foundry-works

Task implementation skill for spec-driven workflows. Reads specifications, identifies next actionable tasks, and creates detailed execution plans. Use when ready to implement a task from an existing spec - bridges the gap between planning and coding. TRIGGER when: user explicitly asks to implement, continue work, or pick up the next task from a spec (e.g., "what's next", "continue", "implement"). Never call mcp task/spec tools directly for task workflow operations. DO NOT TRIGGER when: a <command-name> tag is present in the current turn (skill already loaded), called from a subagent, or model decides on its own that implementation should start.

AI-powered research skill with five workflows - chat (single-model conversation), consensus (multi-model synthesis), thinkdeep (systematic investigation), ideate (creative brainstorming), and deep (multi-phase web research). Supports persistent threads and research sessions. TRIGGER when: user asks for research, consultation, brainstorming, investigation, or deep research. Always route through this skill — never call mcp__plugin_foundry_foundry-mcp__research directly. DO NOT TRIGGER when: a <command-name> tag is present in the current turn (skill already loaded), called from a subagent, simple factual questions answerable from context, or codebase exploration (use Explore agent instead).

Spec-first development methodology that creates detailed specifications before coding. Creates structured specs with phases, file-level details, and verification steps. Includes automatic AI review, modification application, and validation. TRIGGER when: user explicitly asks to create a spec or plan, OR model detects work that would benefit from a spec (new feature, multi-file refactor, API integration, architecture change) — in the latter case, confirm with the user before invoking. Never call mcp authoring/spec/plan tools directly for spec creation workflows. DO NOT TRIGGER when: a <command-name> tag is present in the current turn (skill already loaded), called from a subagent, single-file edits, trivial fixes, or exploratory spikes.

Review implementation fidelity against specifications by comparing actual code to spec requirements. Identifies deviations, assesses impact, and generates compliance reports for tasks, phases, or entire specs.

First-time setup for the foundry plugin (plugin:foundry@claude-foundry)

Ingest a source into the LLM wiki, or re-audit a past ingest. Use when the user asks to ingest a source ("ingest the source at raw/foo.pdf", "/wiki-ingest raw/foo.md", "add this paper to the wiki") or to audit a past ingest ("/wiki-ingest --audit-only <source>", "audit the Williamson ingest", "what did we miss in <source>?"). You compose wiki-ops primitives plus precheck, extraction, and audit judgment. The agent decides the order; common compositions are examples. Does not commit to git; the human reviews first.

Health-check the wiki's structure, provenance, and conceptual coverage. Use when the user says "/wiki-lint", "lint the wiki", "health-check the vault", "what's stale or broken?", or "audit the wiki for gaps". Runs `python3 scripts/wiki-lint.py` first, then performs the judgment-dependent lint passes described in CLAUDE.md. Does not commit to git.

Answer a question from the LLM wiki's accumulated knowledge. Use when the user asks a question that plausibly draws on the wiki ("what does the wiki say about X", "/wiki-query X", "compare Y and Z from the wiki", "summarize what we know about ..."), or asks to produce a standalone artifact (comparison, timeline, slide deck) grounded in existing pages. Reads the index and relevant pages, follows wikilinks, cites specific pages, distinguishes sourced facts from inferences, and — when the answer is reusable — files it as a new page with proper frontmatter and a log entry. Does not commit to git.

Make a scoped corrective edit to existing wiki pages when the problem is already known. Use when the user says "/wiki-repair ...", "fix the audit findings on X", "repair these attribution mismatches", "apply this corrective edit", or "clean up this known page issue" and the work is narrower than a re-ingest. Reads the affected pages plus the relevant source-summary/raw source when needed, applies the smallest defensible patch, updates bookkeeping, and does not commit to git.

Propose updates to purpose.md based on what the wiki has actually accumulated. Use when the human says "review purpose", "/wiki-purpose", "has the wiki drifted from purpose?", "update the research direction", or after a batch of ingests that may have shifted the scope. Reads purpose.md, the synthesis, the index, and recent log entries; proposes targeted edits to Goal, Key Questions, Scope, and Thesis where the wiki's accumulated state has diverged from the stated direction. Never modifies purpose.md without explicit approval — purpose.md is human-owned.

Deep research agent that searches academic databases, downloads sources, and produces structured evidence-backed reports. Use when the user asks for deep research, literature review, systematic investigation, or any question requiring multiple sources and synthesis.

Revision orchestrator that takes an existing research session and runs a structured review-then-revise cycle to improve accuracy, correctness, and clarity. Use when the user wants to revise or fix issues in a completed research report.

Research quality evaluator that analyzes session artifacts and produces structured quality assessments with scores and recommendations. Use when the user asks to evaluate, reflect on, review, or score a completed deep-research session.

Pipeline improvement planner that analyzes multiple research sessions to identify systemic patterns and produce actionable improvement plans. Use when the user wants to improve the research pipeline based on past session performance.

Autonomous single-phase implementation using Foundry MCP session-step orchestration. Runs one spec phase under unattended posture with strict gate enforcement, deterministic exits, and verification receipts. Use when executing autonomous implementation with the autonomy_runner role and session-step protocol.