allura-cowork

Allura Cowork

Use this skill to coordinate Claude Code and Codex without pretending they are the same runtime.

Core Contract

Intent -> Project context -> Allura Brain -> Route -> Work -> Validate -> Receipt -> Handoff or close

Allura Brain is the shared memory layer. Claude and Codex are execution surfaces. Project-local systems, such as Team RAM or Durham, apply only when the repo declares them or Ronin explicitly routes through them.

Required Cowork Header

Before planning or handoff, state:

Cowork active.
Runtime: <Claude|Codex|OpenCode|OpenClaw|other>
Allura: <memory searched/not available + query or reason>
Project overlay: <Team RAM|Durham|TALON|IRIS|none|unknown>
Route: <who owns implementation + who reviews>
Validation: <commands, checks, screenshots, or evidence path>
Handoff target: <Claude|Codex|none>

Runtime Honesty

• Say "Claude perspective" or "Codex perspective" only as a lens unless that runtime actually ran.
• Never claim a subagent, tool, MCP call, test, file edit, or handoff happened unless it actually happened.
• If a handoff is only written instructions, label it Handoff packet.
• If validation did not run, say exactly what is unvalidated.

Allura Brain Rules

• Search before planning when prior work, people, dates, decisions, preferences, todos, or Allura governance are involved.
• Default scope: group_id=allura-system.
• Use the active actor as user_id, for example troy-curator, team-ram, or codex-cowork.
• Write after substantive work with files changed, validation, outcome, and remaining risk.
• Raw episodic memory is evidence. It is not canonical promotion.

Approval Boundaries

Explicit approval is required before:

• runtime or database mutation
• MCP config mutation
• cron mutation
• live hook installation or enforcement
• RuVix runtime enforcement changes
• semantic memory promotion
• Notion sync
• Done/Approved status moves
• public or external sends

Handoff Packet

Use schemas/handoff.schema.json for machine-shaped packets. For chat, use:

## Allura Cowork Handoff

From: <runtime/persona>
To: <runtime/persona>
Goal: <one sentence>
Context searched: <queries + memory ids or unavailable reason>
Project overlay: <Team RAM|Durham|TALON|IRIS|none|unknown>
Files touched/read: <paths>
Decisions: <bullets>
Open risks: <bullets>
Validation run: <commands/results>
Next action: <single concrete step>
Memory status: <written/not written + id if available>
Approval needed: <yes/no + why>

Routing Defaults

• Codex: repository edits, validation loops, code review, practical implementation, commits.
• Claude: synthesis, long-form docs, UX/design review, plugin packaging, critique, prose-heavy planning.
• Team RAM: project-local implementation when the repo declares RAM or Ronin routes there.
• Durham: brand, copy, product feel, design canon, and final brand-sensitive approval.
• TALON: deploy, runtime, API readiness, test gates, runbooks, and guardrails.

User-Facing Promise

Say: "This plugin reduces hallucinated claims by requiring memory search, evidence, validation, and receipts before work is called done."

Do not say: "This plugin prevents hallucinations."