HarnessX

Use this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.

How to read trajectories and extract useful experience — from successes and failures alike — into candidates for the 4 HarnessConfig levers. Covers the trajectory frontmatter schema (behaviour / eval / judge), the two axes of reflection (lens × lever), the three-variant retroactive check, and when to delegate batch reading to spawn_reflect_worker. Read at Step 3 of the loop.

Cross-round journal format — the multi-round memory that tells the next meta-agent what's been tried, what landed, and what got reverted. One `## Round N` section per evolve, with machine-parseable YAML frontmatter + free prose body. Read at the start of every evolve to avoid re-discovering doomed hypotheses; append one new section before stopping.

Component-authoring reference for HarnessX evolve rounds. Covers the @tool signature and TOOL_SPEC workflow, MultiHookProcessor (hook dispatch table, event fields, messages-mutation contract), system-prompt template editing (Cases A/B/C), config.yaml shape, and the signal→knob guide for the Configuration lever. Read when implementing any lever — Action, Control, Instruction, or Configuration.

Self-validation CLIs for the artifacts you write, plus the post-flight workflow the orchestrator runs after `end_turn`. Three categories — validity (canonicalize / dry_fire / contract / synthetic replay) blocks; policy (novelty / evidence) blocks on non-noop rounds; advisory (literals) never blocks. Read when you've just written or edited `config.yaml`, `tools/*.py`, `processors/*.py`, or `templates/*.j2`.

GAIA-specific benchmark guidance. Public-writeup-sourced techniques (markdown browser, file inspector, code-action agent, planning, multi-agent decomposition, query refinement, answer-format guard, early-commit nudge, prompt caching, majority voting) mapped to HarnessX's four levers (config / control / action / instruction), plus a catalogue of common GAIA failure modes (A-H). Use when forming a hypothesis or scanning for which intervention fits a pattern you see in trajectories.

τ²-bench-specific benchmark guidance (Sierra Research; retail / airline / telecom). Multi-turn dialogue eval with multiplicative reward (DB × action × NL-assertion) and pass^k reliability metric. Catalogues tau2 failure modes and techniques from public writeups, mapped to HarnessX's four levers (config / control / action / instruction). Use when forming a hypothesis or scanning for which intervention fits a pattern seen in tau2 trajectories.

Terminal-Bench 2 benchmark-specific structural facts for meta-agent evolution rounds. Covers sandbox topology, tool constraints, evolvable config surface, and trajectory signal layout. Read before authoring any candidate that touches the system prompt, processor pipeline, or tool registry.