Lisa Coding Agent Parity

Lisa ↔ Coding-Agent Parity (RESEARCH)

Lisa is distributed for multiple coding agents: Claude Code is the production path, with Codex, Cursor (cursor-agent), Antigravity (agy), and GitHub Copilot (copilot) as additional install targets. When any agent ships a new capability, or when a Lisa feature exists for one agent ahead of the others, the picture of "what's portable, what isn't, where Lisa polyfills, what's deliberately agent-only" drifts. This skill is the repeatable research protocol for refreshing that picture.

This skill is research-only. Its output is a four-part artifact — a universal feature catalog, a support matrix, a plugin-distributability matrix, and per-cell Lisa polyfill designs. It does not write installer code, build per-agent plugin variants, or modify Lisa's source tree. Implementation work is the sibling skill [[lisa-coding-agent-parity-implement]], which consumes this artifact as its primary input.

The transport for querying agents non-interactively — adapter slots, dry-run, write guards, parallel dispatch — lives in [[harness-parity-council]] (.agents/skills/harness-parity-council/). Use that skill whenever this protocol says "query the agents." Web research uses WebSearch + WebFetch.

Core principle: empirical, not documentary

Do not trust documentation OR assumptions alone. Both lie. Examples from the canonical run that produced this skill:

• Codex's [[reference-codex-hooks-capabilities]] memory said plugin-bundled hooks silently don't run; Codex 0.125.0 actually supports plugin hooks for ten events including SubagentStart, SubagentStop, and PostCompact. The memory was stale.
• agy's agy plugin --help advertises agy plugin import claude as an import source; at runtime that command returns "No claude extensions found" and never imports any of the registered Claude marketplace plugins. Docs implied feasibility; runtime falsified it.
• agy's plugin validator accepts the full Claude 7-event hooks.json schema and agy plugin install correctly writes the file into the installed plugin tree; none of the registered hooks fire during a real agy -p session. Validator and runtime are separate code paths.

Every claim in the four-part artifact must be backed by one of:

1. Documentation read — a publicly-published official doc, GitHub issue, or canonical community example. Tag [VERIFIED-DOC] with the URL.
2. Source-read — inspecting the installed binary, its --help, config home, vendored source, or generated schemas. Tag [VERIFIED].
3. Runtime capture — running a real CLI invocation and observing actual behavior (capture hook stdin, install a probe plugin, observe what loads). Tag [VERIFIED-BY-RUN] with the command and a short observed-output snippet.

Treat [UNKNOWN] and [VERIFIED-DOC]-only claims as preliminary — they should be upgraded to [VERIFIED-BY-RUN] before they're load-bearing for a polyfill design in Step 4.

The coding agents in scope

The skill scopes to whatever five agents Lisa currently distributes to. As of this skill's last revision: Claude Code, Codex, Cursor (cursor-agent), Antigravity (agy), GitHub Copilot. No agent is the baseline. Capabilities flow in any direction — Codex may have a hook event Claude lacks; agy may have a first-class plugin manager Codex lacks; Cursor may have a Tab-specific hook event nobody else has. Every such asymmetry is a first-class row in the catalog.

Claude remains the lead orchestrator (per [[harness-parity-council]]) — it drives the research and writes the final synthesis — but Claude is not the reference implementation. No agent's claim about its own internals is overridden without an empirical counter-example, regardless of which agent makes the claim.

When the fleet changes (new CLI added, existing CLI dropped), update this table as the first action of a research pass.

Prerequisites

• All five CLIs installed and authed on this machine. claude and codex are the historical baseline; cursor-agent, agy, and copilot were added on the coding-agent-parity worktree.
• harness-parity-council available under .agents/skills/ — run node .agents/skills/harness-parity-council/runtime-adapters.mjs to confirm resolved command slots, base flags, and help/version probes before starting.
• Each CLI's non-interactive entry point understood: claude -p / codex exec / cursor-agent -p / agy -p / copilot -p. Resume flags: --continue / resume --last / --connect / --resume / -c.
• Filesystem access is available so probes can install throwaway plugins under isolated config homes (CODEX_HOME, etc.) and observe behavior.

CLI invocation matrix (model + effort)

When this skill drives a CLI to investigate or empirically verify behavior, use the model and reasoning-effort settings below. Do not pick a smaller model or a lower effort silently — research verdicts depend on the other side reasoning carefully about its own internals.

Notes:

• agy exposes no --model/effort flag. Set Gemini 3.5 Flash, medium via the Antigravity app's Model Selection; the CLI honors it. agy is also prone to misinterpreting structured prompts (treats long briefs as meta-questions about CLI flags). Use short, direct prompts; fall back to direct source-read when agy goes off-script.
• cursor-agent has no thinking variant for Composer 2.5; for deep-reasoning questions route to claude or codex.
• For all five, pass the non-interactive print flag when scripting and bound long runs with a timeout (agy in particular can hang).

The protocol

Research proceeds in four steps. Each step's output is appended to a single artifact file at /tmp/parity-research.md (or a Lisa-local temp path) so the four steps form one continuous document. Earlier steps' outputs are inputs to later steps.

Step 1 — Universal feature aggregate

Build the union catalog of every feature each in-scope agent exposes. Cast a wide net: this is the foundation that Steps 2-4 depend on, and omitting features here propagates errors down the chain.

Method — combine three sources for each agent:

1. Web/doc research (WebSearch + WebFetch): official docs, GitHub issues, release notes, community feature catalogs. Sources to canvass per agent:

   • Claude Code: code.claude.com/docs, anthropic blog, GitHub marketplace plugin examples.
   • Codex: developers.openai.com/codex/, GitHub releases for codex-cli, core-plugins/src/manifest.rs schema.
   • Cursor: cursor.com/docs, the bundled index.js (loader source), community marketplace examples.
   • Antigravity (agy): antigravity.google/docs/*, Google Cloud community blog (Medium), agy plugin --help subcommand surface.
   • Copilot: docs.github.com/en/copilot/*, the cli-plugin-reference page, generated TypeScript declarations under ~/Library/Caches/copilot/pkg/.../rpc.d.ts.

2. CLI self-query: <cli> --help and every visible subcommand's --help. Walk the help tree fully.

3. Council brief (via [[harness-parity-council]]): send each agent the same brief asking it to enumerate every feature it exposes — not just plugin-related ones. Include tool surfaces, model selection, output formats, session resumption, sandbox modes, ACP, MCP transports, workspaces/worktrees, keybindings, auth methods, image/voice input, IDE integrations, update mechanism, telemetry, etc. Tag the prompt with the symmetric framing from §1z.

Categories to canvass (use as a checklist — not exhaustive; add as discovered):

• Authoring surfaces: skills, sub-agents, slash commands, hooks, MCP servers, rules/instructions, memory files, custom prompts, workflows, role personas, LSP servers, app connectors.
• Distribution & lifecycle: plugin manifest, plugin marketplace, plugin install/enable/disable/validate, version pinning, update mechanism, session-only loading, marketplace registry.
• Runtime modes: interactive vs print (-p), agent client protocol (ACP), plan mode, ask mode, headless, sandbox modes, trust mode, worktree mode, autopilot mode.
• Inputs & I/O: model selection, reasoning effort, output formats (text/json/stream-json), streaming, image attachments, voice input, context windows, prompt caching, multi-line input, custom keybindings.
• Tools: file edit/write, shell, web fetch, web search, code search, file read, glob, MCP tool exposure, tool allowlist / permission gates, allowed-paths, allowed-URLs, sub-agent invocation.
• Auth & identity: API key, OAuth, login subcommand, account identity, organization policy gates, telemetry/analytics defaults.
• Integrations: IDE integration (VS Code, JetBrains, etc.), shell integration, Git integration, commit attribution, scheduling / cron, notifications.
• Hook events: every event name each agent exposes (PreToolUse, PostToolUse, SessionStart, SessionEnd, SubagentStart, SubagentStop, UserPromptSubmit, Stop, PreCompact, PostCompact, Notification, PermissionRequest, beforeShellExecution, beforeTabFileRead, …).

Output of Step 1 — a flat list of every distinct feature found across the union of all five agents, each with a short description and the agent(s) that surfaced it. Do not classify per-agent support yet — that's Step 2. Aim for breadth; minimum ~60 distinct features for a five-agent fleet. Save to the artifact under heading ## Step 1 — Universal feature catalog.

Step 2 — Support matrix

For each feature in the Step 1 catalog, fill a row showing which agents natively support that feature. No assumptions about plugin distributability yet — that's Step 3. Step 2 is purely "does this agent have this feature, in any form (config, CLI flag, plugin-bundled, anywhere)."

Cell forms:

• ✅ — supported natively. Name the exact mechanism in parentheses.
• ✅* — supported with a caveat (only in interactive mode; only at user level; only when feature-flag X enabled). State the caveat.
• ❌ — explicitly not supported (verified by source-read or runtime capture, not just absence from docs).
• ⚠ unverified — claim from docs only, runtime not probed.

Empirical bar: every ✅ cell needs a [VERIFIED-BY-RUN] tag with a command + observed output, or a [VERIFIED-DOC] tag with URL. [UNKNOWN] cells are allowed but flagged for Step 4 follow-up.

Output of Step 2 — markdown table, rows = Step 1 features, columns = agents. Save under ## Step 2 — Support matrix.

Step 3 — Plugin-distributability matrix

For each ✅ or ✅* cell from Step 2, answer the separate question: can this feature be packaged inside that agent's plugin format, such that a user gets the feature by installing the plugin? This is a different matrix from Step 2 because many features (CLI flags, OS-level integrations, auth methods) are agent-supported but not plugin-distributable.

Cell forms:

• 📦 yes — bundleable in the agent's plugin manifest. Name the manifest field or directory convention (e.g. skills/<n>/SKILL.md, manifest hooks block).
• 📦* yes-with-caveat — bundleable but the runtime behavior differs from the standalone case (e.g. agy's hooks/hooks.json validates and installs but doesn't fire in -p mode — [VERIFIED-BY-RUN]).
• ❌ no — the feature exists on this agent but cannot be carried in its plugin format (e.g. MCP on agy: lives at user-config or project-config scope, NOT in plugin).
• n/a — feature isn't supported on this agent at all (cell was ❌ in Step 2).

Empirical bar: for any 📦 yes cell, an empirical probe should have either installed a plugin carrying that feature and observed it loading, or read a public canonical plugin that does so. 📦* yes-with-caveat cells MUST have [VERIFIED-BY-RUN] evidence — the caveat is the whole point.

Output of Step 3 — same row/column layout as Step 2, different cell semantics. Save under ## Step 3 — Plugin-distributability matrix.

Step 4 — Lisa polyfill designs

For each (feature, agent) cell that is one of:

• ❌ in Step 2 (agent doesn't support the feature at all), or
• ❌ no or 📦* yes-with-caveat in Step 3 (agent supports it but it's not distributable via plugin, or it's distributable but runtime behavior is broken),

design how Lisa would polyfill the gap. This is the design step, not the implementation step — output is prose + sketch, not code.

Per-cell deliverable:

1. Cell ID: <feature> × <agent>
2. Gap classification:
   • Type A: feature absent on this agent entirely.
   • Type B: feature present but not plugin-distributable.
   • Type C: feature plugin-distributable but runtime-broken in headless mode.
3. Polyfill strategy — one of:
   • Translate: emit the feature in a different shape the agent does support (e.g. Codex commands → Lisa skills with lisa- prefix).
   • Wrap: invoke a non-plugin Lisa installer at lisa apply time to write the feature into a non-plugin location (e.g. agy MCP → write ~/.gemini/config/mcp_config.json).
   • Bake: fold the feature's content into a different surface the agent does auto-load (e.g. agy rules → bake rules/eager/*.md bodies into AGENTS.md).
   • Skip: document as agent-only, do not polyfill (e.g. Copilot LSP servers — Lisa doesn't ship LSPs).
   • Block: declare blocked, route to a separate work item to either pressure upstream or rethink Lisa's reliance on the feature.
4. Sketch — 3-5 sentences naming the file path Lisa would change in the implementation phase, the data shape, and any caveats. NOT code.
5. Empirical verification plan — the probe lisa-coding-agent-parity-implement will run to prove the polyfill works.

Polyfill collision check — before finalizing each design, check whether the polyfill would collide with a native mechanism on a different agent loading the same plugin (e.g. Lisa's plugin-bundled SessionStart rules hook on Claude would double-fire on Cursor, which auto-loads rules/ natively). Per-agent plugin artifacts (the Pattern B decision from the plugins-surface pass) are the canonical fix.

Output of Step 4 — one subsection per gap cell. Save under ## Step 4 — Lisa polyfill designs.

Output artifact

The artifact is a single markdown file at /tmp/parity-research.md (or a Lisa-local temp path) with four top-level sections matching Steps 1-4 plus a header summarizing the run (date, agent fleet, CLI versions probed, evidence sources canvassed).

The artifact is the entire output of this skill. It is the input to [[lisa-coding-agent-parity-implement]]. No installer code, no per-agent plugin variants, no commits, no PRs — those belong to the implementation skill.

Where each agent's plugin artifacts live (reference)

A reference for Step 1 / Step 3 work — where to look in Lisa's source tree to understand each agent's existing surface, and where the agent's installed plugins land on disk.

• Claude — Plugin manifest plugins/src/<plugin>/.claude-plugin/plugin.json with hooks block, auto-discovered skills/, agents/, commands/, rules/ (NOT auto-loaded — Lisa polyfills via SessionStart hook). Marketplace at .claude-plugin/marketplace.json at repo root, 8+ plugins listed. Installed: ~/.claude/plugins/. Plugin-root env var: ${CLAUDE_PLUGIN_ROOT}.
• Codex — .codex-plugin/plugin.json pointer (skills + MCP); hooks via .codex/hooks.json per-project, OR — as of codex-cli 0.125.0 — via plugin-bundled hooks/hooks.json or inline manifest hooks field. Hook events: ten total, including SubagentStart, SubagentStop, PreCompact, PostCompact. Sub-agents are config-level [agents.<role>], NOT a plugin component. Installed: ~/.codex/plugins/cache/<marketplace>/<plugin>/<version>/. Marketplace: ~/.codex/config.toml [marketplaces.<n>]. Lisa installs via src/codex/*-installer.ts (skills, agents, hooks, AGENTS.md, marketplace, settings, manifest tracking).
• Cursor — Reads BOTH .cursor-plugin/plugin.json AND .claude-plugin/plugin.json natively (loader's discoverComponents covers skills, agents, commands, rules). Hooks auto-normalized from Claude event names. Installed plugins land under ~/.claude/plugins/cache/ (Cursor shares Claude's tree). Session-only via --plugin-dir <path>. Cursor-unique events: beforeShellExecution, beforeTabFileRead.
• Antigravity (agy) — Plugin manifest is bare plugin.json at plugin root (NOT .claude-plugin/ or .gemini-plugin/). Components: skills/, agents/, commands/ (auto-converted to skills!), hooks/hooks.json. MCP is NOT a plugin component — lives at user ~/.gemini/config/mcp_config.json or project .agents/mcp_config.json. Auto-loads AGENTS.md since v1.20.3. Plugin-bundled hooks validate and install but DO NOT FIRE in -p mode — Lisa cannot rely on agy plugin hooks for headless rule injection (use AGENTS.md bake-in instead). Installed: ~/.gemini/config/plugins/<plugin-name>/. Plugin manager: agy plugin install/uninstall/enable/disable/validate/link/import.
• Copilot — Manifest lookup: plugin.json → .plugin/plugin.json → .github/plugin/plugin.json → .claude-plugin/plugin.json. Components: skills/<n>/SKILL.md (default) or manifest skills path; agents/<n>.agent.md (NOTE the .agent.md extension — adapter needed vs Claude's <n>.md); manifest hooks; manifest mcpServers; LSP servers via lspServers (Copilot-unique). MCP three layers: User ~/.copilot/mcp-config.json, Workspace .mcp.json, Plugin. Installed: ~/.copilot/installed-plugins/<marketplace>/<plugin>/ or _direct/. CodySwannGT/lisa is already a registered marketplace.

Known capability references (memory)

Update these after a research pass surfaces new capabilities. They are the institutional memory that lets the next pass start from current knowledge rather than redoing CLI source-reads from scratch.

• [[reference-codex-hooks-capabilities]] — Codex hook events (currently stale; 0.125.0 expanded the event list — update after next implementation).
• [[reference-codex-plugin-skill-loading]] — .codex-plugin/ skills pointer.
• [[reference-codex-commit-attribution]] — Codex commit-attribution gates.
• [[reference-codex-http-mcp-plugin-shape]] — Codex accepts Claude's {type,url} MCP shape.
• [[reference-codex-automations]] — Codex automations / scheduling.
• [[reference-codex-cli-collaboration]] — codex exec non-interactive patterns.
• [[reference-lisa-hook-delivery]] — how Lisa hooks reach projects.

Per-agent capability memory for Cursor, agy, Copilot does not yet exist — write the first one whenever this skill produces a [VERIFIED-BY-RUN] load-bearing finding about that agent.

Worked example — the plugins-surface research pass (2026-05-28)

The canonical reference run. Topic: Lisa's plugin surface across all five agents.

• Step 1 surfaced ~30 features focused on plugin packaging (skills, sub-agents, slash commands, hooks, MCP, rules, memory, settings, marketplace, version pinning, enable/disable, etc.).
• Step 2 filled the support matrix with a mix of [VERIFIED-BY-RUN] (agy via direct probes + plugin install; Cursor via --plugin-dir smoke test; Copilot via marketplace browse) and [VERIFIED-DOC] (codex hooks event list).
• Step 3 flipped multiple cells that looked positive in Step 2: agy MCP (supported, but ❌ not plugin-distributable — file lives outside the plugin), agy hooks (supported AND plugin-distributable per the validator, but 📦* — runtime DOES NOT FIRE in -p mode).
• Step 4 produced polyfill designs including: Codex commands → Lisa-prefixed skills (Translate); agy MCP → mcp-installer.ts writing user-config (Wrap); agy rules → bake into AGENTS.md (Bake); Copilot LSP servers → not portable (Skip).

The full artifact is at /tmp/plugins-plan.md (the research-output file from that run). Sections 1a, 1z, 5, 5.5, 5.6 of that artifact carry the canonical examples for the matrix-cell formats above.

That pass also surfaced that the previously stale [[reference-codex-hooks-capabilities]] memory needed updating — Codex 0.125.0 expanded the hook event list.

Definition of done

• Step 1 catalog covers a clear majority of each agent's documented feature surface — ~60+ distinct features for a five-agent fleet, with no obvious omissions in any of the canvass categories above.
• Step 2 support matrix has every cell labeled (✅ / ✅* / ❌ / ⚠ unverified), and every load-bearing claim is tagged with [VERIFIED-DOC] (URL), [VERIFIED] (source-read command), or [VERIFIED-BY-RUN] (CLI invocation + observed output).
• Step 3 plugin-distributability matrix has every cell labeled (📦 / 📦* / ❌ / n/a). The split between Step 2 support and Step 3 distributability is explicit — no conflation.
• Step 4 has one polyfill design per gap cell, each with: gap type, polyfill strategy, file/data sketch, collision check, and verification plan for lisa-coding-agent-parity-implement.
• The artifact at /tmp/parity-research.md (or chosen path) is a single coherent markdown document.
• Capability memory entries are updated for any agent whose surface changed materially during the research pass.

Out of scope explicitly: no installer code, no plugin variant generation, no commits, no PRs, no lisa apply modifications. Those belong to [[lisa-coding-agent-parity-implement]].

Handoff to implementation

When this skill completes, the artifact at /tmp/parity-research.md (or chosen path) is the input contract for [[lisa-coding-agent-parity-implement]]. The implementation skill:

1. Reads the artifact.
2. Picks a scope (one polyfill, one full agent, one feature surface, etc. — depends on the implementation skill's own gating).
3. Authors the installers / plugin variants / build-script changes.
4. Runs the per-cell empirical verifications spec'd in Step 4.
5. Lands the work.

This skill never executes any of those steps. If a user invokes this skill and asks for "implementation" or "apply this," reject the request and point at the sibling skill.