evolve: Adopt Claude Code v2.1.122-2.1.157 features (/goal, dynamic workflows, disallowed-tools, args exec form, continueOnBlock, Opus 4.8)

[krishagel]

Source

Auto-created by /evolve running in: psd401/psd-claude-plugins
Classification: Plugin-level

Amended 2026-05-29: The original /evolve run missed the two headline Claude Code features (/goal, dynamic workflows) because the evolve skill fetches a lossy changelog source and has no command/orchestration check category. Tracked separately in #65. Added below.

Findings

36 Claude Code versions shipped since our last adopted feature (v2.1.121, tracked 2026-04-23). Several new capabilities are directly applicable to our plugin architecture. Current install: v2.1.157.

Recommended Changes

High-value (headline features — added in amendment)

• Evaluate /goal for autonomous loops (v2.1.139, May 11) — Set a completion condition; the agent works across turns until a fast model confirms it's met. Up to 4,000-char conditions; works in interactive, -p, and Remote Control. Assess whether /lfg, /optimize, and /test self-healing loops should delegate loop control to /goal instead of hand-rolled iteration. Reduces custom orchestration in skill bodies.
• Evaluate dynamic workflows vs /swarm (v2.1.154, May 28) — JS scripts orchestrate tens–hundreds of subagents in background (16 concurrent, 1,000 total cap), with built-in find → refute → converge verification. Direct functional overlap with our /swarm (Agent Teams) skill. Decision needed: rebuild /swarm on workflows, deprecate it, or keep both? Also assess workflows for /lfg, deep code review, and /evolve's own multi-agent analysis. Requires v2.1.154+; Max/Team/Enterprise plans.

Other adoptions (original analysis)

• Evaluate Opus 4.8 for skills (v2.1.154) — Opus 4.8 is available with high effort default. Test key skills (architect, lfg, evolve) on 4.8 vs 4.6 quality. Update model: frontmatter if warranted.
• Adopt disallowed-tools frontmatter (v2.1.152) — Add to read-only agents (learnings-researcher, spec-flow-analyzer, best-practices-researcher, etc.) to prevent accidental writes. Example: disallowed-tools: Write, Edit
• Adopt args exec form for hooks (v2.1.139) — Convert WorktreeCreate and WorktreeRemove hooks from shell command strings to args: string[] arrays. Eliminates shell injection risk and complex escaping.
• Adopt continueOnBlock for PostToolUse (v2.1.139) — Add to syntax validation hook so syntax errors are fed back to the model as rejection reasons instead of silent failures.
• Adopt $CLAUDE_EFFORT in hooks (v2.1.133) — Available in hook environment. Could skip expensive validation at low effort levels.
• Adopt MessageDisplay / reloadSkills / sessionTitle SessionStart hooks (v2.1.152) — New hook capabilities worth assessing for the plugin.
• Update CLAUDE.md adopted features table — Currently stops at v2.1.121. Document new adoptions.

Evidence

• Canonical changelog: https://code.claude.com/docs/en/changelog (NOT the GitHub blob mirror, which the skill currently uses)
• Dynamic workflows: https://www.marktechpost.com/2026/05/28/anthropic-ships-claude-opus-4-8-alongside-dynamic-workflows-and-cheaper-fast-mode-with-workflows-capped-at-1000-subagents/
• /goal: https://explainx.ai/blog/claude-code-goal-command-long-running-agents-2026
• Current hook config: plugins/psd-coding-system/hooks/hooks.json — uses shell command strings, no args exec form
• Current agents: read-only agents like learnings-researcher have no disallowed-tools restriction
• No breaking changes detected across all 36 versions

---

Auto-generated by /evolve Phase 4.5 — Issue Routing; amended manually 2026-05-29

[krishagel]

🤖 Picked up by the lfg routine. Working on this now. PR will be linked here when ready, or this issue will be marked lfg-blocked with a reason if I can't complete it.

---

Generated by Claude Code

[krishagel]

🤖 PR opened and security-audited: #66

Implemented:

• disallowed-tools: [Write, Edit] on all 6 research agents (v2.1.152)
• Dynamic workflows documented in /swarm as upgrade path (v2.1.154)
• CLAUDE.md adopted features table updated

Requires human implementation (hooks.json is a protected path):

• args exec form for hooks (v2.1.139)
• continueOnBlock on PostToolUse (v2.1.139)
• $CLAUDE_EFFORT in hooks (v2.1.133)
• MessageDisplay/reloadSkills/sessionTitle SessionStart hooks (v2.1.152)

---

Generated by Claude Code

[krishagel]

PR #66 was closed as stale (it targeted the pre-v3.0.0 layout, including the deleted swarm skill). Keeping this issue open, but it needs re-scoping: the adoption analysis should target the current 6-skill structure, and the version range has moved on (adoption table currently ends at v2.1.121). The re-applied /evolve Phase 3B fix (v2.21.5) will surface these features on its next release-check run.