gspec

Platform capabilities

gspec runs across many AI coding harnesses — the tools it installs into, like Claude Code or Cursor — but they are not equivalent. Every harness gets the full spec-authoring surface; where they differ is how strongly they can enforce your specs and whether they support the autonomous build and the learning loop.

Two paths matter when reading this table. The interactive path is hands-on /gspec-* use in the harness. The build path is the autonomous, headless implementation run. A harness can be strong on one and absent on the other.

Today, Claude Code is the only full-fidelity target. The cells below reflect what ships today, not what is planned — expand any harness for the specifics.

Full support
Partial / advisory
Not supported
gspec capability support by AI coding harness. Each cell is one of: full support, partial or advisory support, or not supported.
Harness Spec authoring Skills, agents, and commands for writing and maintaining specs. Autonomous build Headless, spec-driven implementation runs where an engine adapter exists. Competitive research Web-backed competitor research — /gspec-research interactively, --research on the build path. Enforcement — interactive Deterministic spec-integrity floors during hands-on /gspec-* use. Enforcement — build The same floors enforced during autonomous build runs. Learning loop Agents record memories to .gspec/memory/pending/; /gspec-memorize commits them to skills.
Claude Code Full support Full support Full support Full support Full support Full support
Codex Full support Full support Full support Partial / advisory Full support Partial / advisory
Cursor Full support Not supported Partial / advisory Partial / advisory Not supported Partial / advisory
Open Code Full support Not supported Partial / advisory Partial / advisory Not supported Partial / advisory
Pi Full support Full support Full support Partial / advisory Full support Partial / advisory
Antigravity Full support Not supported Partial / advisory Partial / advisory Not supported Not supported

Platform capabilities verified July 2026. These harnesses change monthly — re-verify before relying on a specific cell.

Harness by harness

What each harness can and can’t do, with the reasoning behind every partial. Expand a harness for its platform mechanics and the specifics.

Claude Code

The reference target. Full-fidelity support on both the interactive and build paths.

Spec authoring: Full support Autonomous build: Full support Competitive research: Full support Enforcement — interactive: Full support Enforcement — build: Full support Learning loop: Full support
  • Full support

    Spec authoring. Native skills, agents, and commands.

  • Full support

    Autonomous build. Wired headless engine.

  • Full support

    Competitive research. Interactive /gspec-research plus the build’s --research stage, with native WebSearch/WebFetch for the research agents.

  • Full support

    Enforcement — interactive. Hard, deterministic blocks via Pre/PostToolUse hooks on every file write.

  • Full support

    Enforcement — build. The same floors hold on the autonomous build path.

  • Full support

    Learning loop. Recording, the address-tag guard and /gspec-memorize all run — the only target where the tag is hook-enforced.

File-write hook
Yes — Pre/PostToolUse on Write/Edit
Tool interception
Any tool
Native per-agent memory
Yes — memory: silos
Headless engine
Yes — wired

Codex

OpenAI’s coding agent. Strong build-path parity, but hooks fire on the Bash tool only.

Spec authoring: Full support Autonomous build: Full support Competitive research: Full support Enforcement — interactive: Partial / advisory Enforcement — build: Full support Learning loop: Partial / advisory
  • Full support

    Spec authoring. Native skills, agents, and commands.

  • Full support

    Autonomous build. Wired engine; the runtime driver runs QA gates and verify.sh.

  • Full support

    Competitive research. Interactive command plus the build’s --research stage; the engine opens sandbox network access for the research agents.

  • Partial / advisory

    Enforcement — interactive. Gap: hooks see Bash only, never a file write, so floors ship advisory (inlined). Hard blocks need an MCP write-mediator.

  • Full support

    Enforcement — build. Enforced by the runtime driver, independent of any hook.

  • Partial / advisory

    Learning loop. Recording and /gspec-memorize work; the address tag is a convention here, not a hook (Codex hooks fire on Bash only). Planned.

File-write hook
No — hooks fire on Bash only
Tool interception
Bash only (PreToolUse deny, PostToolUse audit + additionalContext)
Native per-agent memory
No — AGENTS.md instructions
Headless engine
Yes — wired

Cursor

Closest to native enforcement parity via file-edit hooks. No wired build engine yet.

Spec authoring: Full support Autonomous build: Not supported Competitive research: Partial / advisory Enforcement — interactive: Partial / advisory Enforcement — build: Not supported Learning loop: Partial / advisory
  • Full support

    Spec authoring. Native sub-agents, skills, and commands.

  • Not supported

    Autonomous build. Gap: no build-engine adapter wired. cursor-agent -p exists, so adding one is straightforward.

  • Partial / advisory

    Competitive research. Gap: /gspec-research works interactively; build-path research (--research) waits on a build adapter.

  • Partial / advisory

    Enforcement — interactive. Gap: floors are advisory (inlined) today, but onPreEdit / onPostEdit observe edits, so hard-block guards can port. Planned.

  • Not supported

    Enforcement — build. Gap: waits on a build adapter.

  • Partial / advisory

    Learning loop. Interactive recording and /gspec-memorize work; the address tag is unenforced, and build-path recording waits on the build adapter.

File-write hook
Yes — onPreEdit / onPostEdit
Tool interception
Editor-event hooks
Native per-agent memory
No — rules / AGENTS.md
Headless engine
Yes — cursor-agent -p (adapter not yet wired)

Open Code

Open-source agent with a complete plugin system. No wired build engine yet.

Spec authoring: Full support Autonomous build: Not supported Competitive research: Partial / advisory Enforcement — interactive: Partial / advisory Enforcement — build: Not supported Learning loop: Partial / advisory
  • Full support

    Spec authoring. Native sub-agents, skills, and commands.

  • Not supported

    Autonomous build. Gap: no adapter wired. opencode run exists.

  • Partial / advisory

    Competitive research. Gap: /gspec-research works interactively; build-path research (--research) waits on a build adapter.

  • Partial / advisory

    Enforcement — interactive. Gap: advisory (inlined) today. A gspec plugin could intercept tool execution across 25+ lifecycle events to hard-block writes. Planned.

  • Not supported

    Enforcement — build. Gap: waits on a build adapter.

  • Partial / advisory

    Learning loop. Interactive recording and /gspec-memorize work; a plugin-backed address-tag guard is planned alongside the adapter.

File-write hook
Yes — plugin tool-execution interception
Tool interception
Any tool, via 25+ lifecycle events
Native per-agent memory
No — community "mem" plugins
Headless engine
Yes — opencode run (adapter not yet wired)

Pi

Extension-complete agent with a wired build engine. The fullest non-Claude parity today.

Spec authoring: Full support Autonomous build: Full support Competitive research: Full support Enforcement — interactive: Partial / advisory Enforcement — build: Full support Learning loop: Partial / advisory
  • Full support

    Spec authoring. Native sub-agents, skills, and commands.

  • Full support

    Autonomous build. Wired engine via the pi-subagents extension; the runtime driver gates every run.

  • Full support

    Competitive research. Interactive command plus the build’s --research stage; if the research agents’ web tools stall on print-mode approval, pass --pi-permission-level.

  • Partial / advisory

    Enforcement — interactive. Gap: advisory (inlined) today, but the tool_call hook can block or modify any call (Claude-equivalent), so a gspec extension would hard-enforce. Planned.

  • Full support

    Enforcement — build. Enforced by the runtime driver on every build run.

  • Partial / advisory

    Learning loop. Recording and /gspec-memorize work on both paths; the address tag is a convention here, not a hook. Planned.

File-write hook
Yes — tool_call block / modify
Tool interception
Any tool, via the extension API
Native per-agent memory
No — episodic-memory extensions exist
Headless engine
Yes — wired, via pi-subagents

Antigravity

Workflow-based agent with native JSON lifecycle hooks. No wired build engine yet.

Spec authoring: Full support Autonomous build: Not supported Competitive research: Partial / advisory Enforcement — interactive: Partial / advisory Enforcement — build: Not supported Learning loop: Not supported
  • Full support

    Spec authoring. Skills plus one self-contained workflow per capability (sub-agents spawn at runtime, so no agent files).

  • Not supported

    Autonomous build. Gap: no adapter wired. agy -p exists.

  • Partial / advisory

    Competitive research. Gap: research runs interactively in single-file mode; build-path research (--research) waits on a build adapter.

  • Partial / advisory

    Enforcement — interactive. Gap: advisory (inlined in the workflows) today. JSON lifecycle hooks (after-file-edit) run headless, so guards port like Cursor. Planned.

  • Not supported

    Enforcement — build. Gap: waits on a build adapter.

  • Not supported

    Learning loop. Gap: workflows only, with no agent files to record from. Recording lands with the agent adapter.

File-write hook
Yes — JSON hooks incl. after-file-edit
Tool interception
Before-tool / after-edit / session-start
Native per-agent memory
No
Headless engine
Yes — agy -p (adapter not yet wired)