fix(skill-management): remove broken agent refs and widen audit recall (#570)

* fix(skill-management): Opus 5 migration fixes for skill-doctor and migrate-config-to-opus-5

Remove three phantom agent references, make the trigger-quality subagent
dispatch conditional, make the analysis pass coverage-first, and add
deliverable-length lines to both output sections.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fix(skill-management): balance the analysis rubric's precision guard

analysis-rubric.md opened with "don't invent problems" and nothing on the
other side. The intent is right, but in an auditing prompt an unbalanced
precision instruction reads to Opus 5 as license to drop anything it is
uncertain about -- which is a recall filter applied at discovery time.

Added the coverage counterweight: the guard constrains what you assert,
not what you look at. Report every issue found, mark the uncertainty,
and let the reader filter. A silently omitted real finding fails the same
way a fabricated one does, and is harder to notice.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fix(skill-management): drop PROACTIVELY from skill-doctor's description

audit-patterns.md lists "PROACTIVELY in agent descriptions" as a Delta 2
smell to search for, and skill-doctor's own description used it. Shipping
the pattern the plugin tells users to flag undercuts the guidance.

The trigger itself is kept -- skill-doctor still applies after a repeatable
multi-step workflow, especially right after a PR. Only the nudge word is
gone. The PR-time prompt is delivered by hooks/pr-skill-doctor-prompt.cjs,
not by this description, so nothing is lost by removing it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

Author: wkoutre

CLAUDE.md

@@ -228,7 +228,7 @@ All plugins follow semantic versioning (semver). Key versioning rules:
 | development-planning       | 2.0.7   |
 | development-pr-workflow    | 2.4.0   |
 | development-productivity   | 2.4.1   |
-| skill-management           | 1.2.0   |
+| skill-management           | 1.3.0   |
 | spec-workflow              | 2.1.0   |
 | uniswap-integrations       | 2.6.1   |
 

packages/plugins/skill-management/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "skill-management",
-  "version": "1.2.0",
+  "version": "1.3.0",
   "description": "Audit, map, mine, and improve your Claude Code skills, agents, and slash commands. Inventories your whole customization surface, flags overlaps/gaps/weak triggering descriptions, and mines sessions for workflows worth codifying into new skills or agents. Also migrates full Claude Code configs (CLAUDE.md, rules, settings, hooks, CI scripts) for new model generations like Opus 5.",
   "author": {
     "name": "Uniswap Labs",

packages/plugins/skill-management/README.md

@@ -46,8 +46,8 @@ The hook is a dependency-free Node script invoked via `node ${CLAUDE_PLUGIN_ROOT
 
 `/skill-map`, `/skill-mine`, and `/skill-new` are thin prompts that invoke the **skill-doctor** skill
 in a specific run mode. The skill does the _finding_ — what to add, merge, fix, or codify — then hands
-the deep work to the tools that own it (`skill-creator` for drafting/evals, `agent-optimizer` /
-`prompt-engineer` for agent tuning). It never auto-applies a change: it proposes a prioritized menu and
+the deep work to the tools that own it (`skill-creator` for drafting/evals,
+`development-productivity:prompt-engineer-agent` for agent tuning). It never auto-applies a change: it proposes a prioritized menu and
 lets you pick. Personal config under `~/.claude/{skills,agents,commands}` is edited in place; skill /
 agent / command files that live inside a git repo are delivered as a **draft PR** off the repo's
 default branch.

packages/plugins/skill-management/skills/migrate-config-to-opus-5/SKILL.md

@@ -71,3 +71,5 @@ Present the "user decision" findings via AskUserQuestion (batch related ones; re
 ## Output
 
 End with: scope migrated, a per-commit list of what changed and why, findings deliberately left alone (with the policy reason), anything out of reach (other machines, external repos), and the checkpoint date.
+
+Keep it to what the user has to act on — one line per commit, one line per left-alone finding. No restated summary of the audit, no filler sections.

packages/plugins/skill-management/skills/skill-doctor/SKILL.md

@@ -8,9 +8,9 @@ description: >
   skills do I have", "suggest skill improvements", "are any of my skills
   redundant", "clean up my skills", "turn this into a skill", "should this be a
   skill or an agent", "codify this workflow", or invokes /skill-map, /skill-mine,
-  or /skill-new. Also use PROACTIVELY after finishing a multi-step workflow that
-  the user is likely to repeat (especially right after opening a PR) to ask
-  whether it should become a skill or agent. It inventories every
+  or /skill-new. Also applies after finishing a multi-step workflow the user is
+  likely to repeat (especially right after opening a PR), to ask whether it
+  should become a skill or agent. It inventories every
   skill/agent/command across the user's own dirs AND all installed marketplaces,
   flags overlaps / gaps / weak triggering descriptions, and — when there's real
   prior work in the session — proposes new or edited skills/agents grounded in
@@ -26,11 +26,11 @@ the deep work to the tools that already own it. **Reuse, don't reinvent:**
 - **Creating or iterating a skill (drafting, evals, description optimization, packaging)** → invoke the `skill-creator` skill if it's
   installed (it ships in Anthropic's agent-skills marketplace and bundles
   `improve_description.py`, `run_loop.py`, `package_skill.py`, `quick_validate.py`).
-- **Tuning an agent's prompt** → delegate to the `agent-optimizer` or `prompt-engineer` agents.
-- **Scoring an agent's capabilities / team fit** → the `agent-capability-analyst` agent.
-- **Cataloging agents** → the `claude-agent-discovery` agent.
+- **Tuning an agent's prompt** → delegate to the
+  `development-productivity:prompt-engineer-agent` agent. Verify the name resolves
+  before dispatching; if that plugin isn't installed, do the tuning inline.
 - **"This is really a standing rule, not a skill"** → write it into the relevant
-  `CLAUDE.md` (e.g. via an `update-claude-md` skill) or the project's memory store.
+  `CLAUDE.md` (e.g. via the `/update-claude-md` command) or the project's memory store.
 - **Auditing / cleaning up the memory store itself** (orphans vs `MEMORY.md`, stale
   memories, dedupe, "scope memories into skills") → the `memory-doctor` skill if
   installed, the sibling of this one for the memory surface. skill-doctor _reads_
@@ -108,21 +108,29 @@ Reason over the map (not the raw files). Produce a **prioritized, numbered list*
 of suggested improvements. For each: target file, problem, proposed change, and
 whether it's `writable`. Cover:
 
-- **Overlaps / duplicates.** Use the `duplicate_names` and
+The script's flags are a starting point, not the boundary of the audit. Read the
+whole map and report every genuine issue you see, including ones the script did
+not flag — the flags are keyword and threshold heuristics, so a real overlap or a
+weak description can sit in an unflagged row. Rank at the end; do not filter
+during discovery.
+
+- **Overlaps / duplicates.** Start from the `duplicate_names` and
   `near_duplicate_descriptions` flags. Ignore `mirror-only` duplicates
   (marketplace↔plugin-cache are just install mirrors of the same upstream item).
   Focus on `[editable]` collisions and genuinely distinct skills doing the same
   job. Recommend a canonical one + merge/deprecate the rest.
 - **Gaps.** Recurring needs with no skill. Cross-reference the project's memory
   store if one exists — many feedback memories encode repeated corrections that
   may deserve a skill.
-- **Trigger-quality.** Use `weak_or_missing_description` and
-  `no_trigger_language`. For the handful of **writable** items you'll actually
-  propose changing, deep-read them **via a subagent** (Explore or general-purpose)
-  so main context stays bounded — ask the subagent to return the current
-  description + body summary + a tightened description proposal. For rigorous
-  triggering work, hand the item to skill-creator's `improve_description.py` /
-  `run_loop.py` loop.
+- **Trigger-quality.** Start from `weak_or_missing_description` and
+  `no_trigger_language`, and flag any other description that reads weak on the
+  rubric. Deep-read the **writable** items you'll actually propose changing. Read
+  them directly when there are only a few (a handful of reads is cheaper inline
+  than a dispatch); delegate to a single Explore or general-purpose subagent only
+  when the set is large enough that reading it inline would crowd out main
+  context — ask that subagent to return the current description + body summary +
+  a tightened description proposal per item. For rigorous triggering work, hand
+  the item to skill-creator's `improve_description.py` / `run_loop.py` loop.
 - **Oversized bodies.** `oversized_body` items (>500 lines) are candidates for
   progressive disclosure (move detail into `references/`).
 
@@ -153,11 +161,14 @@ turn that motivates it. Vague "you could add a skill for X" proposals aren't use
 ### Step 5 — Present & apply (all modes)
 
 Show the consolidated, prioritized menu. Mark each item `[editable]` or
-`[read-only]`. Let the user choose. Then, for accepted items only:
+`[read-only]`. Keep each entry to a line or two — target, problem, proposed change
+— so the menu stays scannable; detail belongs in the follow-up on the items the
+user picks. Let the user choose. Then, for accepted items only:
 
 - New/iterated skill → `skill-creator` skill.
-- Agent prompt tuning → `agent-optimizer` / `prompt-engineer` agents.
-- Standing rule → an `update-claude-md` skill or a memory file.
+- Agent prompt tuning → the `development-productivity:prompt-engineer-agent` agent
+  (inline if that plugin isn't installed).
+- Standing rule → the `/update-claude-md` command or a memory file.
 - Direct small edits (e.g. a tightened description):
   - under `~/.claude/…` (non-git) → edit in place; show the before/after first.
   - inside a git repo → open a **draft PR** per "Delivering changes once

packages/plugins/skill-management/skills/skill-doctor/references/analysis-rubric.md

@@ -3,6 +3,13 @@
 Heuristics for the analysis + mining passes. Keep judgments concrete and tied to
 the inventory data; don't invent problems.
 
+Precision and coverage are both required, and they pull against each other.
+"Don't invent problems" constrains what you _assert_, not what you _look at_ —
+report every issue you actually found, including ones you are unsure about or
+judge low-severity, and mark that uncertainty rather than dropping the finding.
+An audit that silently omits a real issue fails the same way as one that
+fabricates a fake one, and the omission is harder to notice.
+
 ## What a good triggering description looks like
 
 The `description` is the _only_ thing Claude sees when deciding whether to