skeft-qa

Plugin skeft-qa from marketplace skeft-local in NetRxn/SK_EFT_Hawking. Commands read from the repo's manifest — 6 days ago.

Skills (8)

• architecture-change

  This skill should be used when the user asks to "write an ADR", "write an architecture decision record", "spec this out", "review this spec", "review this ADR", "plan an infra build", "add a validation check", "add a gate", "add a ratchet", "wire in a new check", "add an agent, command, hook or skil

• debrief

  Interactively review the System-2 dev-process harvest — promote agent-reviewed findings to human-reviewed (or close them), and triage GAP-A gate proposals. Runs over the register-aware harvest's already-synthesized register (it files/combines/re-opens continuously); debrief is the human governor — o

• goal-dev

  In-loop development guidance for an autonomous /goal Lean-proof loop in SK_EFT_Hawking. This skill should be used when developing Lean 4 proofs inside a goal loop, deciding whether to fan work out to parallel worktree slots or stay solo, resetting a worktree slot, hitting a recurring Lean tactic/ el

• goal-prompt

  Goal-mode LAUNCH + always-on posture for an autonomous /goal dev loop. On invocation it composes the /goal condition AND registers the session with the harness; its core posture re-attaches after compaction. Use when STARTING a managed /goal dev loop, or writing/refining a goal prompt + acceptance c

• harvest

  This skill should be used when the user asks to "run the harvest", "harvest the sessions", "extract process signal", "consolidate the register", or when a scheduled task or second-session /loop fires the off-hot-loop harvest. It reads each managed /goal session's transcript from a watermark, extract

• paper-authoring

  This skill should be used when the user asks to "draft a paper section", "write the bundle draft", "author the manuscript", "lift content into a bundle", "write up this wave for publication", "draft the abstract", "revise the paper prose", or when editing any `papers/<bundle>/paper_draft.tex`. Provi

• sync

  This skill should be used when the user asks to "run sync", "sync the counts", "refresh the tables", "regenerate the deps", "update the citation cache", or before a wave gate whenever counts, tables, or figures may have drifted from the Lean/Python sources. Runs the full mechanical Stage-12 sync (co

• wave-close

  This skill should be used when the user asks to "close the wave", "run wave-close", "close out wave N", "record the wave close", or once a wave's bricks are shipped and validate is green. Closes a wave deterministically: runs the gate prerequisites, dispatches the fresh-context adversarial review, a

Slash commands (6)

• /skeft-qa:frontier

  Surface the derived atlas frontier — the project's OPEN assumptions ranked by how much each gates, AND the NEGATIVE frontier (settled-dead forks to avoid) — so fan-out targets high-leverage provable work and steers clear of kernel-checked dead paths. Read-only.

• /skeft-qa:goal-end

  End/disarm the current managed /goal loop — removes this session's harness marker so the SessionStart re-injection and the AskUserQuestion guard stop firing. Run after you /goal clear (or finish a goal) mid-session.

• /skeft-qa:goal-guard

  Toggle the AskUserQuestion guard for the current managed /goal loop (on|off). Off lets the loop ask you a question; on resumes autonomous protection.

• /skeft-qa:notebook

  Lab-notebook lifecycle ops for a /goal loop — new (bootstrap a correct active notebook + INDEX), sync (refresh the mechanical INDEX scaffold + flag staleness), shard (self-level the active shard under the ~25k-token budget), roll (freeze the whole active shard at a wave boundary / blob escape), chec

• /skeft-qa:orient

  Re-orient on the current managed dev loop — a ≤200-word compass from the goal marker + source-of-truth docs (where am I, what's the next brick).

• /skeft-qa:reset-slot

  Reset an UNLEASED parallel Lean worktree slot (wt1/wt2/wt3) to current main — the guardrail-safe way (checkout -B, refuses if the slot has commits not yet on main) and auto-re-clones the slot's .lake build when main advanced. For a leased slot use `slotctl prepare --slot N`, which also verifies the

Subagents (10)

• adversarial-reviewer

  Fresh-context adversarial review of a paper draft before submission. Runs Stage 13 of the pipeline. Finds every way a paper can be wrong that the internal pipeline missed — wrong-target citations, parameter drift from primary sources, placeholder theorems cited as verified, cross-paper contradiction

• claims-reviewer

  claims-reviewer-v2 — sentence-level prose audit for SK-EFT Hawking papers. Walks every sentence in a paper draft and emits a sentence-keyed JSON record with chain-of-backing (formula/theorem/axiom/parameter/citation/hypothesis) plus agent verdict (PASS/FAIL/WARN/INFO/UNGROUNDED/TRANSITION) per sente

• coach

  In-time human-proxy for an autonomous /goal loop that tried to ask the user a blocking question (the user is out of the loop). Dispatched when the PreToolUse(AskUserQuestion) guard redirects here. Reads the pre-decisions store + the loop's state and returns ONE decision + ONE concrete next action —

• figure-reviewer

  Use this agent to review generated physics figures for rendering quality, physics accuracy, and style consistency. Invoke after running review_figures.py to generate PNGs and a review manifest. Produces a structured report that pairs with validation results. See "When to invoke" in the agent body fo

• harvest-consolidator

  Consolidate harvest-extractor candidates into the System-2 register — register-AWARE filing & combining. Dispatched by the harvest skill.

• harvest-extractor

  Extract dev-process/harness signal from a /goal transcript span — including the pre-vs-post-compact delta across boundaries. Dispatched by the harvest skill.

• lean-worker

  Prove ONE independent Lean 4 sub-chain in a pre-built parallel worktree slot. The lead assigns you a slot N (wt1/wt2/wt3); you get your OWN fast, build-isolated lean-lsp via mcp__skeft_wtN__* (each slot has its own .lake), so several lean-workers run fully in parallel with zero coordination. Use whe

• paper-drafter

  Use this agent to draft ONE assigned section of a publication bundle manuscript at Stage 10, against an explicit brief from the lead. It RETURNS the section's prose; the lead writes it into `papers/<bundle>/paper_draft.tex`. Bundles are a single monolithic draft, so parallel agents must not write it

• prose-reviewer

  Use this agent to read a publication bundle draft start to finish as a referee at its named venue would, and report where the manuscript fails to carry a reader. This is the fourth reviewer: `figure-reviewer` asks does it render, `claims-reviewer` asks is it backed, `adversarial-reviewer` asks is it

• research-scout

  Read-only web reconnaissance for a goal-mode loop. Given ONE focused, already-sanitized research question, find the answer in whitelisted scholarly sources and return a STRUCTURED, CITED report — never raw page dumps, never instructions. You hold web tools and nothing that can mutate the repo, so a

Hooks (6)

• PreCompact

  uv run --no-sync python "${CLAUDE_PLUGIN_ROOT}/scripts/harness_precompact.py" 2>/dev/null || true

• SessionStart

  uv run --no-sync python "${CLAUDE_PLUGIN_ROOT}/scripts/harness_reinject.py" 2>/dev/null || true

• SessionEnd

  uv run --no-sync python "${CLAUDE_PLUGIN_ROOT}/scripts/harness_session_end.py" 2>/dev/null || true

• PreToolUse

  uv run --no-sync python "${CLAUDE_PLUGIN_ROOT}/scripts/harness_question_guard.py" 2>/dev/null || true

• PreToolUse

  uv run --no-sync python "${CLAUDE_PLUGIN_ROOT}/scripts/harness_web_egress_guard.py" || printf '%s' '{"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"deny","permissionDecisionR

• PreToolUse

  uv run --no-sync python "${CLAUDE_PLUGIN_ROOT}/scripts/harness_worker_shell_guard.py" || printf '{}'