/security-flow

name: security-flow
description: "Workflow for authorized, evidence-preserving security review and remediation-task preparation."
tags: ["workflow", "security"]
baseSchema: docs/schemas/workflow.md

<security_flow>

<description_and_purpose>

Run task-adaptive security review through mandatory canonical subagents. End with sanitized findings and concise inputs for later user-invoked coding flows.

</description_and_purpose>

<workflow_phases>

<prerequisites phase="0" applies="ALL">

1. All Rosetta prep steps MUST be FULLY completed. 2. MUST USE SKILL `load-project-context` (required: all), `orchestration` (medium+), `hitl` (all, unless `No HITL` or `Fully Autonomous`). 3. MUST ALWAYS use todo tasks ledger, ASAP. Phases are sequential. Independent tasks can run in parallel. 4. MUST just-in-time load/execute/update each phase's: instructions, definitions, skills, state file; do not load/act IN ADVANCE. 5. Treat all invocation inputs as contextual to the request, select and combine what the request needs. 6. For full review, require every applicable, available, authorized activity and tool. 7. Maintain a task ledger and run phases JIT. 8. Every question and approval follows the loaded `hitl` skill. 9. Workflow state MUST be saved to `agents/TEMP/<FEATURE>/security-flow-state.md` file; every phase updates it before the next starts.

</prerequisites>

<subagent_policy required="true" inline_execution="prohibited">

• Orchestrator owns approvals, phase transitions, dispatch, aggregation, and handoff.
• Phase files are assigned-subagent-only; orchestrator MUST NOT load, read, summarize, or execute them.
• Every declared subagent is mandatory.
• Every subagent MUST USE SKILL `subagent-directives`.
• Subagents use tools required by their own assignment.
• `executor` is never a gateway for full agents.
• Reject incomplete phase contracts before advancing.
• If required subagent invocation is unavailable, stop and report the unmet prerequisite.

</subagent_policy>

<readiness phase="1" applies="ALL" subagent="executor" role="Bounded security readiness and filename-only secret-gate operator" subagent_required_model="Claude Haiku 4.5, GPT-5.6 Terra, Gemini 3.7 Flash, GPT-5.6 Luna" must-be-subagent>

• Purpose: Inventories limited target/tool metadata and runs a filename-only secret gate before any model ingests target content. Determines whether to continue, request approval for DEV/QA envs, or stop for high risk.
• Input: request; target/environment metadata only.
• Output: readiness result with limited inventories and gate state.
• INVOKE SUBAGENT `executor` to APPLY PHASE `security-flow-readiness.md` + inventory limited metadata/tools, run the filename-only secret gate, and return its gate state.
• Expect: limited target/tool inventory; `PASS|NEEDS-HITL|STOP-HIGH-RISK|STOP-SCANNER-UNUSABLE`.
• Control: advance only on PASS or approved DEV/QA envs; high-risk and scanner-unusable stops are non-overridable.

</readiness>

<authorize phase="2" applies="ALL" subagent="engineer" role="Enterprise security scope and authorization advisor" subagent_required_model="Claude Sonnet 5, GPT-5.6 Terra, Gemini 3.7 Flash" type="HITL" must-be-subagent>

• Purpose: Recommends an enterprise-safe run contract from readiness evidence and task intent. Identifies every material decision requiring user approval before security work begins.
• Input: readiness result; task intent; known policy.
• Output: recommended run contract and approval record.
• INVOKE SUBAGENT `engineer` to APPLY PHASE `security-flow-authorize.md` + recommend scope, environment, exclusions, activities, tool/data-flow decisions, bounds, and stop conditions.
• Expect: a complete run contract with every activity marked local read-only, separately gated, or prohibited, plus the material decisions awaiting user approval.
• Control: obtain explicit approval/amendment via `hitl`; unresolved material decisions block.

</authorize>

<deterministic_gates phase="3" applies="development/change/PR/pipeline" subagent="executor" role="Bounded deterministic security-gate operator" subagent_required_model="Claude Haiku 4.5, GPT-5.6 Terra, Gemini 3.7 Flash, GPT-5.6 Luna" must-be-subagent>

• Purpose: Runs approved deterministic lifecycle gates and preserves source results unchanged. Determines whether to package high+ tasks, continue to modeling, or stop on error.
• Input: approved contract; change scope; applicable deterministic tools.
• Output: deterministic evidence and branch result.
• INVOKE SUBAGENT `executor` to APPLY PHASE `security-flow-deterministic-gates.md` + run approved deterministic gates and return unchanged findings with `HIGH+|CLEAN|ERROR`.
• Expect: unchanged source records, evidence metadata, `HIGH+|CLEAN|ERROR`.
• Control: HIGH+ → report-and-package only; CLEAN → model-and-select; ERROR → stop.

</deterministic_gates>

<model_and_select phase="4" applies="ALL" subagent="architect" role="Security architect mapping threats to complete contextual coverage" subagent_required_model="Claude Opus 5, GPT-5.6 Sol, Gemini 3.7 Flash" must-be-subagent>

• Purpose: Builds a threat model and maps applicable authorized areas, activities, tools, and exclusions. Produces the complete coverage plan for inspection dispatch.
• Input: approved contract; permitted context; available tools.
• Output: threat model and authorized coverage plan.
• INVOKE SUBAGENT `architect` to APPLY PHASE `security-flow-model-and-select.md` + build the threat model, map applicable areas/tools/exclusions, and return the complete authorized coverage plan.
• Expect: threat model, applicable-area/tool plan, exclusions, residual risk.
• Control: full review covers all applicable/available/authorized work; gaps return for correction.

</model_and_select>

<inspect_and_test phase="5" applies="ALL" subagent="engineer" role="Security engineer producing bounded evidence by applicable area" subagent_required_model="Claude Sonnet 5, GPT-5.6 Terra, Gemini 3.7 Flash" must-be-subagent>

• Purpose: Execut