workbench-core

Review and compact accumulated skill learnings. For workbench plugin skills, optionally integrate proven learnings into the SKILL.md itself. Triggered when a learnings file exceeds 30 entries, or run manually any time.

Interactive interview to build or refine the user's profile.md. Asks about role, working style, technical stack, and communication preferences. Focuses on facts about the user — agent behavioral rules belong in soul-hot.md via /workbench:define-soul.

Interactive onboarding and refinement for agent identity — soul-hot.md and soul-core.md. Walks the user through defining who their agent is through a conversational series of questions, pushing back on vague or contradictory answers until each file is solid. Works as both first-time setup and refinement of existing files.

Grade recorded decisions and memory entries for decision quality — correctness against outcomes, accuracy/efficiency/speed, consistency/recurrence, and missing process — and write a learnings report. Gear 2 ("Evaluate") of the decision-quality learning loop; feeds /workbench-core:propose-upgrades. Run manually any time.

Install the plugin's shipped persona into your live Claude Code locations — soul files to the memory vault, the output style to ~/.claude/output-styles, and the outputStyle setting in ~/.claude/settings.json. Non-destructive — hand-edited soul files are diffed and confirmed before any overwrite. Use when setting up the agent persona, or re-syncing after a plugin update.

Install workbench-* plugin skills into Claude Chat (Mac app) via .skill packaging. Discovers all eligible skills in installed @claude-workbench plugins (excluding workbench-core itself), packages them with skill-creator's package_skill.py, and opens each .skill file with the Mac app to trigger the install dialog. Use this skill whenever the SessionStart warmup output mentions new Chat-installable skills, or to manually re-sync Chat skills after installing or updating a workbench plugin.

Log the current session segment right now — dump the raw log and write the narrative summary + any decision promotions inline. Use this when you want to snapshot mid-conversation, or when you want a richer summary than the auto-generated one.

Periodic health-and-repair pass over the memory vault — rescue files skipped for missing frontmatter, repair broken links, conservatively connect orphans, repair vault-index drift, flag duplicates for human review, and write an audit report. Run monthly via the scheduled-tasks MCP, or manually any time.

Report the shared memory server's facts — vault/cache paths, probe health, bearer token, live session refs, server binary, and index/maintenance state. Use when memory search/write is failing or to confirm the server is wired.

Process any pending session summaries by dispatching background agents. Use when the warmup notices unprocessed markers, or run manually to clear the backlog. Does NOT block the session — dispatches agents in the background and moves on.

Turn a decision-quality evaluation into concrete proposals — corrections to existing memories and new process recordings — then walk human sign-off and apply only what's approved. Gears 3+4 ("Propose" + "Sign-off") of the decision-quality learning loop; consumes /workbench-core:evaluate-decisions output. Run manually any time.

Configure the workbench — agent name, memory paths, MCP server name, identity file paths, and the permission safety rails (defaultMode plus deny/ask rules) written to ~/.claude/settings.json. Config lives in the plugin data directory and is read at MCP start time, so plugin updates never clobber settings.

Manually summarize a specific session by ID or by picking from recent unsummarized logs. Use when the auto-summarizer missed a session, when you want to re-summarize with better context, or when you want to summarize an older session.