Incrementally decompose monolithic workflow_orchestrator.md

[barkain]

Context

After the native plan mode refactor (PR #30), workflow_orchestrator.md grew to 1000+ lines by absorbing the task-planner skill's planning logic. While functional, this creates maintainability concerns.

Proposal

Incrementally move deterministic logic out of the prompt and into code/dynamic loading:

Phase 1: team_mode_score calculation → hook

• The 8-factor scoring table is deterministic logic better suited as a Python hook
• A PreToolUse or PostToolUse hook could calculate the score and inject the result
• Removes ~50 lines of scoring instructions from the prompt

Phase 2: Agent selection keywords → dynamic from agent files

• Agent keyword table is currently hardcoded in the prompt
• Could be dynamically loaded from agent YAML frontmatter at planning time
• Keeps agent definitions as single source of truth
• Removes ~30 lines from the prompt

Phase 3: Evaluate contextual injection

• Investigate whether planning instructions can be injected on-demand (e.g., when EnterPlanMode is detected)
• Current constraint: SessionStart hook fires once, no mid-conversation injection mechanism
• May require Claude Code platform support for event-based prompt injection

Non-goals

• Full file split into many small files (coordination overhead outweighs benefits)
• Moving LLM reasoning guidance into code (decomposition rules, atomicity validation are inherently prompt-based)

References

• PR refactor: replace task-planner skill with native plan mode (EnterPlan… #30 review comment from Qoso
• system-prompts/workflow_orchestrator.md (Planning Instructions section, lines 376+)