Workflow Routing

Overview

Route user requests to the right workflow based on their goal. Keep main context for coordination; delegate work to agents.

Goal Detection

Detect the user's primary goal from their message:

If intent is clear from context, infer the goal. Otherwise, ask:

What's your primary goal?
1. Research — understand/explore something
2. Plan — design and spec a solution (brainstorming → premortem → writing-plans)
3. Encode — decompose a spec into beads (beadcraft)
4. Build — execute beads and ship
5. Fix — debug/fix an issue

Workflow Chains

Research

1. •explore skill at appropriate depth
2. •Document findings
3. •Suggest: "Ready for planning?"

Plan

Runs automatically as a single chain — do not stop between steps.

1. •brainstorming — explore requirements, generate design options, make decisions (includes per-decision premortems)
2. •premortem — full plan-level risk analysis on the chosen design
3. •writing-plans — produce a spec document incorporating brainstorming output and premortem mitigations
4. •Suggest: "Ready to encode into beads?"

Output: A spec document ready for decomposition.

Encode

Runs automatically — invoke beadcraft and present the tree.

1. •beadcraft (decompose mode) — parse spec into epic + task beads with full quality (Rule of Five, acceptance criteria, Read/Signature/Edges)
2. •Present bead tree for user confirmation
3. •Suggest: "Ready to build?"

Input: Spec from Plan phase. Output: Bead dependency graph.

Build

Runs automatically — execute beads in dependency order.

1. •executing-beads — TDD cycle per bead, quality gate, atomic commit
2. •requesting-code-review — review between batches
3. •finishing-work — integrate and clean up

Input: Bead graph from Encode phase.

Fix

1. •systematic-debugging — find root cause
2. •test-driven-development — write failing test, fix
3. •verification-before-completion — verify fix
4. •finishing-work — integrate

Parallel Detection

When tasks are independent, suggest parallel agents:

User: "Research auth patterns and fix the login bug"

→ Detect: 2 independent tasks (Research + Fix)
→ Suggest parallel agents
→ Synthesize results

Main Context = Coordination

Delegate to agents:

• •Reading 3+ files → agent
• •External research → agent
• •Implementation → agent
• •Running tests → agent

Keep in main context:

• •Understanding user intent
• •Workflow selection
• •Agent coordination
• •Presenting summaries

After Completing a Workflow

Suggest the natural next step: