup script substitutes __CREWRIG_REPO_DIR__ with the repo root at install time and applies the same bash + tls-exec.sh prefix. Shared ChromaDB HTTP daemon (chroma run) supervised by launchd (macOS) or systemd-user (Linux); replaces concurrent PersistentClient instances (#98). Guard + diagnostic per the mechanism cell — no CLI-specific code ✅ Same wrapper invoked from mcpServers.mempalace in ~/.gemini/config/mcp_config.json — the file setup-antigravity-interactive.sh actually writes, NOT ~/.gemini/antigravity-cli/settings.json, which carries no mempalace key (verified); __CREWRIG_REPO_DIR__ substituted at install time (spec 0054), same bash + tls-exec.sh prefix. Guard + diagnostic per the mechanism cell — no CLI-specific code
7e Org-rules extension delivery (AGENTS.org.md) (spec 0020) AGENTS.org.md, AGENTS.md (@ import), scripts/setup-{gemini,copilot,antigravity}-interactive.sh ✅ Native recursive @AGENTS.org.md import (CLAUDE.md → @AGENTS.md → @AGENTS.org.md); no setup deployment needed ✅ Fallback: setup-gemini-interactive.sh deploys AGENTS.org.md → ~/.gemini/66_ORG_RULES.md (priority 66, after 65 org-tools). Gemini resolves @file imports only in GEMINI.md (absent at repo root), so the directive does not resolve — setup copy is the equivalent, enrolled in context.fileName alongside the other overlays (config/gemini/settings.json; #576) ✅ Fallback: setup-copilot-interactive.sh deploys AGENTS.org.md → ~/.copilot/instructions/66-org-rules.instructions.md. Copilot does not resolve @file includes in instruction files and auto-reads only the standard AGENTS.md name. Fallback drift: the deployed copy is taken at setup time; editing AGENTS.org.md requires re-running setup (org rules change rarely; setup is idempotent) ✅ Fallback: setup-antigravity-interactive.sh deploys AGENTS.org.md → ~/.gemini/antigravity-cli/66_ORG_RULES.md (priority 66); included in the ~/.gemini/config/AGENTS.md concatenation (spec 0061 delta-01). Same pattern as Gemini CLI (spec 0054)
7f User-space system-context store — on-demand direct read (spec 0068) artifacts/core/system-context/, scripts/lib/common.sh (install_dir), scripts/setup-*-interactive.sh, artifacts/core/rules/60-tools.md (retrieval protocol + stubs) ✅ ~/.crewrig/system-context/ installed by setup-claude-interactive.sh (install_dir, link/copy duality). Direct read PASS-default — the Step 1 probe read the store from an untrusted scratch dir with bare claude -p, no flags/config (the Read tool has unrestricted absolute-path access). Evidence: docs/research/system-context-sandbox-probe.md §4.1 ⚠️ ~/.crewrig/system-context/ installed by setup-gemini-interactive.sh (same install_dir call). Direct read could-not-probe — auth ineligible on the authoring machine (IneligibleTierError, free-tier deprecated), so the read is unverifiable; Gemini also gates tool use on workspace trust by default. [R6 trigger candidate] for the deferred dedicated retrieval service; covered at runtime by the store's explicit-signal fallback. No unverified Gemini-specific config is written. setup-gemini-interactive.sh now prints actionable store-access guidance naming the per-invocation grant (--include-directories / GEMINI_CLI_TRUST_WORKSPACE=true / --skip-trust), still writing no durable config (spec 0075 R6). Evidence: probe doc §4.4 ⚠️ installed by setup-copilot-interactive.sh (same install_dir call). Direct read PASS only with a path grant — reads via per-invocation --add-dir/--allow-all-paths or interactive approval; path verification denies the out-of-workspace read otherwise, and trustedFolders in ~/.copilot/config.json does not grant a durable cross-project read (tested, denied). No ineffective durable config is written; headless-deny covered by the fallback stub. [R6 trigger candidate]. setup-copilot-interactive.sh now prints actionable store-access guidance naming the per-invocation grant (--add-dir / --allow-all-paths), still writing no durable config (spec 0075 R6). Evidence: probe doc §4.3 ✅ ~/.crewrig/system-context/ installed by setup-antigravity-interactive.sh (same install_dir call). Direct read PASS-default — the Step 1 probe read the store from an untrusted scratch dir with bare agy --print, no flags/config. Evidence: probe doc §4.2
7g MemPalace layered wake-up (L0–L3) MemPalace mempalace.layers exposes Layer0/1/2/3 + MemoryStack.wake_up()/recall()/search()/status(); the mempalace CLI exposes wake-up; the mempalace.mcp_server tool registry exposes none of these. Installed 3.6.0 (pinned >=3.6.0,<3.7); issue #415 spike, re-verified 2026-07-20 against the live 36-tool MCP surface (issue #566, e2e sidecar image docker/e2e/mempalace.Dockerfile) ❌ via MCP — uniform (same MCP surface, per row 7d); ✅ in CLI/library but forbidden as an ad-hoc agent action ❌ via MCP — uniform (same MCP surface, per row 7d); ✅ in CLI/library but forbidden as an ad-hoc agent action ❌ via MCP — uniform (same MCP surface, per row 7d); ✅ in CLI/library but forbidden as an ad-hoc agent action ❌ via MCP — uniform (same MCP surface, per row 7d); ✅ in CLI/library but forbidden as an ad-hoc agent action
7g Custom root-CA / native-TLS delegation (spec 0084) — opt-in setup offer writes only ~/.crewrig/tls-env.sh (never the shell profile); MCP registrations and the launchd/systemd ChromaDB daemon are routed through the shared scripts/lib/tls-exec.sh wrapper so runtime network inherits the trust. Never disables verification. Shared mechanism ⇒ full parity. scripts/lib/tls-delegation.sh, scripts/lib/tls-exec.sh, scripts/setup-*-interactive.sh, docs/runbooks/custom-ca-tls-trust.md ✅ offer in setup-claude-interactive.sh; MCP via claude mcp add … -- bash tls-exec.sh … ✅ offer in setup-gemini-interactive.sh; mcpServers.* in ~/.gemini/settings.json routed through tls-exec.sh ✅ offer in setup-copilot-interactive.sh; mcpServers.* in ~/.copilot/mcp-config.json routed through tls-exec.sh ✅ offer in setup-antigravity-interactive.sh; mcpServers.* in ~/.gemini/config/mcp_config.json routed through tls-exec.sh
7h Org-declared MCP server delivery (spec 0091) — the org-owned root manifest mcp-servers.org.json (a CLI-agnostic neutral schema keyed by name: transport stdio/http/sse → command/args/env or url/headers) is translated into each CLI's native MCP shape and folded after the spec-0089 operator merge, so precedence is framework-reserved > org > operator-pre-existing. R10 — an org declaration under a reserved name (mempalace/sequentialthinking) is NOT applied (framework wins), warned. R11 — a non-reserved org name that collides with an operator entry wins, warned. stdio and http/sse reach all four CLIs (grounded). The neutral url maps to each CLI's native remote key: Antigravity serverUrl (no transport type), Copilot/Claude {type,url}, and — repaired by spec 0180 PLAN v5 step 4 (2026-08-25 maintainer arbitration) — Gemini httpUrl (http) or url (sse), with no type/transport key at all; Gemini infers the transport from WHICH key is present. Claude's R13 coverage is realized by the hermetic-equivalent — a pure org_mcp_to_claude_argv unit test + structural call-site assertions — the sanctioned shape, NOT a silent parity gap (mirrors how spec 0089 made Claude its R5 reference rather than an end-to-end test; the imperative claude mcp add path is undrivable hermetically without the claude binary + a live ~/.claude.json). mcp-servers.org.json, scripts/lib/common.sh (read_org_mcp_manifest, org_mcp_to_native, apply_org_mcp_servers, org_mcp_to_claude_argv, register_org_mcp_claude), scripts/setup-*-interactive.sh, docs/org-mcp-declaration.md ✅ imperative: register_org_mcp_claude iterates the manifest and issues claude mcp add --scope user per entry (stdio [-e K=V]… <name> -- cmd args; http/sse --transport <t> <name> <url> [--header …] — grounded against claude mcp add --help). R11 remove-then-add is guarded — the prior entry is captured from ~/.claude.json and restored if the re-add fails, so a partial run never drops the operator's server ✅ org_mcp_to_native gemini → native entry folded into ~/.gemini/settings.json by apply_org_mcp_servers. stdio {command,args,env}; http/sse {httpUrl|url,headers} — no type/transport key at all, transport inferred from which key is present. Re-grounded (spec 0180 PLAN v5 step 4, 2026-08-25): gemini-cli 0.46.0's own schema (bundle/docs/reference/configuration.md:2251-2287) lists command, args, env, cwd, url, httpUrl, headers, timeout, trust, description, includeTools, excludeTools — no type key anywhere. The earlier "grounded empirically" run (gemini mcp add --scope user -t http -H …, temp HOME) actually observed gemini mcp add's OWN CLI-flag echo, not the schema the extension/settings file itself accepts, and the shape it recorded ({type,url}) does not appear in that schema — it was a defect, not a second valid shape, now repaired ✅ same applier into ~/.copilot/mcp-config.json. stdio {type:"stdio",command,args,env} (the shipped template convention; the CLI's own mcp add writes type:"local", both accepted); http/sse {type,url,headers} — grounded empirically (copilot mcp add --transport http --header …, temp HOME) ✅ stdio delivered into ~/.gemini/config/mcp_config.json as {command,args,env}; http/sse delivered as the Antigravity-native {serverUrl,headers} shape — the neutral url is translated to serverUrl (not url), with no transport type field (one shape covers http + Streamable-HTTP/SSE). Grounded against the official Antigravity MCP docs (mcp_config.json structure: stdio {command,args,env,cwd?} + remote {serverUrl,headers?}), which supersede spec 0054's stale "format not publicly documented" note — closing the former remote-transport gap. Auth extras (authProviderType/oauth) are out of scope for this base declaration
8 Hook-integration manifest — prompt submissions distinguish human prompts ([USER] / user-prompt) from automated harness injections ([HARNESS] / harness-injection) via structural pattern matching (spec 0161) hooks/ ✅ hooks/claude-transcript-hooks.json (PreToolUse for git guard; UserPromptSubmit, PostToolUse, Stop, SessionEnd for transcripts; setup rewrites transcripts to installed target and guard to in-repo absolute path — spec 0169) ✅ hooks/gemini-transcript-hooks.json (BeforeTool for git guard; BeforeAgent, AfterTool, AfterModel, SessionEnd for transcripts; setup rewrites transcripts to installed target and guard to in-repo absolute path — spec 0169) ✅ hooks/copilot-transcript-hooks.json — deployed as opt-in by setup-copilot-interactive.sh to: ~/.copilot/hooks/copilot-transcript-hooks.json (user-level, all projects). Committed settings.json has "hooks": {} by design and is NOT mutated at workspace level. The manifest uses the documented object schema: hooks is an object keyed by camelCase config key (preToolUse for git guard; sessionStart, userPromptSubmitted, postToolUse, agentStop, sessionEnd for transcripts), each mapping to an array of {"type": "command", "bash"/"command": "…"} entries; version is the integer 1. Setup rewrites transcripts to installed target and guard to in-repo absolute path (spec 0169). The hook payload's hook_event_name stays PascalCase regardless, so hooks/mempalace-transcript.sh is unaffected. ✅ hooks/antigravity-transcript-hooks.json (Stop only) — deployed as opt-in by setup-antigravity-interactive.sh to: ~/.gemini/config/hooks.json (the GLOBAL customization root, all projects), with every command rewritten to the absolute installed path ~/.gemini/antigravity-cli/hooks/mempalace-transcript.sh and the event name appended as an argument. NOT the Gemini shape: the top level is a map of NAMED hooks ({"<name>": {"<Event>": [...]}}), and the CLI's only events are PreToolUse/PostToolUse (grouped, matcher+hooks) and PreInvocation/PostInvocation/Stop (flat). The four names spec 0056 shipped do not exist; spec 0116 supersedes 0056 R2/R4/R6. Stop is the only event registered under the named transcript hook crewrig-mempalace-transcript, and it is the only one that fires once per turn. The manifest also carries the named hook crewrig-worktree-git-guard (PreToolUse grouped run_command, deployed to the repository path, not installed — spec 0116 delta-03 R27/R28; spec 0169). Measured on agy 1.0.16 (spec 0116 delta-02 R26 — the exercise, not just the counts). Exercise A, one agy --print turn issuing three shell commands: PreInvocation 4×, PostInvocation 4×, PreToolUse 3×, Stop 1× — the other four have roughly per-tool-round cardinality and the CLI runs hooks synchronously, blocking the agent loop. Exercise B, two turns in ONE conversation (agy --print, then agy --continue --print): Stop 2×, same conversationId, executionNum 0 on both. B is the one that clears R23's bar — A alone cannot separate once-per-turn from once-per-execution-process, since it contains one turn. B also shows executionNum does NOT accumulate, so conversationId+executionNum is not a unique key and per-turn idempotence is not buildable from the payload at all (spec 0116 delta-01; the parent spec mislabelled PreInvocation as a start-of-turn event, on the strength of a single-prompt probe in which one invocation equals one turn). [GAP-content] Antigravity entries are turn markers only — [AGENT] Session turn completed (<terminationReason>) — where Claude, Gemini and Copilot persist actual prompt and response text: no Antigravity hook payload carries either, so the content is unavailable rather than dropped (specs/0116 → Out of scope). Evidence: the CLI's own ~/.gemini/antigravity-cli/builtin/skills/agy-customizations/docs/hooks.md, confirmed by live agy 1.0.16 probes (spec 0116 / issue #724)
8b Transcript hook write path (spec 0164) — routes through the shared MCP HTTP daemon via curl (mempalace_add_drawer JSON-RPC) instead of writing to the palace directly via Python, replacing the spec-0110 in-process lock bypass. Fails loud (exit 4, DAEMON_UNREACHABLE) if the daemon is offline, or (exit 5, ADD_FAILED) if the tool call fails. Cap on execution is 5.0 seconds via curl --max-time 5. hooks/mempalace-transcript.sh ✅ All CLIs share the same bash hook; the transport shift applies universally without CLI-specific configuration ✅ same ✅ same ✅ same
9 Project-dir env var consumed by hooks hooks/*-transcript-hooks.json ✅ $CLAUDE_PROJECT_DIR ✅ ${GEMINI_PROJECT_DIR} ❌ No $COPILOT_PROJECT_DIR — [GAP]. Workaround: hook reads workspace path from stdin JSON, falls back to $PWD ❌ No $ANTIGRAVITY_PROJECT_DIR — [GAP], but no longer worked around with $PWD, which was actively wrong here: a handler's working directory is the directory holding hooks.json, so $PWD resolved under ~/.gemini/config/ rather than under any project. Since spec 0116 the hook reads the project path from the stdin payload's workspacePaths[0], falling back to the git root then $PWD. Evidence: docs/hooks.md ("The working directory is set to the directory containing hooks.json") and a captured payload carrying workspacePaths, conversationId, transcriptPath. Caveat: workspacePaths was [] in a headless --print run outside a workspace, hence the retained fallback chain
10 Interactive setup script scripts/ Spec 0113 delta-02: each script makes the shared MemPalace MCP HTTP daemon the default outcome of setup (ensure_mempalace_http, no opt-in prompt): it probes the daemon with an authenticated accept probe (never /healthz), installs and starts it when it is not serving, and registers the HTTP entry — falling back to the stdio arrangement only on a reported failure, naming the repair (task mempalace:status / task mempalace:switch-http, then re-run setup). Wiring all four is mandatory, not tidiness: mempalace is in MCP_RESERVED_NAMES, so merge_preexisting_mcp_servers deliberately does not preserve an operator entry under that name and the framework write that replaces it is stdio-shaped — a setup run on an unwired CLI would silently revert the entry to stdio. A single-CLI run switches that CLI only and reports which others it leaves behind; the machine-wide all-or-nothing obligation belongs to task mempalace:switch-http. See docs/runbooks/mempalace-mcp-server.md for the conversion command, daily operations, and the token-replacement procedure (spec 0136). ✅ scripts/setup-claude-interactive.sh (offers MemPalace auto-install via pipx; MemPalace itself now defaults to the shared HTTP daemon — see the delta-02 clause; installs built artifacts from dist/<tier>/ to ~/.claude/skills+agents — library automatically, community/org per opt-in prompt — ADR-0011, spec 0019; the library tier's dist/library/.claude staging output is now auto-built on the fly via bash scripts/build-components.sh --target claude if missing, before the install step — spec 0107, issue #618) ✅ scripts/setup-gemini-interactive.sh (same HTTP-by-default MemPalace registration; installs from dist/<tier>/ to ~/.gemini/skills+agents — library auto, community/org opt-in; dist/library/.gemini is auto-built via bash scripts/build-components.sh --target gemini if missing, before the install step — spec 0107, issue #618) ✅ scripts/setup-copilot-interactive.sh (same HTTP-by-default MemPalace registration; installs skills from dist/<tier>/.github/skills/ to ~/.copilot/skills — library auto, community/org opt-in; agents skipped, see row 4 [GAP-confirmation]; dist/library/.github/skills is auto-built via bash scripts/build-components.sh --target copilot if missing, before the install step — spec 0107, issue #618) ✅ scripts/setup-antigravity-interactive.sh (same HTTP-by-default MemPalace registration; installs from dist/<tier>/.agents/ to ~/.gemini/config/skills+agents — the machine-local customization root the vendor documents — library auto, community/org opt-in. Spec 0123 moved both kinds here from ~/.gemini/antigravity-cli/, where the assistant never looked: R12 evidence is asymmetric and both halves are recorded. Skills — documented AND observed: the vendor's own ~/.gemini/antigravity-cli/builtin/skills/agy-customizations/SKILL.md lists skills/<name>/SKILL.md relative to the global root ~/.gemini/config/, and a sentinel there is found while sentinels at the two other candidate roots are not. Agents — OBSERVATION ONLY: that same table lists five customization types (Rules, Skills, Plugins, Hooks, MCP Servers) with agents absent entirely, so nothing about agents can be read out of the vendor documentation at all. Measured agy 1.1.11, 2026-08-11, two reproducible runs; method, full matrix and the re-runnable procedure (R13/R14) in docs/runbooks/antigravity-discovery-probe.md, driven by scripts/probe-antigravity-discovery.sh. The install step is now a VERIFIER rather than a reporter of intent: install_antigravity_tier_to_home() in scripts/lib/common.sh asserts each component's presence at the destination before printing its Installed line (R5), reports any staged-vs-placed shortfall (R4), and returns non-zero for a kind that staged components and placed none (R6) — the superseded code globbed "$staging/agents"/*.md for a flat shape the build has never produced, so it installed no agent at all and exited 0. migrate_antigravity_superseded_components() clears the superseded placement on re-run (R8/R9), sweeping all three non-core tiers regardless of which the run installs. Both helpers are hermetically covered by scripts/tests/test-antigravity-component-install.sh; generates ~/.gemini/config/AGENTS.md by concatenating deployed context files; spec 0116 adds the MemPalace transcript opt-in its three siblings already had — two fzf prompts, then deploy_antigravity_transcript_hooks() in scripts/lib/common.sh. The deployment lives in a helper rather than inline, because the interactive scripts cannot run end-to-end in CI (fzf, the agy guard, the chroma daemon) and spec 0116 R17 requires hermetic coverage; the prompts are asserted structurally by scripts/tests/test-setup-antigravity-transcript.sh; spec 0054 / PR #443, spec 0061 delta-01 / issue #478; dist/library/.agents is auto-built via bash scripts/build-components.sh --target antigravity if missing, before the install step — spec 0107, issue #618)
11 Transcript backfill script scripts/ ✅ scripts/import-claude-history.sh (reads ~/.claude/projects) ✅ scripts/import-gemini-history.sh (reads ~/.gemini/tmp) ✅ scripts/import-copilot-history.sh (reads ~/.copilot/session-state/<session-id>/events.jsonl) ✅ scripts/import-antigravity-history.sh (reads ~/.gemini/antigravity-cli/history.jsonl; spec 0055 / PR #446)
12 Component-management script — since spec 0119 every command serves the ordered overlay tier set library → community → org and never core, on both request shapes, through the shared resolver scripts/lib/component-resolve.sh. A named request that resolves in no served tier fails, naming every root examined; a name resolving to more than one component in one landing zone installs nothing under that name and is reported (R15). The compiled types prune and rebuild dist/<tier>/<cli-root> on demand — on a miss for a named request, up front for an unnamed one. scripts/ ✅ scripts/manage-claude-component.sh — claude-skills → ~/.claude/skills, basis dist/{library,community,org}/.claude/skills; policies → ~/.claude/rules, basis artifacts/{library,community,org}/policies; mcp-servers → ~/.claude.json via claude mcp add, basis artifacts/{library,community,org}/mcp-servers ✅ scripts/manage-workspace-component.sh — skills → ~/.gemini/skills and agents → ~/.gemini/agents, basis dist/{library,community,org}/.gemini/{skills,agents} (R2: the assisted setup reads the same); commands, hooks, policies → ~/.gemini/<type> and mcp-servers/themes → ~/.gemini/settings.json, basis artifacts/{library,community,org}/<type>. [GAP] link mode warns but does not prompt, unlike the other three: scripts/install-workspace.sh drives this script seven times per run, so a per-type confirmation would fire seven times for one task link-workspace. The disclaimer closes the harm; the prompt asymmetry is recorded here rather than left silent. ✅ scripts/manage-copilot-component.sh — skills/commands → ~/.copilot/skills (was .github/skills in the committed tree, which spec 0119 R3 forbids for a non-core tier), basis dist/{library,community,org}/.github/skills; mcp-servers → ~/.copilot/mcp-config.json, basis artifacts/{library,community,org}/mcp-servers. [GAP] agents is refused with a non-zero status: R3 bars the project tree for a non-core tier, R6 bars core, and R4 declines to oblige a landing zone where the assisted setup installs none — see rows 4 and 10 for the underlying repo-level-layout gap. ✅ scripts/manage-antigravity-component.sh — antigravity-skills → ~/.gemini/config/skills (spec 0123 R7: the same location a full setup run places it, row 10), basis dist/{library,community,org}/.agents/skills, followed by a per-name sweep of the superseded placement. Only the skills DEST moved: ANTIGRAVITY_HOME also resolves policies and mcp-servers below, both excluded by spec 0123 for want of a probe, so the obvious one-line edit to that constant would have relocated two kinds silently; policies → ~/.gemini/antigravity-cli/rules and mcp-servers → ~/.gemini/antigravity-cli/settings.json, basis artifacts/{library,community,org}/<type> (both were hardcoded to community alone before spec 0119); spec 0058
13 Plugin / extension build script scripts/ ✅ scripts/build-claude-plugin.sh → dist-claude-plugin/.claude-plugin/marketplace.json; also renders extension command pivots commands/*.md → ephemeral plugin skills (row 5b); copies the extension's hooks/ handler tree into the output when the generic hooks section is declared (spec 0179 R9/v2-F1 — the hook FILE, hooks/hooks.json, is emitted separately by render_plugin, AFTER this builder runs; this builder no longer reads a per-CLI claude.hooks key). Spec 0180 (issue #1006): generates .mcp.json from the generic mcpServers section via the shared translator scripts/lib/extension-manifest.sh's ext_mcp_native claude, rewriting the one neutral path token ${extensionRoot} to ${CLAUDE_PLUGIN_ROOT} — resolved by Claude Code itself at plugin-load time (no render-time absolute path, R8); copies dist/ + package.json alongside it (R16) ✅ scripts/build-extension.sh --target gemini — Gemini-side symmetric counterpart to build-claude-plugin.sh (spec 0042, generalized by spec 0173). Renders the complete installable tree — the verbatim source plus gemini-extension.json and commands/*.toml — into build/extensions/<name>/; nothing is committed and nothing is installed by this script (that is row 14's install-extension.sh). Also renders the generic hooks section (spec 0179) via the shared translator scripts/lib/extension-hooks.sh, which reads its per-target constants (shell-tool name, time unit, root-token form, hook file, match-all form) from scripts/lib/extension-targets.json — a descriptor shared across S2/S3/S4 (issues #1005/#1006/#1007). Spec 0180: mcpServers is translated via ext_mcp_native gemini (the neutral ${extensionRoot} rewritten to Gemini's own ${extensionPath}, which gemini-cli itself resolves when it loads the extension in place) instead of the prior verbatim-passthrough; the shared translator (scripts/lib/common.sh org_mcp_to_native) also repairs a live defect in the same diff — Gemini's remote transport now emits httpUrl/url with no type key, instead of the {type,url} shape that landed a declared http server on Gemini's SSE transport (row 7h). ✅ scripts/build-copilot-plugin.sh — mirrors build-antigravity-extension.sh; resolves extension name across tiers, reads extension.json (no gemini-extension.json fallback — retired, spec 0183 R14: an extension with no committed extension.json fails naming scripts/migrate-extension.sh), generates plugin.json at output root, copies skills, flattens agents from agents/<name>/AGENT.md → agents/<name>.agent.md, renders command pivots as skills when convertToSkills is true, copies the extension's hooks/ handler tree into the output when the generic hooks section is declared (spec 0179 R9/v2-F1 — the hook FILE itself is emitted separately, by render_plugin in row 13's own build-extension.sh, AFTER this builder runs, via scripts/lib/extension-hooks.sh; this builder no longer reads copilot.hooks, retired by 0065 delta-01). Output: dist-copilot-plugin/<name>/ (spec 0065). Spec 0180 (issue #1006): generates .mcp.json at the plugin root ("source": "plugin" in Copilot's own registry — probed empirically, 1.0.80) via ext_mcp_native copilot, rewriting the neutral token to ${COPILOT_PLUGIN_ROOT} — confirmed by a live non-interactive session actually spawning the declared stdio server with the token resolved (Copilot defaults a plugin server's cwd to its own plugin root, so even a bare relative form resolves); copies dist/ + package.json alongside it (R16) ✅ scripts/build-antigravity-extension.sh — mirrors build-claude-plugin.sh; resolves extension name across tiers, reads extension.json (no gemini-extension.json fallback — retired, spec 0183 R14: an extension with no committed extension.json fails naming scripts/migrate-extension.sh), generates plugin.json at output root, copies skills/agents, renders command pivots as skills when convertToSkills is true, renders the context file via the shared resolver when a context source is declared (row 5d, spec 0181, issue #1007 — no longer a copy from a per-CLI antigravity.contextFileName key, retired), copies the extension's hooks/ handler tree into the output when the generic hooks section is declared (spec 0179 R9/v2-F1 — the emitted command is working-directory-relative, so the handler MUST land at hooks/<handler>, not beside hooks.json itself; the hook FILE is emitted separately by render_plugin, AFTER this builder runs; this builder no longer reads antigravity.hooks, retired by 0063 delta-01). Output: dist-antigravity-plugin/<name>/ (spec 0063). Spec 0180 (issue #1006), superseding the prior "copies MCP artifacts" note: generates mcp_config.json at the plugin root (plugins.md/mcp_servers.md's own documented location) via ext_mcp_native antigravity, which LEAVES the neutral ${extensionRoot} token UNRESOLVED — a live probe proved a relative command/args does NOT resolve against the plugin directory (Antigravity spawns a plugin's MCP server with the CLI's own launch directory as CWD, unlike Copilot's default, and unlike how Antigravity itself fires a hook). scripts/install-antigravity-extension.sh's POST-INSTALL step resolves the token once agy plugin install has placed the plugin at its real, deterministic destination (~/.gemini/config/plugins/<pluginName>/, confirmed to copy mcp_config.json verbatim, no rewriting of its own); copies dist/ + package.json alongside it (R16). Evidence: docs/runbooks/extension-mcp-token-probe.md (2026-08-25)
14 Plugin / extension install script scripts/ ✅ scripts/install-claude-plugin.sh — builds into a single SHARED local marketplace home (${CLAUDE_CONFIG_DIR:-$HOME/.claude}/local-marketplace/) OUTSIDE the working tree, NOT in-tree dist-claude-plugin/ (spec 0045): multiple extensions coexist in one marketplace, the manifest upserts each by name, and installs survive branch switches. Known limitation: the home is keyed only on CLAUDE_CONFIG_DIR, so two different repos under the same config root share one marketplace.json (top-level name is last-writer-wins) — single-repo assumption, multi-repo namespacing out of scope. The bare build path (row 13) still emits in-tree. ✅ scripts/install-extension.sh (tasks install-gemini-extension / install-gemini-extensions, debugging symlinking via link-gemini-extensions / link-gemini-extension-build / scripts/link-extensions.sh, and cross-CLI umbrella install-extension-all / scripts/install-extension-all.sh) — renders the extension (scripts/build-extension.sh --target gemini), then copies or links the resulting build directory into $GEMINI_HOME/extensions/<name> where Gemini CLI auto-discovers it (spec 0177, as amended by spec 0173 delta-01 R20). link-gemini-extension-build is a dedicated, documented debugging path pointed at the build directory, not an install path. ✅ scripts/install-copilot-plugin.sh — resolves the extension name across tiers, builds into dist-copilot-plugin/<name>/, and invokes copilot plugin install <output-dir> (spec 0065) ✅ scripts/install-antigravity-extension.sh — resolves the extension name across tiers, builds into dist-antigravity-plugin/<name>/, and invokes agy plugin install <output-dir> (spec 0063)
15 Build-components target and tier flags — --target <cli> selects the CLI; --tier <name> (repeatable, spec 0119) narrows the tier set and is CLI-agnostic, so it applies identically to all four targets. It exists for the per-component install path: filtered to the served overlay tiers a rebuild takes 0.57 s and writes nothing git-visible, where an unfiltered one takes 4.31 s and, on a branch carrying an unbuilt core source edit, leaves four unrequested tracked modifications per edited skill behind. scripts/build-components.sh ✅ --target claude · --tier <name> ✅ --target gemini · --tier <name> ✅ --target copilot · --tier <name> ✅ --target antigravity (spec 0053 / PR #441) · --tier <name>
15b REPO_DIR build-root override scripts/build-components.sh ✅ REPO_DIR env var — redirects the inferred repo root to an arbitrary directory (e.g. a temp dir for test isolation); introduced in spec 0018; CLI-agnostic ✅ same ✅ same ✅ same
15c Build-components declared query mode (--list-output-dirs) — queries declared relative output directories fast without building or checking prerequisites (yq/jq), consumed by scripts/check-core-paths.sh to enforce output governance (spec 0125); CLI-agnostic scripts/build-components.sh list_output_dirs ✅ same ✅ same ✅ same ✅ same
16 Taskfile entries Taskfile.yml ✅ setup-claude-interactive, install-claude-workspace, link-claude-workspace, install-claude-component, link-claude-component, build-claude-plugin, install-claude-plugin, import-claude-history, build-components-claude ✅ setup-gemini-interactive, import-gemini-history, build-components-gemini, install-gemini-extension, install-gemini-extensions, link-gemini-extensions, link-gemini-extension-build (debugging path, spec 0173 delta-01 R20), unlink-gemini-extensions, install-extension-all (plus deprecated install-extension, install-extensions, link-extensions, unlink-extensions) (spec 0177), migrate-extension (spec 0183, issue #1008 — converts an extension source tree off the retired components.<subject>.enabled shape; cross-CLI, listed here beside the other extension-lifecycle tasks) ✅ setup-copilot-interactive, import-copilot-history, build-components-copilot, build-copilot-plugin, install-copilot-plugin ✅ setup-antigravity-interactive, import-antigravity-history, build-antigravity-extension, install-antigravity-extension, build-components-antigravity, install-antigravity-component, link-antigravity-component (spec 0054/0055/0057/0058/0063)
17 Per-CLI extension manifest. Two committed manifests under version lockstep (spec 0044, as amended by delta-01), one built and asserted separately. An extension's package.json version is authoritative; the committed extension.json of the SAME extension MUST declare that SAME version — scripts/check-extension-manifest-version.sh is a static CI guard that fails on a divergent extension.json, and scripts/monorepo-release.sh rewrites it to nextRelease.version at release time. gemini-extension.json is no longer a committed sibling: it is a BUILD output (scripts/build-extension.sh --target gemini, spec 0173 delta-01), and its version lockstep is asserted separately, non-vacuously, by bash scripts/build-extension.sh --check's VERSION-DRIFT arm against the built manifest. extensions/<tier>/<name>/ ✅ extension.json (+ CLAUDE.md, RENDERED from the generic context.source — spec 0181, no longer a per-CLI claude.contextFileName key) — version MUST equal authoritative package.json version (consumed by build-claude-plugin.sh:64 → generated plugin.json) ✅ extension.json's generic sections and gemini: section render build/extensions/<name>/gemini-extension.json (+ GEMINI.md, RENDERED from context.source, spec 0181 — no longer a per-CLI gemini.contextFileName key nor a verbatim copy) — the built manifest's version MUST equal authoritative package.json version, asserted by --check (row 13/14; spec 0173 delta-01 R11) ✅ Copilot CLI plugin uses the same extension.json as Claude Code, with no copilot: section at all — copilot.pluginName was the section's only key and is retired, reducible from .name (spec 0183 R7/R8); build-copilot-plugin.sh reads .name directly. hooks was never admissible there (0065 delta-01 R1); hooks are declared once in the generic top-level hooks section (spec 0179) and translated by scripts/lib/extension-hooks.sh. ✅ Antigravity CLI uses the same extension.json as Claude Code, with no antigravity: section at all — antigravity.pluginName was the section's only surviving key (contextFileName already retired, spec 0181 R10) and is now itself retired, reducible from .name (spec 0183 R7/R8); build-antigravity-extension.sh reads .name directly (+ rules/AGENTS.md, RENDERED from context.source — spec 0181, row 5d). hooks was never admissible there (0063 delta-01 R1); hooks are declared once in the generic top-level hooks section (spec 0179) and translated by scripts/lib/extension-hooks.sh.
17b Extension hook subject (spec 0179, issue #1005) — a single generic hooks section, translated for every target by scripts/lib/extension-hooks.sh (invoked from scripts/build-extension.sh, row 13). The normative per-target correspondence — event names, matcher form, hook file, time unit, extension-root token — is docs/extension-hook-events.md, kept honest against the translator by scripts/check-extension-hook-map.sh (R6) and against unresolved tokens by scripts/check-extension-hook-tokens.sh (R12); both are re-grounded, per target, by scripts/probe-extension-hooks.sh / docs/runbooks/extension-hook-probe.md. Per-target constants live in scripts/lib/extension-targets.json, shared additively with #1006/#1007. scripts/lib/extension-hooks.sh, scripts/lib/extension-targets.json, docs/extension-hook-events.md ✅ PreToolUse/UserPromptSubmit → hooks/hooks.json at the plugin root, envelope-grouped, matcher on tool name (Bash), root token ${CLAUDE_PLUGIN_ROOT} ✅ PreToolUse/UserPromptSubmit → BeforeTool/BeforeAgent in hooks/hooks.json in the extension directory, envelope-grouped, matcher run_shell_command, root token ${extensionPath} ✅ PreToolUse/UserPromptSubmit → preToolUse/userPromptSubmitted in hooks.json at the plugin root, flat (inline matcher per handler), matcher bash, root token ${COPILOT_PLUGIN_ROOT} — grounded live (B1A), not by analogy (0065 delta-01 R12) ✅ PreToolUse only (UserPromptSubmit has no counterpart) → named-hook map in hooks.json at the plugin root, matcher run_command, no root token (working-directory-relative; handler MUST resolve to hooks/<handler>, v2-F1)
17c Extension-scoped MCP server subject (spec 0180, issue #1006; spec 0185, issue #1035) — a single generic mcpServers section, translated for every target by the SAME translator the org-level channel (spec 0091, row 7h) uses — scripts/lib/common.sh's org_mcp_to_native, reached through scripts/lib/extension-manifest.sh's ext_mcp_native (invoked from row 13's four builders). The neutral vocabulary and the positional path-token rule are validated by ext_validate_mcp_shape / ext_validate_mcp_tokens / ext_validate_mcp_names in the same file; both are re-grounded by scripts/probe-extension-mcp-token.sh / docs/runbooks/extension-mcp-token-probe.md. Framework-reserved names (mempalace, sequentialthinking) are refused everywhere (R12). cwd (for stdio) and timeout (for stdio, http, sse) are admitted into the neutral vocabulary (spec 0185 / issue #1035), translated symmetrically across both extension and org declaration channels. Peripheral fields (description, includeTools, excludeTools, trust) remain refused. scripts/lib/extension-manifest.sh (ext_mcp_native, the three validators), scripts/lib/common.sh (org_mcp_to_native), scripts/lib/extension-targets.json (mcpDelivery column), scripts/probe-extension-mcp-token.sh ✅ File .mcp.json at the plugin root. Resolver/moment — the neutral ${extensionRoot} is rewritten to ${CLAUDE_PLUGIN_ROOT} at RENDER time; Claude Code resolves that token itself when it LOADS the installed plugin (evidenced: the official marketplace plugin discord/.mcp.json, and CLAUDE_PLUGIN_ROOT substitution entries in the installed changelog). R13 collision — an extension-declared server is NAMESPACED as plugin:<pluginName>:<serverName>, never colliding with a bare name at user/project scope (both listed side by side) — confirmed live, 2026-08-25. R14 layout — src/ (source) + dist/ (build output) at the extension root; dist/ + package.json copied alongside .mcp.json ✅ File gemini-extension.json's own mcpServers key (no separate file — it is one of the six fields the built manifest carries, row 17). Resolver/moment — rewritten to ${extensionPath} at RENDER time; gemini-cli resolves that token itself when it LOADS the extension from the installed tree. R13 collision — settings.json (where the org/operator declaration lives) wins over an extension-declared server of the same name (bundle/docs/extensions/reference.md:151, citation). R14 layout — same as Claude. This subject ALSO repairs a live defect in the shared translator: Gemini's remote transport now emits httpUrl/url with no type key (row 7h) ✅ File .mcp.json at the plugin root, "source": "plugin" in Copilot's own registry (probed empirically, 1.0.80). Resolver/moment — rewritten to ${COPILOT_PLUGIN_ROOT} at RENDER time; Copilot resolves it AND defaults the server's cwd to its own plugin root when it SPAWNS the server (confirmed live: all three candidate forms spawn correctly, 2026-08-25). R13 collision — the PLUGIN-sourced entry wins over a same-named user-scope entry, which is written successfully but never surfaces in copilot mcp list/get — confirmed live, 2026-08-25. R14 layout — same as Claude ✅ File mcp_config.json at the plugin root (plugins.md/mcp_servers.md's own documented location). Resolver/moment — the neutral token is LEFT UNRESOLVED at render time (a relative command/args does NOT resolve against the plugin directory when Antigravity spawns a plugin's MCP server — confirmed live, no error surfaced, 2026-08-25); resolved by scripts/install-antigravity-extension.sh's POST-INSTALL step, once agy plugin install has placed the plugin at its real destination (~/.gemini/config/plugins/<pluginName>/, confirmed deterministic and copied verbatim). R13 collision — plugin tools "namespaced if necessary to prevent collisions" (plugins.md, citation). R14 layout — same as Claude
18 CI workflow targeting the CLI .github/workflows/ ✅ claude.yml (anthropics/claude-code-action on @claude mentions) ❌ No gemini.yml equivalent ✅ copilot.yml (placeholder — native @copilot mention trigger is set at repo / org level, not via this workflow) ❌ No antigravity.yml — [GAP]. agy --print is confirmed non-interactive (exit 0), but no published GitHub Action (analogous to anthropics/claude-code-action) exists for Antigravity CLI. Evidence: gh api /repos/antigravity-labs/agy-code-action → HTTP 404; gh api /users/antigravity-labs → HTTP 404 (org not found on GitHub); gh search repos "agy-action" — no relevant results
19 Documentation prose references README.md, AGENTS.md, CONTRIBUTING.md, DEVELOPMENT.md ✅ Documented throughout ✅ Documented throughout ✅ Documented throughout ✅ AGENTS.md pillars 1 and 5 updated in this PR (spec 0058)
20 .gitignore carve-outs .gitignore ✅ dist-claude-plugin/, .claude/settings.local.json, CLAUDE.local.md ❌ No Gemini-specific entries ✅ .github/copilot/settings.local.json, copilot-instructions.local.md, dist-copilot-plugin/ (spec 0065) ✅ dist-antigravity-*/ — covers dist-antigravity-plugin/<name>/ emitted by build-antigravity-extension.sh (spec 0063). Already present in .gitignore as dist-antigravity-*/ (added with spec 0057 / PR #449).
21 e2e dedicated-account auth flow scripts/e2e/auth-<cli>.sh, Taskfile.yml → e2e:auth:*, ADR 0002 ✅ scripts/e2e/auth-claude.sh — interactive claude /login in crewrig/e2e-claude:latest, persists ~/.crewrig-e2e/claude/{.credentials.json,.claude.json} (RW one-shot, RO at scenario time) ✅ scripts/e2e/auth-gemini.sh — interactive gemini "Login with Google" in crewrig/e2e-gemini:latest. Persists ~/.crewrig-e2e/gemini/ with oauth_creds.json, settings.json, google_accounts.json, google_account_id, installation_id, gemini-credentials.json, extension_integrity.json, trustedFolders.json, acknowledgments/, history/, projects.json, state.json, etc. — full ~/.gemini minus the antigravity* and tmp/ subtrees and backup leftovers (.bak, .ori, .orig). Bundle dir is chmod 0700. Mounted :ro at /run/gemini-creds; a container-side wrapper copies it to /home/agent/.gemini before exec gemini (see issue #147 §5). ⚠️ scripts/e2e/auth-copilot.sh — guidance-only, no container. Prints PAT setup steps for the test account; scenarios consume $COPILOT_GITHUB_TOKEN at run time via docker run -e. Nothing persisted under ~/.crewrig-e2e/copilot/. [GAP-soft] — empirically justified divergence (see Parity gaps). The e2e runner (#78) consumes this asymmetry first-class via mounts = [] in tests/e2e/defaults.toml [cli.copilot], so no CLI-branching is needed in tests/e2e/run.sh. When local.toml routes Copilot through Ollama Cloud, task e2e:auth:ollama persists an Ed25519 keypair under ~/.crewrig-e2e/ollama/ that is mounted RO into the copilot container (see ADR 0002 Decision 8). ❌ No scripts/e2e/auth-antigravity.sh — [GAP-soft]. No crewrig/e2e-antigravity:latest Docker image or e2e auth flow has been defined for Antigravity CLI. Adding e2e coverage is explicitly out of scope for spec 0058
22 e2e pillar 01 — layered context tests/e2e/scenarios/01-layered-context/, ADR 0005 ✅ host-orchestrated probe via claude -p; mounts ${CREWRIG_E2E_HOME}/claude ro; structural + LLM-judge assertions ✅ host-orchestrated probe via gemini -p; dual :ro mounts — ${CREWRIG_E2E_HOME}/gemini at /run/gemini-creds (auth bundle) and ${HOME}/.gemini at /run/gemini-rules (layered-context rules, [0-6]0_*.md only); container-side bootstrap copies both into /home/agent/.gemini before exec gemini (see issue #148 Decision 5 Revision 2) ✅ host-orchestrated probe via copilot -p; mounts ${CREWRIG_E2E_HOME}/copilot ro (auth-copilot.sh currently writes nothing here — scenario soft-skips if the dir is absent) ❌ No e2e coverage — [GAP-soft]. No crewrig/e2e-antigravity:latest image or dedicated auth flow exists. Out of scope per spec 0058
23 e2e pillar 02 — cross-tool memory tests/e2e/scenarios/02-cross-tool-memory/, ADR 0005 ✅ MemPalace sidecar (per-run volume); drawer write + read across two short-lived containers ✅ same path ❌ [GAP-soft] — crewrig/e2e-copilot:latest (docker/e2e/copilot.Dockerfile) does not embed the mempalace CLI, and Copilot CLI ships no MemPalace MCP integration in artifacts/community/. Empirical reproduction: docker run --rm crewrig/e2e-copilot:latest which mempalace → exit 1 (binary absent). Scenario excluded via applies_to = ["claude","gemini"] until the Copilot image gains a MemPalace surface. ❌ No e2e coverage — [GAP-soft]. Out of scope per spec 0058
24 e2e pillar 03 — skill build tests/e2e/scenarios/03-skill-build/, ADR 0005 ✅ runs scripts/build-components.sh --target claude inside the container; asserts .claude/ populated + first SKILL.md has metadata: frontmatter ✅ same with --target gemini → .gemini/ ✅ same with --target copilot → .github/ ❌ No e2e coverage — [GAP-soft]. Out of scope per spec 0058
25 e2e pillar 04 — harness loop tests/e2e/scenarios/04-harness-loop/, ADR 0005 ✅ simulated path (drawer write to wing=harness-friction + curator-style search); MemPalace sidecar; LLM-judge on the friction summary ✅ same simulated path ✅ same simulated path. [GAP-soft] — full interactive harness-report skill invocation is not yet driven non-interactively from any CLI; the simulation matches what the curator reads and is parity-equivalent for v1. ❌ No e2e coverage — [GAP-soft]. Out of scope per spec 0058
26 idea skill — IDEA lifecycle stage (spec 0066) artifacts/core/skills/idea/SKILL.md ✅ .claude/skills/idea/SKILL.md (built by scripts/build-components.sh --target claude; row 3 core path) ✅ .gemini/skills/idea/SKILL.md (built by scripts/build-components.sh --target gemini; row 3 core path) ✅ .github/skills/idea/SKILL.md (built by scripts/build-components.sh --target copilot; row 3 core path) ✅ .agents/skills/idea/SKILL.md (built by scripts/build-components.sh --target antigravity; row 3 core path)
27 User-gate validation backend (spec 0080, spec 0189) — the user-validate skill is the sole implementation of the backend-configurable gate protocol (R9); the per-user backend and cross-cutting options (translate / pedagogy / illustration / theme) are captured symmetrically by the four setup-*-interactive.sh scripts through the shared configure_validation_backend helper (scripts/lib/common.sh) and persisted to ~/.crewrig/validation.conf — per-user, machine-local, outside the core layer (R15/R16). When cross-cutting options are active, a bespoke presentation document is generated honoring translation, pedagogy, and theme; illustrations (R14, spec 0189 R4) are embedded as inline base64 data URIs to avoid browser sandbox security blocks. artifacts/library/skills/user-validate/SKILL.md, scripts/lib/common.sh (configure_validation_backend), scripts/setup-*-interactive.sh, ~/.crewrig/validation.conf ✅ built to dist/library/.claude/skills/user-validate/SKILL.md (--target claude; row 3 non-core tier routing to dist/<tier>/), auto-installed to ~/.claude/skills/user-validate/ by the interactive setup (row 10, library tier); setup opt-in via configure_validation_backend → ~/.crewrig/validation.conf; plannotator annotate <file> --gate --json detached invocation verified (Bash run_in_background re-invokes the agent on process exit; foreground sleep blocked — use Monitor for conditional waits); internal guaranteed floor ✅ built to dist/library/.gemini/skills/user-validate/SKILL.md (--target gemini; row 3 non-core tier routing to dist/<tier>/), auto-installed to ~/.gemini/skills/user-validate/ by the interactive setup (row 10, library tier); same setup opt-in → ~/.crewrig/validation.conf; plannotator annotate <file> --gate --json detached invocation verified (degrades to internal at gate time if infeasible — R4); internal guaranteed floor ✅ built to dist/library/.github/skills/user-validate/SKILL.md (--target copilot; row 3 non-core tier routing to dist/<tier>/), auto-installed to ~/.copilot/skills/user-validate/ by the interactive setup (row 10, library tier); same setup opt-in → ~/.crewrig/validation.conf; plannotator annotate <file> --gate --json detached invocation verified (same as Gemini); internal guaranteed floor ✅ built to dist/library/.agents/skills/user-validate/SKILL.md (--target antigravity; row 3 non-core tier routing to dist/<tier>/), auto-installed to ~/.gemini/config/skills/user-validate/ by the interactive setup (row 10, library tier; moved there by spec 0123 — under the superseded placement this skill was installed but never discovered); same setup opt-in → ~/.crewrig/validation.conf; plannotator annotate <file> --gate --json detached invocation verified (same as Gemini); internal guaranteed floor
27b Internal-backend interactive prompt (spec 0080 R5) — the internal backend realises the gate through a structured interactive prompt, never a prose question (preserves spec 0037) artifacts/library/skills/user-validate/SKILL.md (internal backend) ✅ AskUserQuestion (confirmed, used across the repo) ⚠️ [GAP-confirmation] — structured interactive-prompt equivalent unverified; evidence at DEV ⚠️ [GAP-confirmation] — same ⚠️ [GAP-confirmation] — same
28 Spec-id reservation (spec 0112 + delta-01) — scripts/reserve-spec-id.sh secures an id before the spec file exists (parentless git object pushed under the namespace named by .crewrig/spec-id-carrier; the guarantee is the remote's create-only compare-and-swap via --force-with-lease=<ref>:), and scripts/check-spec-id-reserved.sh is the pull-request backstop. CLI-agnostic by construction, so full parity is structural, not asserted: both are plain git-only Bash 3.2 scripts (spec 0111) invoked from the shell, with no per-CLI surface, no MCP tool, and no forge API — requirement 4 makes forge-independence a requirement, and AGENTS.md → Forge Access keeps the credential the one the contributor already holds for git. The behavioural contract each CLI's spec-author consumes is the exit code only: 0 secured, 3 allocated-but-unsecured (stdout carries unsecured-id: true verbatim), 1 genuine failure. Enforcement is CI-side, not CLI-side: the check-spec-id-reserved capability in ci/ci-capabilities.yml (portable, history-depth: full, no tools: key — git is installed conditionally as command[0] so the GitLab debian:stable-slim job needs no generator change) becomes the GHA job of the same name plus the generated .gitlab-ci.yml job. scripts/reserve-spec-id.sh, scripts/check-spec-id-reserved.sh, .crewrig/spec-id-carrier, ci/ci-capabilities.yml, docs/spec-pr-workflow.md → Reserving the spec id ✅ consumed by .claude/skills/spec-author/SKILL.md + .claude/agents/spec-author.md (built from artifacts/core/, row 3 core path) ✅ same contract via .gemini/skills/spec-author/SKILL.md + .gemini/agents/spec-author.md ✅ same contract via .github/skills/spec-author/SKILL.md + .github/agents/spec-author.md ✅ same contract via .agents/skills/spec-author/SKILL.md + .agents/agents/spec-author/AGENT.md
29 Tool-boundary git worktree guard (spec 0153) — pre-tool hook intercepting prohibited whole-tree git operations in shared ticket worktrees unless an exclusive claim is held hooks/worktree-git-guard.sh, hooks/*-transcript-hooks.json ✅ hooks/claude-transcript-hooks.json (PreToolUse on Bash) ✅ hooks/gemini-transcript-hooks.json (BeforeTool on transcript-git-guard) ✅ hooks/copilot-transcript-hooks.json (preToolUse hook) ✅ hooks/antigravity-transcript-hooks.json (PreToolUse grouped run_command guard — spec 0153 R4 / spec 0116 delta-03 R27; fires once per tool step, on run_command only)
30 Reviewer seat contract (spec 0166) — every review pass is attributed to one reviewer seat keyed <surface>/<ticket> on the specs / plan / review surfaces; each verdict carries a seat: line, the instantiation brief is references-only, and from a seat's second pass the mandatory reading is bounded to the change since the revision it last examined. Parity is uniform, not per-CLI: every obligation is dischargeable with the forge's own CLI and a self-contained brief, so nothing depends on a message bus, a surviving process, a shared memory service, or a CLI-exclusive primitive (spec 0166 R19). No CI check observes the contract — spec 0166 → Out of scope rules one out by name. docs/reviewer-seat.md (normative contract); sources artifacts/core/agents/pr-reviewer/AGENT.md, artifacts/core/skills/{pr-reviewer,architect,spec-author}/SKILL.md ✅ .claude/agents/pr-reviewer.md + .claude/skills/{pr-reviewer,architect,spec-author}/SKILL.md (built by scripts/build-components.sh --target claude; row 3 core path) ✅ .gemini/agents/pr-reviewer.md + .gemini/skills/{pr-reviewer,architect,spec-author}/SKILL.md (built by scripts/build-components.sh --target gemini; row 3 core path) ✅ .github/agents/pr-reviewer.md + .github/skills/{pr-reviewer,architect,spec-author}/SKILL.md (built by scripts/build-components.sh --target copilot; row 3 core path) ✅ .agents/agents/pr-reviewer/AGENT.md + .agents/skills/{pr-reviewer,architect,spec-author}/SKILL.md (built by scripts/build-components.sh --target antigravity; row 3 core path)
31 e2e probe A — Copilot BYOK model-routing (spec 0194 R8-R11) — differential probe for the upstream defect tracked as github/copilot-cli#4437: does a model: hint in a repository agent declaration break subagent routing on a bring-your-own-key session? Copilot-only by construction (applies_to = ["copilot"]) — no cross-CLI parity is claimed. Skips (records no verdict) unless tests/e2e/local.toml declares a backed byok_provider/byok_model pair. Verdict published as a comment on issue #1103 and cited from row 4's gap note below. tests/e2e/scenarios/05-copilot-model-routing/, tests/e2e/lib/probe_a_resolve.sh, scripts/e2e/publish-probe-verdict.sh ❌ N/A — probe is Copilot-only by design ❌ N/A — probe is Copilot-only by design ✅ host-orchestrated three-leg differential probe (control / hint-efficacy / model-bearing) against the effective [cli.copilot] configuration; lands declarations at the workspace-relative .claude/agents/probe-router/AGENT.md surface; verdict vocabulary BUG-PRESENT|BUG-ABSENT|INDETERMINATE ❌ N/A — probe is Copilot-only by design
32 e2e probe B — agent-surface consumption (spec 0194 R12-R15) — per CLI, per per-file-layout determination of whether .claude/agents/ repository declarations (nested <n>/AGENT.md and flat <n>.md) are actually consumed. Covers copilot and claude (R12 "at least"; neither Gemini CLI nor Antigravity CLI is a consumer of epic #1100's compiled-layout seam, so out of scope). Each cell carries an in-cell session-liveness baseline plus, for copilot, one cross-cell control at the only documented user-level layout (~/.copilot/agents/<n>.md). Outcomes published as a comment on issue #1103 and cited from row 4's [GAP-confirmation] note and epic #1100. tests/e2e/scenarios/06-agent-surface-consumption/, tests/e2e/lib/probe_b_resolve.sh, tests/e2e/lib/copilot_ephemeral_home.sh ✅ two cells (nested, flat), same mechanism as Copilot — no user-level Claude agent layout was documented in docs/ or artifacts/ when probe B ran, so its cells record control: in-cell-liveness-baseline-only (weaker attribution); spec 0201 has since named ~/.claude/agents/ in this table's own row 4 and in docs/agent-profile-migration.md, and the recorded control value stands as the record of that run rather than a still-true absence. Probe B run 20260902T154627Z-0d89 (issue #1103, comment 5512320726): both cells recorded indeterminate (the workstation Claude credential had expired), so this run establishes nothing either way about whether Claude Code consumes the flat .claude/agents/<name>.md form (spec 0201); a re-run after re-authentication is needed to ground it. ❌ N/A — not a consumer of epic #1100's compiled-layout seam (out of scope) ✅ three cells: nested and flat (recorded, workspace-relative .claude/agents/) plus a cross-cell control at ~/.copilot/agents/<n>.md via an ephemeral-home substitution (never the persisted credential bundle); outcome vocabulary consumed|not-consumed|indeterminate; a not-consumed nested outcome is recorded against row 4's [GAP-confirmation] and against specs/0143-copilot-subagent-model-fallback.md's ## Intent, without resolving the contradiction here. Probe B run 20260902T154627Z-0d89 (issue #1103, comment 5512320726): all three cells recorded consumed — .claude/agents/ nested, .claude/agents/ flat, and the documented ~/.copilot/agents/<n>.md surface. Consequence for spec 0201 (epic #1100 seam (g)): flattening the compiled Claude Code output does not remove Copilot's consumption of the .claude/agents/ surface under either per-file layout — the requirement 8 shared-read guard of specs/0143-copilot-subagent-model-fallback.delta-01.md remains necessary regardless. ❌ N/A — not a consumer of epic #1100's compiled-layout seam (out of scope)
33 Per-agent model-mapping declaration surface (spec 0197) — the per-CLI mapping that translates a spec 0195 capability profile into a target's native fields and prose; read since spec 0198 landed the build-time resolution (row 34), and populated against a real declaring source since spec 0200 migrated the 22 core agent sources onto metadata.model: (docs/agent-profile-migration.md). model-mappings/<target>.yml, docs/model-mapping-format.md, scripts/check-model-mappings.sh ✅ frontmatter model and effort; the model item alone is withheld by the shared-read guard (spec 0143 delta-01 R8, scripts/lib/model-resolve.sh's _resolve_item) while its state is withheld, with the declination stated on the guidance surface instead — effort reaches frontmatter regardless of the guard's state ✅ frontmatter model/temperature/max_turns; reasoning ⚠️ out-of-band in settings.json modelConfigs.overrides only, no guidance surface ⚠️ zero offerings, unconfigured by design — the served model lineup is deployment knowledge the core layer does not have (spec 0197 R41) ⚠️ guidance surface only — no per-agent frontmatter surface exists on this target (spec 0197 Premise 1)
34 Build-time resolution of a declared profile against the mapping in force (spec 0198) — realized against the 22 core agent sources migrated onto metadata.model: by spec 0200 (docs/agent-profile-migration.md); a resolution never fails a build (spec 0197 R15), degrading a mapping's unreadable cells and a target's unserved axes to drop records on the build's diagnostic stream instead. Since spec 0200, each migrated source's compiled output consequently carries: guidance prose naming the resolved alias on Claude Code and Antigravity CLI, a model: field on Gemini CLI, and no change on GitHub Copilot CLI (beyond the metadata.provenance.version line every version bump reaches). scripts/lib/model-resolve.sh, scripts/build-components.sh build_agents (resolve_agent), scripts/check-agent-profiles.sh ✅ writes effort: (frontmatter), and model: only while the guard's state is directed; while withheld, the model item is directed onto the compiled description instead (row 33's guard) ✅ writes model:/temperature:/max_turns: (frontmatter) — re-probed live for Gemini CLI 0.46.0 (issue #1116, spec 0198 R22) confirming those three keys are accepted; withholds reasoning and every tuning knob the mapping's out-of-band surface alone names ⚠️ withholds every item — model-mappings/copilot.yml declares zero offerings (row 33), so every declared axis and knob drops unsupported-on-cli ✅ writes the model item onto the compiled description alone (no frontmatter surface exists on this target, row 33); withholds every tuning knob the same way
35 Organization-level override of the mapping in force (spec 0199) — an org-owned model-mappings/<target>.org.yml beside each core mapping, merged into it inside mapping_in_force before row 34's resolution runs. Upstream ships all four silent (no offering, no surface, no guard state), so a fresh adopter's compiled agent outputs are byte-for-byte unaffected. Because a populated override legitimately changes what a compiled agent output contains, the four compiled agent output trees carry the sync policy regenerable (not strict) — an upstream synchronization restores and reports a diverged agent output rather than aborting on it; the compiled skill and command trees reach no mapping and stay strict. scripts/lib/model-resolve.sh (mapping_in_force), scripts/check-model-mappings.sh (pass 3), scripts/sync-from-upstream.sh, scripts/check-core-paths.sh, docs/org-model-mapping-override.md ✅ .claude/agents/ — regenerable ✅ .gemini/agents/ — regenerable ✅ .github/agents/ — regenerable (.github/skills/ stays strict, no mapping reaches it) ✅ .agents/agents/ — regenerable

Cross-reference — CI-engine × capability axis (spec 0047, ADR-0012). The CI-engine × CI-capability axis is a different axis from the CLI × feature matrix above and is self-documented by its own source of truth: ci/ci-capabilities.yml, normatively described by docs/ci-reference-format.md. This row is a voluntary cross-reference for discoverability, not a triggered CLI-matrix obligation — ci/ does not appear on the CLI Matrix Maintenance literal trigger list. Consult the reference, not this matrix, when adding or mapping a CI engine.

GitLab pipeline generator (spec 0048). The GitLab pipeline is derived from that reference, not hand-authored: scripts/build-ci.sh reads ci/ci-capabilities.yml and emits one GitLab job per portable capability into .gitlab-ci.yml at the repo root (job key == capability id; requires: → image/before_script/GIT_DEPTH; env: → variables:; command: → script:). Engine-specific capabilities are skipped with no placeholder (spec 0048 R4); the GitHub Actions workflows under .github/workflows/ are not generated (spec 0048 R5). scripts/build-ci.sh --check runs in this repo's own CI (.github/workflows/build.yml job gitlab-ci-check, itself a contract-described portable capability) and fails on any drift between the committed .gitlab-ci.yml and a fresh derivation. scripts/build-ci.sh is a scripts/build-*.sh script and therefore on the CLI Matrix Maintenance trigger surface; this entry satisfies that obligation.

CI agentic surface (spec 0050). The agentic layer over the CI machinery — the new gitlab-ci knowledge skill (artifacts/core/skills/gitlab-ci/, with a seven-file references/ corpus and two offline check scripts), the ci-configurator agent generalised to be engine-aware (it hand-authors a GitHub Actions workflow but derives a GitLab pipeline by composing scripts/build-ci.sh), and the new ci-parity agent (artifacts/core/agents/ci-parity/, which operates scripts/check-ci-parity.sh and reconciles drift) — rides the existing generic component-build mechanisms with full three-CLI parity: the skill ships through the generic skill-build (row 3, including propagate_skill_resources for its references/ and scripts/), and both agents ship through the generic agent-build (row 4). No new CLI mechanism row is created — these are ordinary artifacts/core/** components that scripts/build-components.sh compiles symmetrically to .claude/, .gemini/, and .github/. The agentic surface only composes the spec-0048 generator (build-ci.sh) and the spec-0049 harness (check-ci-parity.sh); it introduces no CLI-specific integration point. This paragraph discharges the artifacts/** CLI-matrix-update obligation for spec 0050.

CI execution speed (spec 0147). The former monolithic check-components capability (one job, ~85 sequential steps) is decomposed into focused, parallel, changeset-gated capabilities in ci/ci-capabilities.yml, each with its own paths: trigger filter and a changeset-gated: true marker. On both engines a job runs only when a changed file matches its paths: set (GitLab changes:; GitHub Actions dorny/paths-filter at the job level). Heavy groups (e.g. e2e) declare a manual trigger so they are when: manual / workflow_dispatch on non-main runs while staying enforced on push to main. Cached groups declare a cache: block (the key-derivation need: which files + env vars feed the key); scripts/build-ci.sh derives the GitLab cache:key:files mechanism and scripts/check-ci-parity.sh asserts semantic key agreement with the GitHub Actions hashFiles(...) inputs. A portable changeset-coverage capability with no paths: filter runs scripts/ci-changeset-coverage.sh on every change as the fail-safe (spec 0147 R5): when a changed file is not covered by the union of the focused paths: sets it executes the full check suite; otherwise it is a fast no-op. This paragraph discharges the ci/ CLI-matrix-update obligation for spec 0147.

test-wiring fine-grained cache (spec 0157). The test-wiring capability adds a cache-guard: false field and a fine-grained, content-addressed per-suite cache inside scripts/check-test-strays.sh. cache-guard: false tells scripts/build-ci.sh to emit the strays command bare (no coarse ci-cache-guard.sh wrapper), because the coarse wrapper's single engine-level key is incompatible with the per-suite verdict cache: keyed on the suites it invalidates the engine cache on every suite change (defeating the cache), and keyed on lib+script it skips the whole command on a suite-only change (skipping the re-validation the fine-grained cache must perform). The engine-level cache: block (GitLab cache:key:files / GitHub Actions actions/cache hashFiles(...)) is still emitted and persists .ci-cache/ across runs, so the per-suite verdicts survive; the fine-grained cache is the sole gate. The capability also declares requires.history-depth: full so the base ref resolves for the changeset short-circuit. This paragraph discharges the ci/ CLI-matrix-update obligation for spec 0157.

Parity gaps

The following features exist for one CLI but not the other. Each gap is a candidate follow-up issue. Do not fix them in this ticket.

  • [GAP] Copilot CLI model-mapping offerings (row 33) — model-mappings/copilot.yml declares zero offerings by design: the model lineup actually served to GitHub Copilot CLI is deployment knowledge (a GitHub-routed deployment and a bring-your-own-key deployment reach disjoint sets), not something the core layer can name without reproducing the silent routing failure of github/copilot-cli#4437. Empirical reproduction: probe run 20260902T153919Z-0aa9 of issue #1103, verdict BUG-PRESENT, plus the unknown BYOK lineup recorded in the live-verification research report (issue #1111 comment 5523492666). Spec 0197 R42 states the condition a later delta must meet to lift this.

  • [GAP] Antigravity CLI per-agent model frontmatter surface (row 33) — no per-agent model: field exists on this target; model-mappings/antigravity.yml states its declination on the guidance surface instead. Empirical reproduction: agy --help, agy models, ~/.gemini/antigravity-cli/settings.json, and ten on-disk documentation files read in full, none naming a per-agent model field (spec 0197 Premise 1).

  • [GAP] CI workflow — present for Claude Code (.github/workflows/claude.yml), missing for Gemini CLI.

  • [GAP, CLOSED — spec 0183, issue #1008] Plugin/extension build script — was present for Claude Code (scripts/build-claude-plugin.sh), missing for Gemini CLI. scripts/build-extension.sh --target gemini (spec 0173) is the render-side symmetric counterpart (row 13); scripts/release-package-extension.sh (spec 0183) is the RELEASE-artifact-packaging counterpart, since Claude's own release path never needed one — its plugin is installed via the marketplace mechanism (row 14), not a published archive. Left recorded here as history rather than deleted outright: the extensions-are-consumed-in-place framing was the reason no path existed, and this row is where a future reader would look for it.

  • [GAP] In-repo workspace settings file — present for Claude Code (.claude/settings.json), missing for Gemini CLI (settings live only in ~/.gemini/settings.json).

  • [GAP] .gitignore carve-outs — Claude-specific entries (dist-claude-plugin/, .claude/settings.local.json, CLAUDE.local.md) have no Gemini equivalents (e.g., a .gemini/settings.local.json analog).

  • [GAP] Top-level entry point file — CLAUDE.md lives at the repo root; GEMINI.md only appears inside extensions (no repo-root GEMINI.md).

  • [GAP] Slash commands — Copilot CLI has no user-definable slash-command file format. CrewRig commands compile as user-invocable skills under .github/skills/ (same workaround Claude uses). Same trade-off, documented and accepted.

  • [GAP] Project-dir env var — Copilot does not export a $COPILOT_PROJECT_DIR. hooks/mempalace-transcript.sh extracts the workspace path from the hook stdin JSON payload and falls back to $PWD.

  • [GAP-confirmation] Repo-level Copilot agent layout — the public Copilot reference only documents the user-level ~/.copilot/agents/ directory. CrewRig adopts .github/agents/<name>.md by parallelism with the skills layout; implementers must verify loading behavior and downgrade to a hard [GAP] if the path is rejected. Probe B run 2026-09-02 (spec 0194 R12-R15, issue #1103, comment 5507202742): all five cells (copilot nested/flat/cross-cell-control, claude nested/flat) recorded indeterminate — neither CLI session established (unrelated environment issues, not a defect in the probe), so no cell reached consumed or not-consumed. This gap stays a [GAP-confirmation]; a re-run with working sessions for both CLIs is needed to resolve it, and to ground epic #1100's compiled-layout seam.

  • [GAP-confirmation] Copilot agent provenance carrier — CrewRig currently emits the same metadata.provenance YAML block in .github/agents/<name>.md that it emits for Claude. Copilot CLI's tolerance for non-standard frontmatter keys is undocumented; if Copilot rejects the block (as Gemini CLI 0.42.0 did, see issue #54), migrate Copilot agents to the same HTML-comment carrier used by Gemini. Follow-up ticket to be filed.

  • [GAP-soft] e2e pillar 02 (cross-tool memory) for Copilot CLI — the crewrig/e2e-copilot:latest image does not embed the mempalace CLI (docker/e2e/copilot.Dockerfile installs only Node, gh, and @github/copilot; verified via docker run --rm crewrig/e2e-copilot:latest which mempalace → exit 1) and no MemPalace MCP wiring ships in artifacts/community/ for Copilot. The scenario excludes Copilot via applies_to = ["claude","gemini"] in tests/e2e/defaults.toml. Re-evaluate when either the Copilot image gains the mempalace binary or a MemPalace MCP integration appears for Copilot.

  • [GAP-soft] e2e pillar 04 (harness loop) full interactive leg — the scenario currently uses the simulated path (direct drawer write to wing=harness-friction, curator-style search) for all three CLIs. The fully interactive path that drives the harness-report skill end-to-end through the CLI's slash-command surface is not yet supported non-interactively from any CLI. The simulated path is parity-equivalent for v1: it writes the same drawer shape the skill would write and the curator reads the same lane regardless of the writer's identity.

  • [GAP-soft] e2e auth flow for Copilot CLI (spec 0194 R1-R7, issue #1103) — auth-claude.sh and auth-gemini.sh persist a credential via a fresh, interactive container login; auth-copilot.sh instead passes through the workstation's own already-authenticated ~/.copilot/config.json (copied to ~/.crewrig-e2e/copilot/config.json, mounted read-write at /home/agent/.copilot), alongside the pre-existing env-var PAT ($COPILOT_GITHUB_TOKEN / $GH_TOKEN) and Ollama Cloud keypair paths. e2e_auth_ready (scripts/e2e/lib/auth-common.sh) recognizes all four. Empirical evidence for why no interactive container login exists is captured in docs/adr/0002-e2e-auth-flow.md Decision 4: the crewrig/e2e-copilot:latest image lacks libsecret, so the upstream desktop-keychain path is unavailable, and the documented plaintext-fallback device-flow adds login friction without scenario-coverage gain over the passthrough. Re-evaluate if a scenario ever needs a credential distinct from the developer's own workstation session.

  • [GAP] Copilot CLI subagent model routing on BYOK/Ollama (row 4) — on a BYOK provider (e.g. Ollama), spawning a repo-level Copilot custom agent as a task-tool subagent silently fails (Agent completed but produced no response. / Model '...' not found on provider ... (HTTP 404)) when the session model differs from the subagent's default. Root cause: the default subagent-model routing (flight copilot_cli_gpt_5_4_for_subagents in ~/.copilot/config.json) and hardcoded model: hints in repo agent files. Solution: scripts/build-components.sh omits hardcoded model: frontmatter from the two compiled agent surfaces GitHub Copilot CLI reads, .claude/agents/ and .github/agents/ — since spec 0200 migrated the 22 core agent sources onto metadata.model:, 22 .gemini/agents/*.md files carry a model: field (docs/agent-profile-migration.md), so the claim no longer holds tree-wide and is narrowed to the surfaces this hazard actually reaches. Workaround if using custom agent files: pass model: <session-model> explicitly to the task tool call. Tracked upstream at github/copilot-cli#4437. Remove this note when the upstream fix lands. Probe A run 2026-09-02 (spec 0194 R8-R11, issue #1103, comment 5506957758): INDETERMINATE / surface-not-read-or-session-broken — Copilot CLI 1.0.82 via the Ollama Cloud BYOK path; the run did not discriminate because the BYOK session itself never established (an unrelated Ollama-client write-path regression against the read-only credential mount, not the routing defect this note tracks). This note stays in force per R11 (an INDETERMINATE verdict neither discharges nor re-dates it); a re-run on a working BYOK session (or a COPILOT_GITHUB_TOKEN-backed provider) is needed for a discriminating verdict. Probe B run 20260902T154627Z-0d89 (spec 0201, epic #1100 seam (g), issue #1103, comment 5512320726): all three Copilot cells recorded consumed for the .claude/agents/ surface — nested, flat, and the documented ~/.copilot/agents/ surface — confirming the shared-read hazard this note tracks holds on both compiled per-file layouts. Flattening the compiled Claude Code output to .claude/agents/<name>.md (spec 0201) therefore does not retire this note or the requirement 8 shared-read guard of specs/0143-copilot-subagent-model-fallback.delta-01.md; neither the Copilot emission nor that guard changes, and both remain necessary regardless of per-file layout.

  • [GAP] MemPalace layered wake-up (row 7g) — MemPalace's MemoryStack.wake_up() (L0–L3) is exposed by the Python library and the mempalace CLI (wake-up) but is absent from the agent-facing MCP tool surface on the installed 3.6.0 (pinned >=3.6.0,<3.7; re-verified 2026-07-20 against the live 36-tool surface, issue #566). Unlike the gaps above, this is not a cross-CLI asymmetry: all four CLIs share the one MCP surface (row 7d), so the gap is uniform across them. The MCP-only access guard in artifacts/core/rules/60-tools.md forbids reaching it ad-hoc, and the deterministic session-start sweep (that file's Wake-up budget) is its bounded substitute. Candidate follow-up: an upstream request to expose the wake-up through MCP (consent-gated, out of scope for spec 0069).

<!-- Antigravity CLI gaps (spec 0058) -->

  • [GAP] Settings template in config/ (row 6) — Antigravity CLI writes settings directly to ~/.gemini/antigravity-cli/settings.json via setup-antigravity-interactive.sh; no config/antigravity/ repo template is defined. Same pattern as Gemini CLI's row-7 gap.
  • [GAP] In-repo workspace settings file (row 7) — Antigravity CLI reads from ~/.gemini/antigravity-cli/settings.json only; no .agents/settings.json analog written to the project tree.
  • [GAP] Project-dir env var (row 9) — no $ANTIGRAVITY_PROJECT_DIR is injected. Since spec 0116 the project path is read from the hook payload's workspacePaths[0], with the git root and $PWD as fallbacks. Evidence: a captured hook payload carries workspacePaths; no env var appears in any handler's environment.
  • [GAP] Workspace-level hook discovery (row 8) — the vendor documents a per-workspace customization root (.agents/ at the repository root) alongside the global one, but a manifest registered there did not fire. Evidence: two agy --print runs from a git init-ed workspace containing .agents/hooks.json, with a handler that appends its stdin to a log, produced no log entry; the identical manifest at ~/.gemini/config/hooks.json fired PreInvocation, PostInvocation and Stop on the next run. setup-antigravity-interactive.sh therefore deploys only to the global root. Whether the per-workspace root fails in every configuration or only in the probed one (headless, single prompt, freshly initialized repository) is not established.
  • [GAP-soft] Duplicate agent names across discovery roots (rows 4, 10) — Antigravity CLI does not de-duplicate an agent name discovered at two roots at once: agy agents lists it twice. Neither copy shadows the other and the CLI does not refuse the pair. Evidence: one probe sentinel installed simultaneously at ~/.gemini/config/agents/ and ~/.gemini/agents/, listed twice while its unique-named siblings at the same two roots were each listed once (agy 1.1.11, 2026-08-11, two runs; docs/runbooks/antigravity-discovery-probe.md). The live instance is harness-curator, which after spec 0123 exists at ~/.gemini/agents/harness-curator.md from the Gemini setup and ~/.gemini/config/agents/harness-curator/AGENT.md from the Antigravity setup. Both are the same crewrig component rendered for two CLIs, so either resolution yields the same agent and the only effect is a doubled line in the listing. Recorded rather than worked around: renaming a shipped component to dodge a vendor listing artifact would cost more than the artifact does.
  • [GAP-confirmation] Two agent-discovery instruments, and they disagree (rows 4, 10) — agy agents enumerates agents from all three candidate roots (~/.gemini/antigravity-cli/agents/, ~/.gemini/config/agents/, ~/.gemini/agents/), in both the flat and the directory shape. A forced-choice prompt in the same session, against the same sentinels, reports only those under ~/.gemini/config/ as available. The two are not contradicting each other about facts — the subcommand reads a registry, the prompt measures what the assistant can reach in a session — but a discovery claim resting on agy agents alone rests on the more permissive of the two. The first probe on issue #761 concluded from agy agents alone that the application-data root "is a discovery location"; that conclusion is instrument-dependent and is corrected here. ~/.gemini/config/agents/ is the only root that satisfies both instruments, which is why spec 0123 R2's single documented root is the right target under either reading. Evidence: agy 1.1.11, 2026-08-11, two reproducible runs; docs/runbooks/antigravity-discovery-probe.md.
  • [GAP] Extension skill/agent loader (row 5c) — Antigravity CLI has no in-repo extension skill/agent loader. scripts/install-antigravity-extension.sh builds a plugin package and calls agy plugin install — the plugin is consumed as a package, not via per-skill directory links.
  • [GAP] CI workflow (row 18) — agy --print is confirmed non-interactive (exit 0; output: Pong! How can I help you today?), but no published GitHub Action (analogous to anthropics/claude-code-action) exists for Antigravity CLI. Evidence: gh api /repos/antigravity-labs/agy-code-action → HTTP 404; gh api /users/antigravity-labs → HTTP 404; gh search repos "agy-action" — no relevant results.
  • [GAP-soft] e2e dedicated-account auth flow (row 21) — no scripts/e2e/auth-antigravity.sh or crewrig/e2e-antigravity:latest Docker image has been defined. Adding Antigravity CLI e2e coverage is explicitly out of scope per spec 0058.
  • [GAP-soft] e2e pillar 01 — layered context (row 22) — no Antigravity CLI e2e container or scenario variant exists. Out of scope per spec 0058.
  • [GAP-soft] e2e pillar 02 — cross-tool memory (row 23) — no Antigravity CLI e2e container or scenario variant exists. Out of scope per spec 0058.
  • [GAP-soft] e2e pillar 03 — skill build (row 24) — no Antigravity CLI e2e container or scenario variant exists. Out of scope per spec 0058.
  • [GAP-soft] e2e pillar 04 — harness loop (row 25) — no Antigravity CLI e2e container or scenario variant exists. Out of scope per spec 0058.

<!-- Validation backend gaps (spec 0080) -->

  • [GAP-confirmation] Internal-backend structured interactive prompt on Gemini / Copilot / Antigravity CLI (row 27b) — AskUserQuestion is confirmed on Claude Code; a structured interactive-prompt equivalent that avoids a bare prose question is unverified on the other three CLIs. Evidence to be gathered at DEV; downgrade to a hard [GAP] if none exists. The gate SHALL NOT silently degrade to a prose question (spec 0037).
  • [GAP] Illustration image display for the plannotator backend (row 27) — the illustration option (spec 0080 R14) renders only when a browser image surface is available. Unlike the per-CLI gaps above, this is not a cross-CLI asymmetry: Plannotator renders the artifact as HTML in a browser it launches itself, so image display is a function of browser availability, uniform across all four CLIs (spec 0080 OQ4). On the internal backend there is no image surface on any CLI, and illustration is silently ignored (R14). Recorded as a documented uniform limitation — like row 7g (MemPalace wake-up) — so it is not mistaken for a per-CLI parity gap.

<!-- Solo-merge classifier gap (spec 0102) -->

  • [GAP-confirmation] Solo-maintainer self-merge classifier block — Claude Code's auto-mode permission classifier can deny a gh pr merge invocation when the merging identity equals the pull-request author (confirmed on Claude Code: the denial text itself names "the Claude Code auto mode classifier"). Whether Gemini CLI, GitHub Copilot CLI, or Antigravity CLI exposes an equivalent classifier is unverified and out of scope for spec 0102 — no negative is asserted about those three. This is an external CLI runtime behaviour with no CrewRig source-of-truth artifact behind it, hence a Parity gaps note rather than a feature-matrix row. The reactive workaround (self-retry once → hand the merge to the user, no laundering to a sibling agent) and the two optional prevention paths are documented in AGENTS.md → Branching Strategy → On Claude Code CLI — solo-maintainer self-merge block.

Adding a new CLI

Every checklist item below is governed by the Symmetric-script rule and Gap-acceptance evidence rule in AGENTS.md → CLI Matrix Maintenance. Unchecking an item requires the evidence the gap rule demands; it is not a free-form choice.

Use this checklist when onboarding a new CLI. Each item maps to a row in the feature matrix above — leaving one unchecked produces a parity gap.

  1. [ ] Create the workspace config root (./.<cli>/).
  2. [ ] Add a top-level agent-context entry point that re-exports AGENTS.md (analogous to CLAUDE.md).
  3. [ ] Extend scripts/build-components.sh with a --target <cli> branch and emit skills via the tier-routed output root (output_root_for_tier): core → .<cli>/skills/<name>/SKILL.md in the project tree; non-core → dist/<tier>/.<cli>/skills/<name>/SKILL.md (or the native form). The build is tier-agnostic — do not hardcode tier names.
  4. [ ] Extend build_agents in the same script to emit agents in the CLI's native layout.
  5. [ ] Extend build_commands in the same script to emit commands in the CLI's native form.
  6. [ ] Add a per-CLI settings template under config/<cli>/.
  7. [ ] Decide whether a checked-in workspace settings file at .<cli>/settings.json is required, and add it if so.
  8. [ ] Provide a hook-integration manifest under hooks/<cli>-transcript-hooks.json covering the CLI's lifecycle events.
  9. [ ] Document the CLI's project-dir env var convention and wire hooks/mempalace-transcript.sh to read it.
  10. [ ] Add an interactive setup script scripts/setup-<cli>-interactive.sh, including the artifact-install phase that reads dist/<tier>/ and routes by tier scope: library auto-installed to the user home, community/org each behind an opt-in prompt (ADR-0011, spec 0019). core is never installed here — it ships in the project tree.
  11. [ ] Add a transcript backfill script scripts/import-<cli>-history.sh.
  12. [ ] Add or extend a component-management script for the CLI.
  13. [ ] If the CLI requires a build step before install, add scripts/build-<cli>-plugin.sh (or equivalent).
  14. [ ] If a separate install step is required, add scripts/install-<cli>-plugin.sh.
  15. [ ] Register --target <cli> invocations in Taskfile.yml, plus setup-, import-, install-, link-, and build-components- task entries.
  16. [ ] Add a per-CLI extension manifest convention to extension-skeleton/ and update extensions/core/hello-world/ with an example.
  17. [ ] Add a CI workflow .github/workflows/<cli>.yml if the CLI offers an automation entry point.
  18. [ ] Update prose references in README.md, AGENTS.md, CONTRIBUTING.md, and DEVELOPMENT.md.
  19. [ ] Add .gitignore carve-outs for the CLI's local settings and any generated dist-<cli>-* directories.
  20. [ ] Append a new column to this matrix and re-check every row.