Complex generative work must be decoupled into independent nodes optimized for specific computational traits.

The core differentiator of the Brainstorm Reactor is the deployment of pre-validated Orchestration Patterns across its execution layer. Single-shot prompts consistently fail at maintaining high degrees of structural complexity. Orchestration paradigms resolve this by isolating specific generative loads.

Pattern A: The "Shadow Pass" & Selective Amnesia

The Shadow Pass isolates raw idea generation from final presentation to eliminate context contamination and user prompt bias.

1. Step 1 (Visible Baseline): Extracts facts with context_mode: 'full'.
2. Step 2 (Silent Compiler): Evaluates deep abstractions with is_silent: true.
3. Step 3 (Main Task): Executes the target transformation with context_mode: 'last_step'.

By switching to 'last_step', the main agent completely forgets the user's original raw input and works only on the distilled baseline.

Pattern B: Parallel Swarm & Synthesis

Linear thinking compounds strategic bias. Pattern B addresses strategic narrowness by executing simultaneous, adversarial reasoning paths.

• The Swarm: Multiple distinct agents are assigned divergent behavioral directives and run simultaneously via parallel_group.
• The Synthesizer: Once the slowest node terminates, the independent logic outputs are compressed into an XML matrix and submitted to a centralized synthesis step. The Synthesizer is purposefully denied original user context, forcing it to generate consensus based strictly on the structured output of the Swarm.

Pattern C: Autonomous Critique & Revise (Aletheia)

Fully autonomous self-correction loop using logic_branch with mode: "llm_auto_route" and max_loops.

1. Generation: A creative node draft is produced.
2. AI Check: An auditor node evaluates the output against strict constraints and criteria.
3. Loop Control: If verification fails, a backward jump is triggered. The engine automatically prunes the conversation history (Context Collapse) to prevent token explosion and loops until the quality threshold is met.

Pattern D: User Decision Branching (Interactive Router)

Uses logic_branch with mode: "keyword_match" to pause execution and await user feedback.

• Pause and Present: The workflow pauses at defined decision points, presenting the generated choices as Quick Reply buttons.
• Routing: The user's button click triggers the execution of the corresponding routed branch.
• Pattern D2 (Split Router): Splits this into a silent generator (high-temperature ideation) and a visible presenter (cool-temperature curation) to guarantee formatting precision.

Pattern E: Unconditional Forward Jump

Used to merge converging paths or skip steps in a workflow.

• Convergence: When multiple slots or independent branches need to converge to a single final synthesis step, they invoke a logic branch with empty routes and a fallback_step_order.
• Flow Control: This forces an unconditional jump forward to the synthesis step once the branch execution terminates, bypassing unused steps.

Pattern F: Epistemic Anchor (Auto-Pin)

Prevents critical scope constraints from being lost during selective amnesia passes.

• Auto-Pinning: When auto_pin_output = true, the engine automatically pins the step's output message upon completion.
• Persistent Context: The prompt assembler injects pinned messages directly into the <pinned_context> block (Layer 7). This data survives subsequent context_mode: 'last_step' pruning passes and remains directly accessible to all downstream agents.