workbench-core

Plugin workbench-core in mike-bronner/workbench-core. Commands read from the repo's manifest — yesterday.

Skills (15)

• check-a-live-source

  Use before you name or rely on a fact about the world that goes stale and that you know from training, not from a source read in this task, such as the newest or latest model, a model ID, a library or tool version, a price, a limit or quota, a launch or release date, or what a product's current docs

• compact-learnings

  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.

• cross-session-messaging

  Messaging another Claude Code session with SendMessage, and handling a peer message. Use when findings affect a related session, a SendMessage was refused or flagged, or another session's message arrives.

• evaluate-decisions

  Grade recorded decisions and memories for decision quality, and write the learnings report propose-upgrades reads. Use when asked to evaluate decisions, and in the nightly decision-quality run.

• install-chat-skills

  Install the skills of installed workbench-* plugins into the Claude Mac app. Use when the warmup notices say new Chat-installable skills exist, or to re-sync after installing or updating a workbench plugin.

• intake

  The task-intake routine for an interactive session. It sets the goal, asks about gaps through AskUserQuestion, writes acceptance criteria, and grades three options. Run it before the first work on a task that ends in a work product (a change, file, document, or plan). Skip it for trivial asks and fo

• log-now

  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.

• memory-lint

  Monthly health-and-repair pass over the memory vault, ending in an audit report. Use when the monthly memory-lint scheduled task runs, or when asked to lint or repair the vault.

• propose-upgrades

  Turn the latest decision-quality report into proposed memory corrections and process notes, and apply only what the user signs off. Use after evaluate-decisions, and in the nightly decision-quality run.

• read-the-contract

  Use when a field would go out as {}, an empty value, a placeholder or a guess, above all one whose description says another tool hands it out, even when the user asks for it. Also use when a required field has no sourced value, or you would offer the user a choice of guessed values. Use before a wri

• recheck-before-done

  Use before you submit, claim, publish, release or close work, mark a pull request ready, or otherwise report work as done or send it out, when that work was built from a page, list, file, ticket or message read earlier in the session, minutes or hours ago. Above all, use it when its owner updates th

• session-timeline

  Use before you rely on what earlier agent sessions said, decided or did when their logs are on disk, such as rebuilding work from past sessions or finding which draft came last. It reads Claude Code transcripts under ~/.claude/projects, each sub-agent's own file under <session>/subagents included, a

• setup

  Configure the workbench — agent name, memory paths, MCP server name, the permission safety rails (defaultMode plus deny/ask rules) written to ~/.claude/settings.json, and the shipped output style. Re-run after a plugin update to re-sync the output style. Config lives in the plugin data directory and

• show-every-cut

  Use before you shorten, summarize, condense, trim, paraphrase or rewrite text a person gave you (a message, notes, a document, a post, a quote), before you fit their text to a character or word limit, and when some of their passages must stay word for word.

• use-the-owners-logic

  Use before you give a number, count or category that must match the owner's table, page or report, or comes from data that ships its own logic, views or build (a script, a component, a query, a report generator, a test suite), such as a function or SQL view their dashboard reads. Run their code and

Subagents (2)

• meeting-writer

  Background agent that writes the vault note for one finished FluidVoice meeting, with its summary, decisions and action items, and promotes its decisions. Dispatched by scripts/meeting-notes.sh, which the hooks module runs at session start and when the FluidVoice Meetings folder changes.

• summary-writer

  Background agent that processes a pending session summary. Reads the raw log, writes a narrative summary to the memory vault, promotes decisions, and releases the marker. Dispatched by the session-log hook when WORKBENCH_AUTO_SUMMARIZE=1, by the session warmup's drain, or manually with /process-pend

MCP servers (1)

• memory

  connects to: http://127.0.0.1:${WORKBENCH_MEMORY_PORT:-8765}/mcp

Hooks (11)

• SessionStart

  bash "${CLAUDE_PLUGIN_ROOT}/hooks/memory-server-up.sh"

• SessionStart

  bash "${CLAUDE_PLUGIN_ROOT}/hooks/session-warmup.sh" --defer

• SessionStart

  command -v node >/dev/null 2>&1 && node "${CLAUDE_PLUGIN_ROOT}/hooks/fallback/session-start.ts" || true

• SubagentStart

  command -v node >/dev/null 2>&1 && node "${CLAUDE_PLUGIN_ROOT}/hooks/fallback/subagent-start.ts" || true

• Stop

  command -v node >/dev/null 2>&1 && node "${CLAUDE_PLUGIN_ROOT}/hooks/fallback/stop.ts" || true

• PreToolUse

  command -v node >/dev/null 2>&1 && node "${CLAUDE_PLUGIN_ROOT}/hooks/fallback/pre-tool-use.ts" || true

• FileChanged

  command -v node >/dev/null 2>&1 && node "${CLAUDE_PLUGIN_ROOT}/hooks/fallback/file-changed.ts" || true

• PreCompact

  bash "${CLAUDE_PLUGIN_ROOT}/hooks/session-log.sh"

• PostCompact

  bash "${CLAUDE_PLUGIN_ROOT}/hooks/session-warmup.sh"

• SessionEnd

  bash "${CLAUDE_PLUGIN_ROOT}/hooks/session-log.sh"

• SessionEnd

  bash "${CLAUDE_PLUGIN_ROOT}/hooks/memory-server-release.sh"