Skip to content

PLATFORM_SPEC.md — SubQ (Subquadratic Inc.) / "SubQ Code"

New to gald3r on this platform? Start at Use gald3r with SubQ Code for a short overview before the full breakdown below.

SubQ is a subquadratic long-context LLM from Subquadratic Inc. (SubQ 1M-Preview; Subquadratic Sparse Attention / SSA, ~O(n); 1M tokens production, up to ~12M research). Its developer-facing surface is an OpenAI-compatible Chat Completions API (docs.subq.ai). "SubQ Code" is NOT a standalone CLI or coding agent — the official product page (subq.ai/code) describes it as a plugin / "long-context layer for coding agents" that you "Plug into Claude Code, Codex, and Cursor" via a "one-line install", whose job is to "auto-redirect expensive model turns" for cheaper/faster whole-repo exploration.

Because SubQ Code runs inside a host agent, it exposes none of its own extension mechanisms: commands, rules, sub-agents, skills, hooks, and MCP are all properties of the host tool (Claude Code / Codex / Cursor), not of SubQ. As of June 2026 there is still no public standalone SubQ CLI, no published config-folder schema (.subq/), and no SubQ-native instruction-file convention. A gald3r install therefore targets the host platform, never SubQ itself.

Authoring path: UPDATE — supersedes the prior heavy-❓ spec (last_doc_scan: never, status ❓). That spec correctly refused to fabricate, but treated SubQ-native conventions as merely undocumented. The June-2026 official docs (subq.ai/code, docs.subq.ai) make clear they are not-applicable: SubQ Code is a plugin with no own extension surface, so the six gald3r mechanisms are downgraded from ❓ to ❌ (none). Verified 2026-06-02 (see Verification Evidence). This is the honest state — ❌ here means "the host owns it", not "broken".

Access status change. May 2026: API/SubQ Code/SubQ Search were waitlist / private-beta with no public config docs (prior spec). June 2026: the API, console (console.subq.ai, keys), playground (playground.subq.ai), and docs (docs.subq.ai) are publicly reachable — but the coding offering remains a plugin, not an agent with its own extensibility. So the six gald3r mechanisms are not-applicable to SubQ as a host.

1. Folder Hierarchy

Status: ❌ no SubQ-native config folder.

SubQ publishes no .subq/ config-folder schema and ships no standalone CLI that would read one. SubQ Code is a plugin loaded by a host agent, so the only config tree that matters is the host's (.claude/, .codex/, .cursor/). gald3r writes nothing SubQ-native.

<project-root>/
└── (no SubQ-native config tree)
    # SubQ Code is a plugin inside the host agent (Claude Code / Codex / Cursor).
    # gald3r targets the HOST's config tree (.claude/ | .codex/ | .cursor/), not .subq/.

gald3r writes: nothing SubQ-native. Install gald3r into the host platform instead. SubQ owns: the model API (docs.subq.ai), the console (keys), and the plugin's internal turn-redirection — none of which is a gald3r-writable surface. Source: https://subq.ai/code

2. AI Instruction File

Status: ❌ none-native (inherits the host's instruction file).

SubQ publishes no native instruction-file convention — there is no SUBQ.md and no .subq/ instruction schema. Because SubQ Code plugs into Claude Code / Codex / Cursor, it inherits whatever instruction file the host uses (e.g. AGENTS.md, or CLAUDE.md under Claude Code). gald3r context reaches SubQ today only indirectly, via the host tool's AGENTS.md / CLAUDE.md. There is no SubQ-specific instruction file to write. Source: https://docs.subq.ai/overview/

3. Agents Support — ❌ none

  • SubQ Code is itself a layer/plugin invoked by an existing coding agent; it is not a multi-agent framework and documents no sub-agent / agent-role / mode definitions. Any subagents are the host's (e.g. Claude Code .claude/agents/).
  • Evidence: subq.ai/code — "The long-context layer for coding agents. Plug into Claude Code, Codex, and Cursor."
  • Source: https://subq.ai/code

4. Skills Support — ❌ none

  • No SKILL.md / Agent-Skills / reusable-workflow-file mechanism is documented on any official SubQ page. The whole-repo-context value proposition explicitly de-emphasizes chunking/skill machinery, but no skills API exists. gald3r g-skl-*/SKILL.md are not ingested by SubQ; they load only via the host (e.g. Claude Code skills discovery).
  • Source: https://subq.ai/code

5. Commands / Workflows — ❌ none

  • The official SubQ Code page describes only a plugin that "Auto-redirects expensive model turns" via "One-line install"; no slash / custom-command system is documented. docs.subq.ai covers only the Chat Completions API. Any slash commands a user sees come from the host tool (Claude Code / Codex / Cursor), not SubQ.
  • Source: https://subq.ai/code

6. Hooks System — ❌ none

  • No lifecycle / event-hook system (session-start, pre-tool, pre-commit, file-watch, or a hooks.json equivalent) is documented. The only "automatic" behavior is internal turn-redirection inside the plugin, which is not a user-configurable script hook. gald3r's g-hk-*.py hooks do not run on SubQ; they must run via the host (e.g. Claude Code .claude/settings.json hooks) or via git core.hooksPath / manual invocation.
  • Source: https://subq.ai/code

7. Rules / Memory — ❌ none

  • No persistent-rules / memory / always-on-instructions mechanism is documented anywhere on subq.ai or docs.subq.ai. SubQ's pitch is that a 1M–12M token window loads the whole repo directly, so there is no rules/memory file convention; persistent instructions live in the host tool's own files (e.g. AGENTS.md / CLAUDE.md), not SubQ's. gald3r g-rl-* rules have no SubQ-native injection point.
  • Source: https://docs.subq.ai/overview/

8. MCP Support — ❌ none

  • docs.subq.ai documents an OpenAI-compatible Chat Completions API (Base URL, API key, /v1/chat/completions) with "Streaming + tool use" and "OpenAI-compatible endpoints" — an HTTP contract, NOT Model Context Protocol. No MCP client/server support is mentioned. MCP, if used, is provided by the host agent SubQ plugs into.
  • Source: https://docs.subq.ai/overview/

9. Plugin Caveat vs. Cursor Reference

Compared to the Cursor reference (g-skl-platform-cursor/PLATFORM_SPEC.md), SubQ has no native gald3r-relevant extension surface at all — not because it is undocumented, but because SubQ Code is a plugin, not an agent. Every gald3r mechanism (commands / rules / agents / skills / hooks / MCP) belongs to the host the plugin is installed into.

  1. What SubQ actually is — a subquadratic long-context LLM (SubQ 1M-Preview; SSA ~O(n); 1M production / ~12M research), sold as model API + a coding plugin, not a coding agent.
  2. Developer surface — OpenAI-compatible Chat Completions API only (docs.subq.ai; console.subq.ai for keys; playground.subq.ai to try). Marketed gains for the Code plugin: "~25% lower bill, 10x faster exploration".
  3. Output formats / workflows — streaming responses and OpenAI-style tool calling at the API level only; no agent-level workflow / output-format extensibility is exposed.
  4. Commands ❌ / Rules ❌ / Agents ❌ / Skills ❌ / Hooks ❌ / MCP ❌ — all none on SubQ itself; each is a host property (Claude Code / Codex / Cursor).
  5. Instruction file ❌ (none-native) — no SUBQ.md; inherits the host's AGENTS.md / CLAUDE.md.
  6. Install targets the host, not SubQ — to bring gald3r to a SubQ-accelerated workflow, install gald3r into the host (Claude Code / Codex / Cursor) and let SubQ Code accelerate it underneath.

Hook System

  • Type: none ❌
  • Config file: none — SubQ documents no hook wiring file (no hooks.json equivalent); the plugin exposes no user-configurable script hooks
  • Events available: none — no published lifecycle-event taxonomy; the only "automatic" behavior is internal turn-redirection, not a hook
  • Event payload format: n/a
  • OS limits: n/a (no hook runtime to constrain)
  • gald3r hook files: g-hk-*.py do not run on SubQ — wire them on the host (e.g. Claude Code .claude/settings.json) or via git core.hooksPath / manual invocation instead

Atypical Handling

  • SubQ Code is a plugin, not an agent. It runs inside Claude Code / Codex / Cursor and contributes long-context turn-redirection only — it owns no commands/rules/agents/skills/hooks/MCP surface.
  • No .subq/ config tree and no SUBQ.md. gald3r targets the host platform's config tree.
  • The public SubQ developer surface is the model API (OpenAI-compatible HTTP), not an agent SDK.

gald3r Integration Notes

  • To bring gald3r to a SubQ-accelerated session, install gald3r into the host (Claude Code / Codex / Cursor) — see g-skl-platform-claude, g-skl-platform-codex, g-skl-platform-cursor.
  • Do not assume any .ps1 hook, rule, skill, command, agent, or MCP wiring works on SubQ — all of it lives on the host.
  • Re-verify on the next @g-platform-scan-docs subq (crawl_max_age_days: 30) only to catch a future standalone SubQ CLI; today there is no SubQ-native config surface to crawl.

Capability Summary

Hooks Rules Skills Commands MCP Docs Fresh
❌ ❌ ❌ ❌ ❌ ✅

Legend: ✅ verified working · ⚠️ partial / Cursor-generic · ❌ not supported (host owns it) · ❓ untested.

Overall platform status: ❌ — instruction-only. SubQ Code is a confirmed real product (Subquadratic Inc.), but it is a plugin / long-context layer, not a standalone agent, so it exposes no native gald3r-writable extension surface. The six mechanisms are none because the host (Claude Code / Codex / Cursor) owns them — install gald3r into the host. ❌ here is the honest mapping of "not-applicable to SubQ as a host", not a fabricated negative.

Verification Evidence (docs crawl 2026-06-02, https://subq.ai + https://docs.subq.ai)

Capability How verified
What SubQ is subq.ai/introducing-subq + docs.subq.ai/overview — subquadratic long-context LLM (SSA ~O(n); 1M prod / ~12M research); developer surface is an OpenAI-compatible Chat Completions API
SubQ Code is a plugin subq.ai/code — "The long-context layer for coding agents. Plug into Claude Code, Codex, and Cursor"; "one-line install"; "auto-redirects expensive model turns" (~25% lower bill, 10x faster exploration)
Commands ❌ subq.ai/code — only a plugin + one-line install; no slash/custom-command system documented; commands come from the host
Rules ❌ docs.subq.ai/overview — no persistent-rules/memory/always-on convention; whole-repo context replaces it; persistent instructions live in host AGENTS.md/CLAUDE.md
Agents ❌ subq.ai/code — SubQ Code is a layer invoked by an existing agent; no sub-agent/role/mode definitions; not a multi-agent framework
Skills ❌ subq.ai/code — no SKILL.md / Agent-Skills / reusable-workflow mechanism documented
Hooks ❌ subq.ai/code — no lifecycle/event-hook system (no hooks.json equivalent); only internal turn-redirection, not a user hook
MCP ❌ docs.subq.ai/overview — OpenAI-compatible Chat Completions API (Base URL, key, /v1/chat/completions, streaming + tool use); HTTP contract, NOT MCP; no MCP client/server
Instruction file ❌ docs.subq.ai/overview + subq.ai/code — no native SUBQ.md / .subq/ schema; SubQ Code inherits the host's AGENTS.md/CLAUDE.md
Access status June 2026: API/console (console.subq.ai)/playground (playground.subq.ai)/docs (docs.subq.ai) publicly reachable; coding offering still a plugin (was waitlist/private-beta in May 2026)
Docs freshness last_doc_scan: 2026-06-02 — crawl of subq.ai, subq.ai/code, docs.subq.ai/overview, console.subq.ai, playground.subq.ai

Reference page — canonical source: g-skl-platform-subq/PLATFORM_SPEC.md in the engine repo, generated by scripts/generate_platform_support_docs.py. Pages marked curated are hand-assessed and not yet live-verified.