Refactor to native routing: skill-owned dispatch, consolidated meta-skills, openspec removal

[cpliakas]

Summary

Restructures the plugin around Claude Code's native mechanisms, removing the hand-rolled orchestration machinery while keeping the advisory-agent premise intact. Net effect excluding the openspec removal: 35 files, +1,214/−1,935.

1. Plugin hook infrastructure + explicit tool grants (1227592)

• Adds .claude-plugin/plugin.json and hooks/hooks.json so the markdownlint PostToolUse hook actually registers and fires (the script existed but was never wired).
• Hardens the hook script: lint findings (exit 1) are distinguished from tool failures (exit ≥ 2, now surfaced instead of silently disabling linting), output is jq-escaped, truncation is marked.
• Every agent now declares an explicit tools grant: advisory agents get Read, Glob, Grep, Skill (the skills: frontmatter only preloads content — the Skill tool is what allows invocation); Bash only where the body runs commands. Codified in CLAUDE.md.

2. Routing inversion (010da7d) — breaking

• /plan-implementation now owns all mechanical routing: it reads the Tech Lead's registry directly, matches specialists semantically, classifies the tier via the Signals Catalog, and dispatches all five target types itself (subagent/external-agent spawned in parallel, skills invoked, docs read, human questions surfaced as open items).
• The Tech Lead is invoked once, for judgment only: escalation flags (tier-3 one-way doors must name chief-architect), conflict resolution between specialists, constraints, and the recommended approach. tech-lead.md drops from 782 to ~440 lines.
• The hand-parsed markdown contract ("Parseable Phase 1/2 Output Contracts") and every parse-failure path built around it are deleted.
• Routing-outcome capture is removed outright rather than closed into a feedback loop: an auto-tuning registry would violate the plugin's advise-never-mandate stance, and the write-only log had no feedback path. /audit-routing-quality is deleted with it.

3. Consolidation + openspec removal (fc6b95b) — breaking

• /re-onboard merges into /onboard --check-drift (same diff pass, same confirm-before-write flow, optional agent-name scoping).
• /audit-routing-table's structural checks fold into /audit-agent-memory as a tech-lead-only section; the legacy-format MIGRATION.md moves alongside.
• openspec/ is removed along with every live reference — proposals no longer flow through openspec.
• Version bumps to 0.14.0.

4. Review-gate fixes (d17877b)

Cross-cutting fixes from the final adversarial review, dominated by one theme: the target-type registry format wasn't understood by every consumer. The drift check and the audit checks are now target-type aware (previously the drift check could batch-delete valid human/skill/external-agent registrations), agent-file resolution falls back to .claude/agents/ for consuming projects, unknown target types warn-and-exclude instead of coercing to subagent, write-convention spawns the domain owner directly, and agent bodies search with Glob/Grep instead of delegating to a subagent they cannot spawn.

Breaking changes

• tech-lead no longer emits consultation-request output; callers that drove the two-phase protocol manually must use /plan-implementation.
• Removed skills: /audit-routing-quality, /re-onboard (→ /onboard --check-drift), /audit-routing-table (→ /audit-agent-memory tech-lead).

Review process

Each batch ran through an explore → implement → review → fix → test workflow (Opus adversarial review of the staged diff) followed by an independent /code-review high gate; all findings from both layers were addressed before each commit.

After merge

• Tag v0.14.0 per the repo's versioning convention.
• Reinstall/update the locally installed plugin (currently pinned at 0.13.0) so the new skill set takes effect.

🤖 Generated with Claude Code

https://claude.ai/code/session_016myji5spWpb6qwqUQesXCr