Agent Skills: Session Report — Intent-Based Report Generator

Session Report — Intent-Based Report Generator

Generate a polished, self-contained HTML report based on this session's work. The workflow adapts to what the operator needs rather than forcing a fixed section structure.

Follow these four phases in order. Do not skip phases.

Phase 1 — Intent

Ask the operator what kind of report they need. Present ONE AskUserQuestion with a single-select of report intents:

| Intent | Description | |--------|-------------| | Session Log | Chronological record of what happened — channels read, values observed, tools used. Minimal interpretation. | | Analysis Report | Technical deep-dive with charts, data tables, and narrative observations about patterns and anomalies. | | Executive Briefing | High-level summary for management — KPIs, status cards, key observations in non-technical language. | | (Other) | Operator describes a custom report structure. |

Do NOT ask follow-up questions after this unless the operator response is ambigious.

Phase 2 — Inventory & Structure

2a. Call session_summary

Use the session_summary MCP tool to get a compact inventory of all data context entries and artifacts in the workspace. This tells you exactly what data is available.

2b. Draft dynamic structure

Based on the operator's intent and the inventory, draft a report structure. There is no fixed section list — the structure should be driven by what data exists and what the intent calls for.

Guidance (not constraints) on block types that work well:

• KPI row — Good for any intent when you have countable metrics (channels read, values observed, artifacts created)
• Data table — Natural for Session Log intent; useful whenever channel values were read
• Chart.js line/scatter — Natural for Analysis Report when timeseries data exists. Use archiver_downsample to get chart-ready data (never embed raw timeseries).
• Narrative paragraphs — Observations about what was seen. Essential for Analysis Report and Executive Briefing.
• Card grid — Good for Executive Briefing status overview; good for linking to artifacts
• CSS Timeline — Good for Session Log to show a simple chronological list of actions (timestamps + descriptions)
• Mermaid event timeline — Use when the session involved grouped or multi-phase event sequences. Mermaid's timeline diagram type renders events organized by time periods with grouping, making complex sequences easier to scan than a flat list. Prefer this over the CSS timeline when there are 5+ events or natural phase groupings.
• Mermaid state diagram — Use when the session revealed state transitions in a control system (e.g., beam states, interlock chains, operational modes). Mermaid's stateDiagram-v2 renders state machines with transitions and guards. Only include when actual state changes were observed — never fabricate state models.
• Collapsible details — Good for Analysis Report when there's dense technical data to organize

Keep it proportional: A 10-minute session with 2 channel reads doesn't need 8 sections. A 2-hour investigation with archiver data, plots, and multiple analyses might warrant a rich structure. Match the report's complexity to the session's complexity.

2c. Gather chart data (if needed)

If the structure includes charts and the inventory shows timeseries data, call archiver_downsample for each relevant data context entry now. This gives you chart-ready payloads (labels + datasets) that fit inline.

Phase 3 — Delegate

Spawn a Task subagent to generate the report. Pass it:

1. Intent — the operator's chosen intent
2. Inventory — the session_summary output
3. Structure — your drafted report structure
4. Chart data — any archiver_downsample results (if applicable)
5. Conversation context — key observations and findings from this session
6. Safety rules — the content safety rules below

The subagent should:

• Read the reference file at .claude/skills/session-report/reference.md for CSS/JS patterns
• Use workspace MCP tools (artifact_save, session_log, archiver_downsample) as needed
• Generate a single self-contained HTML file and save it via artifact_save
• Block access to Bash, Read, Write, Edit (subagent uses only MCP tools)

Content safety rules (MUST be included in subagent prompt)

CRITICAL — Observation-only reporting

This report documents what was observed during the session. It must NEVER generate recommendations, prescriptive advice, or action items.

• ALLOWED: "Beam current was observed at 302.1 mA" / "A downward trend was noted between 14:00–15:00" / "The value exceeded the nominal range"
• FORBIDDEN: "Investigate the corrector magnets" / "Consider adjusting the RF frequency" / "It is recommended to..."

Exception: If the operator explicitly stated an action item during the session (e.g., "I need to check the vacuum pump tomorrow"), it may be included as an attributed quote: "Operator noted: 'Need to check vacuum pump tomorrow'".

HTML requirements (pass to subagent)

• Single <!DOCTYPE html> file — all CSS in <style>, all JS before </body>
• Responsive sidebar TOC (desktop) / horizontal bar (mobile) using the reference pattern
• Light AND dark theme via prefers-color-scheme media query
• Pick a font pairing from the rotation table in the reference — never Inter/Roboto/Arial
• Use depth tiers: hero for header, elevated for KPIs, default for content, recessed for code/details
• Staggered fadeUp and fadeScale animations with prefers-reduced-motion respect
• Overflow protection globals (min-width: 0, overflow-wrap: break-word)
• Google Fonts via CDN with display=swap
• Chart.js via CDN (only if charts are included)
• Footer with generation timestamp and "Generated by OSPREY Session Report"
• Data integrity: every number, channel name, and timestamp must come from actual session data. Never fabricate values.

Phase 4 — Register & Open

After the subagent returns:

1. Focus the artifact using artifact_focus so it appears in the gallery
2. Open in browser using Bash: open <artifact_path>
3. Confirm to operator: Report generated, list key sections, provide file path

Anti-Patterns

Do NOT:

• Ask multiple rounds of questions — Phase 1 is ONE single-select, then generate
• Generate recommendations or prescriptive advice (see safety rules above)
• Over-engineer simple sessions — match report complexity to session complexity
• Fabricate data — every value must come from the actual session
• Embed raw timeseries data — always use archiver_downsample for charts
• Use emoji anywhere in the report
• Use Inter, Roboto, Arial, or plain system-ui as the primary font
• Skip reading the reference file — patterns drift from memory
• Use a generic dark theme — both light and dark must look intentional
• Omit the prefers-reduced-motion media query
• Hardcode colors instead of using CSS custom properties
• Skip the overflow protection globals