feat(llm-assistants): enrich subagent config with effort, memory, and team presets

[hakula139]

Summary

• Add six new frontmatter parameters (effort, memory, permissionMode, background, maxTurns, isolation) to all Claude Code subagent definitions, enabling cost-aware model routing, persistent cross-session knowledge, and formal read-only enforcement
• Assign per-agent effort levels based on task complexity: high for deep-reasoning agents (architect, reviewer, debugger, implementer), medium for balanced agents (tester, usability-reviewer), low for speed-oriented agents (researcher, codex-worker)
• Enable persistent local memory on all seven agent roles so each accumulates project-specific institutional knowledge across sessions (architectural patterns, recurring issues, file maps, coding conventions, test patterns, debugging dead ends)
• Formalize read-only enforcement via permissionMode: plan on the four agents that were previously read-only by behavioral convention only (architect, researcher, reviewer, usability-reviewer)
• Set background: true on researcher and codex-worker as default non-blocking execution for exploratory tasks
• Add maxTurns: 30 safety cap on codex-worker to prevent runaway haiku-tier delegation loops
• DRY up the frontmatter renderer with a declarative renderField helper that reduces 8 repetitive lib.optional lines to a single concatMap over a field spec list, making new fields a one-liner addition
• Switch tools rendering from YAML list to comma-separated format (tools: Read, Grep, Glob) matching Anthropic's docs style, applied to both the renderer and codex-worker.md
• Merge "Coordination Patterns" and "Team Presets" in CLAUDE.md instructions into a single section with inline mode annotations (Subagents vs Agent Team), eliminating duplication between the previous three overlapping sections
• Add persistent memory prompts to all seven agent role definitions, instructing each to consult and update its agent memory with role-appropriate knowledge

Architecture

• Why effort per agent: The blog post and Anthropic docs describe effort as a cost-quality lever orthogonal to model selection. A haiku researcher at low effort returns fast lookups cheaply; an opus architect at high effort produces thorough analysis. Combined with model choice, this gives fine-grained control over cost per delegation.
• Why memory: local: Local scope (.claude/agent-memory-local/) keeps agent knowledge project-specific without adding files to version control. All agents benefit from accumulated knowledge — the debugger remembers failed approaches, the implementer remembers codebase conventions, the researcher remembers file locations. The cost is a directory on disk; the benefit compounds over sessions.
• Why permissionMode: plan over disallowedTools: The four read-only agents were previously constrained by prompt instructions alone. permissionMode: plan formalizes this at the runtime level while still allowing boundary exceptions when genuinely needed (unlike a hard tool denylist).
• Why merge coordination patterns and team presets: The previous structure had three overlapping sections describing the same workflows at different granularity (abstract patterns, concrete team compositions, detailed step-by-step). Merging into a single list with mode annotations (Subagents / Agent Team) per entry eliminates duplication while preserving all information.
• Why renderField helper: The previous 8 lib.optional lines followed an identical pattern differing only in field name and formatter. The declarative spec list makes the field set scannable, and adding a new field is a one-liner. String entries default to toString (works for strings and ints); only background needs lib.boolToString.

Test plan

• nix fmt passes on all modified Nix files
• statix check passes (no anti-patterns)
• deadnix --fail passes (no unused bindings)
• Pre-commit hooks all pass
• nix build '.#homeConfigurations.hakula-linux.activationPackage' --dry-run succeeds
• nix eval '.#nixosConfigurations.us-4.config.system.build.toplevel.drvPath' succeeds