skeft-qa

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).

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 contradictions, narrative overclaims, stale counts, production runs claimed as evidence without a successful backing run. Emits structured ReviewFinding records that the graph extractor consumes automatically. Finding severity determines gate impact: any finding reopens the relevant ReadinessGate and the paper cannot advance to submission until remediation lands and Stage 13 re-runs clean. <example> Context: User has completed a paper and run the full validate.py suite green. user: "Run adversarial review on paper 12 before I submit" assistant: "I'll use the adversarial-reviewer agent to audit paper 12 with fresh context." </example> <example> Context: Periodic sweep of all draft papers before a submission window. assistant: "Running adversarial review on every paper whose last-reviewed timestamp is older than 7 days." </example>

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 sentence. Covers five finding classes: internal arithmetic drift (IA), toolchain pin drift (TP), stealth pipeline-vs-prose drift (SD), theorem-name reference drift (TN), and hypothesis disclosure gap (HD). Emits reconciliation records (not silent supersession) for prior findings that don't reproduce. Invoke after updating a paper draft or before submission. <example> Context: User has updated Paper 1 with new parameter values user: "Review Paper 1 claims after the parameter corrections" assistant: "I'll use the claims-reviewer agent to walk Paper 1 sentence by sentence and verify every chain." </example> <example> Context: Proactive review before paper submission assistant: "Paper draft updated. Let me run the claims-reviewer sentence-walker pass before submission." </example> <example> Context: After a Lean refactor renames/removes theorems user: "We renamed three theorems in ScalarRungInterpretation — audit paper20" assistant: "I'll use the claims-reviewer agent; Class TN will auto-catch any stale \\texttt{...} references." </example>

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. <example> Context: User has run review_figures.py and wants figures checked user: "Review the generated figures for issues" assistant: "I'll use the figure-reviewer agent to visually inspect each figure against the manifest." <commentary> Figures have been generated and need systematic visual review before paper submission or notebook finalization. </commentary> </example> <example> Context: User has updated visualizations and wants to verify output user: "Check that the new CGL figures look correct" assistant: "I'll use the figure-reviewer agent to review the figures against their expected properties." <commentary> After modifying visualization code, visual review catches rendering issues that automated checks miss. </commentary> </example> <example> Context: Proactive review after figure generation completes assistant: "Figures generated successfully. Let me review them for rendering quality and accuracy." <commentary> Proactively invoke after figure generation to catch issues before they reach the paper or notebooks. </commentary> </example>

Consolidate harvest-extractor candidates into the System-2 register. Dispatched by the harvest skill.

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

Interactively review the System-2 dev-process harvest — promote agent-reviewed findings to human-reviewed (or close them), and triage GAP-A gate proposals. Use when you want to sign off on what the autonomous loop learned about HOW it ran.

Toggle the AskUserQuestion guard for the current managed /goal loop. Turn it off when you want the loop to be able to ask you a question; turn it back on to resume autonomous protection. Use when you want to enable or disable being asked questions during an autonomous dev loop.

The unified goal-mode skill for an autonomous /goal dev loop. Its core is the always-on operating posture (re-attaches after compaction); on invocation it composes the /goal condition AND registers the session with the harness. Use when starting a managed /goal dev loop, or when asked to write/refine a goal prompt or acceptance criteria for autonomous development.

Off-hot-loop System-2 harvest. Reads each managed /goal session's transcript from a byte-offset watermark, extracts process/harness signal (Haiku) — crossing the compact boundary for the pre-vs-post-compact delta and ingesting the guard's blocked-question log — and consolidates it (Opus) into agent-reviewed findings, refreshing the active-issues view. Invoked by a Desktop scheduled task or a second-session /loop — never inside a /goal session.

Run the full mechanical Stage-12 sync (counts, tables, deps, citation cache) in one command. Use before a wave gate, or whenever counts/tables/figures may have drifted from the Lean/Python sources.

2 hooks across 2 events