Brigade in Cortex Code: yes — CoCo ships a Claude-Code-compatible plugin/skill/subagent model, which unblocks P3/P4
The question
"Can Brigade skills/plugins actually run inside Snowflake Cortex Code, and what does the packaging support look like?" Context: this cell was flagged unevaluated in the 2026-07-03 Hex deployment matrix and blocks CAF deployment profiles P3 (Snowflake-native) / P4 (Cowork sales org) before the next phData DSA client deployment conversation.
What we already know (from the vault)
- The [[2026-07-03-hex-deployment-matrix]] explicitly flags this exact cell: Cortex Code as a packaging option is "⚠️ unevaluated: CoCo is real as a runtime (Gary built the KB4 demo in it) but skills/plugin packaging support is unverified — do not assume." That flag is what this brief resolves.
- The matrix splits packaging (Claude Code plugin = LIVE via skill-agent-brigade; Cowork org-admin skills; Cortex Code = unevaluated) from runtime. It defaulted P3's brigade runtime to an SPCS container (implicitly hosting a headless Claude Code session) and P4 to Cowork/cloud scheduled sessions — Cortex Code was never selected as the runtime, only listed as an unverified candidate.
- The Brigade already ships as a LIVE Claude Code plugin: SKILL.md station skills (spec-author → test-author → code-author → critic), station subagents, a deterministic Gate-A lint, and a five-command public surface (BRIGADE-INTERFACE.md, PR #9). See [[2026-06-30-anthropic-skill-guide-vs-brigade-comparison]].
- Snowflake's Cortex stack (Cortex Agents GA Nov 2025, Cortex Search, Cortex Analyst, SPCS) is documented in the vault as the client-governed data-gravity target: [[2026-04-28-snowflake-enterprise-knowledge-management]].
- The predicted hurdle for a Snowflake-native brigade runtime was C1 — SPCS egress: a brigade-in-SPCS must reach the Anthropic API via an external-access integration + network rule + secret, gated on client-admin approval.
What the web says
- Cortex Code CLI plugins bundle skills, subagents, slash commands, hooks, and MCP servers under a single manifest — the identical component set to a Claude Code plugin (Cortex Code CLI plugins docs).
- The plugin manifest is accepted at
.cortex-plugin/plugin.jsonor.claude-plugin/plugin.json— CoCo reads the Claude Code plugin manifest directly (docs, same page). - Skills are
SKILL.mdfiles in./skills/; subagents are markdown in./agents/; discovery paths include~/.claude/skills/and~/.claude/agents/— Claude Code directories are read as-is. Subagents run in their own context with amodelfield (e.g.claude-sonnet-4-5) and tool list (Cortex Code CLI extensibility docs). - Distribution: marketplace name, GitHub shorthand (
owner/repo), git URLs, or shipped as part of a Snowflake connection profile for org-wide, client-governed rollout. Pre-ship validation viacortex plugin validate(plugins docs). - Hooks fire on PreToolUse/PostToolUse/PermissionRequest and can be shell (deterministic) or prompt type — a direct home for the Brigade's Gate-A deterministic lint (extensibility docs).
- Cortex Code also interoperates outward: MCP + Agent Client Protocol (ACP) support, a Claude Code plugin (preview), a VS Code extension (private preview), and a Python/TypeScript Agent SDK. Snowflake-Labs'
subagent-cortex-codelets Claude Code/Cursor/Codex route Snowflake ops to CoCo headless (Snowflake blog: One Governed Agent). - Surfaces: CLI, Desktop, Snowsight (in-browser), and Cloud Agents (private preview) with "dedicated and isolated compute inside Snowflake" where "the same governance framework that governs your data governs your agent."
Convergences and contradictions
- Convergence (resolves the flag): the vault's "do not assume" caution is now answerable — YES. CoCo's packaging model isn't merely analogous to Claude Code; it is compatible by design (accepts
.claude-plugin/plugin.json, reads~/.claude/skills/and~/.claude/agents/). The Brigade's LIVE Claude Code plugin is close to drop-in. - Correction to the matrix's P3 shape: the matrix modeled P3 as an SPCS container hosting a Claude Code session. Reality is simpler — CoCo can be both the packaging and the runtime, so P3 need not smuggle Claude Code into a container.
- Unresolved: model routing / egress. Cloud Agents run "inside Snowflake" governance, which plausibly dissolves C1 (Snowflake-hosted Claude via the Cortex/Anthropic partnership, no external egress), but no doc I fetched states the routing explicitly. Treat C1-dissolution as promising-but-unconfirmed.
Synthesis for RDCO
Brigade CAN run inside Snowflake Cortex Code, and the packaging support is strong — effectively Claude Code plugin compatibility. The 2026-07-03 matrix cell should flip from "⚠️ unevaluated — do not assume" to "supported, with light adaptation." Cortex Code CLI packages exactly what the Brigade is made of (SKILL.md skills, ./agents/ subagents, slash commands, hooks, MCP servers) under one manifest, and it reads the Claude Code manifest path and skill/agent directories directly. The Brigade's existing LIVE Claude Code plugin is the starting artifact, not a rewrite target.
This unblocks P3 and P4 and, more importantly, simplifies them. P3 no longer has to be "an SPCS container hosting a headless Claude Code session"; the client-governed path becomes ship the Brigade plugin via a Snowflake connection profile (org-admin-published, governed distribution) and run it on a CoCo surface (CLI in SPCS, or Cloud Agents once out of private preview). The Brigade's public five-command interface maps to CoCo slash commands, Gate-A lint maps to a PreToolUse hook plus cortex plugin validate, and the rail/cellar ports still bind to Snowflake stage + table as already specced. P4's Cowork story is unchanged but gains a second option: CoCo's MCP/ACP interop means the sales-seat surface could call the Brigade rather than re-hosting it.
The one load-bearing caveat before a client conversation is C1 (egress) and model routing. If CoCo Cloud Agents serve Claude models from within Snowflake's boundary, then the single hardest objection in the matrix — external Anthropic API egress requiring client-admin approval — largely evaporates, which materially strengthens the P3 pitch to Snowflake AEs. That's the highest-value thing to confirm. The honest framing for the DSA conversation: "the Brigade packages as a native Cortex Code plugin (Claude Code-compatible manifest), distributable through a governed connection profile; the remaining spike is confirming in-boundary model routing and remapping any station subagent tools/model-IDs to CoCo's surface."
Net: this is not a non-starter and does not force a different profile. It converts P3/P4 from "gated on unverified packaging" to "gated on a small confirmation spike (egress + a portability smoke test)." Recommend a one-session spike: install the existing skill-agent-brigade plugin into a Cortex Code CLI against the phData sandbox, run cortex plugin validate, and fire one end-to-end station pass. That single test converts every "should" in this brief to "does."
Open follow-ups
- Model routing / egress: does running the Brigade inside CoCo (esp. Cloud Agents) use Snowflake-hosted Claude with no external Anthropic egress, dissolving C1? Highest-value confirmation.
- True drop-in vs light-adapt: does the LIVE skill-agent-brigade Claude Code plugin run unmodified in CoCo, or do station subagents need tool-name / model-ID remapping to CoCo's tool surface? (Run the smoke test.)
- Surface selection for P3/P4: CoCo CLI-in-SPCS vs Cloud Agents (private preview) vs Snowsight — which is client-deployable today given preview gating?
- Gate-A fidelity: can
cortex plugin validate+ a PreToolUse hook reproduce the Brigade's deterministic lint exactly, or is a shim needed? - Governed distribution + secrets: does shipping via a Snowflake connection profile satisfy client-governed rollout AND the no-secrets-on-disk rule (MCP creds via OAuth / env expansion, never in the manifest)?
- KB4 July 15 demo: ride P1 (safe) or stage a P3/CoCo slice for the Snowflake AE story? Decide against confirmed preview status.
Related
- [[2026-07-03-hex-deployment-matrix]]
- [[2026-07-01-caf-ecosystem-map-and-brigade-restructure-read]]
- [[2026-06-30-anthropic-skill-guide-vs-brigade-comparison]]
- [[2026-04-28-snowflake-enterprise-knowledge-management]]
- [[2026-06-14-caf-restructure-organizing-brief]]
Sources
- Vault:
~/rdco-vault/01-projects/phdata/2026-07-03-hex-deployment-matrix.md— [[2026-07-03-hex-deployment-matrix]] - Vault:
~/rdco-vault/01-projects/phdata/2026-07-01-caf-ecosystem-map-and-brigade-restructure-read.md— [[2026-07-01-caf-ecosystem-map-and-brigade-restructure-read]] - Vault:
~/rdco-vault/08-tooling/2026-06-30-anthropic-skill-guide-vs-brigade-comparison.md— [[2026-06-30-anthropic-skill-guide-vs-brigade-comparison]] - Vault:
~/rdco-vault/06-reference/research/2026-04-28-snowflake-enterprise-knowledge-management.md— [[2026-04-28-snowflake-enterprise-knowledge-management]] - Vault:
~/rdco-vault/01-projects/phdata/2026-06-14-caf-restructure-organizing-brief.md— [[2026-06-14-caf-restructure-organizing-brief]] - Web: Cortex Code CLI plugins — Snowflake docs
- Web: Cortex Code CLI extensibility — Snowflake docs
- Web: Cortex Code CLI MCP support — Snowflake docs
- Web: Cortex Code Expands: One Governed Agent for Your Entire Data Stack — Snowflake blog
- Web: Snowflake-Labs/subagent-cortex-code — GitHub