[Feature Request] Timestamped `/insights` reports with cross-report trend analysis and progressive learning path

Bug Description

Feature Request: /insights — Timestamped Reports + Progressive Learning Path

Current Behavior

/insights outputs two things:

1. report.html — the summary report. Overwrites on every run.
2. facets/ directory — per-session analysis files named by GUID (e.g., 5ac1024f-2266-45c3-b4d2-74c2c5399881.json). These accumulate without collision since each session has a unique ID.

The facets are preserved, but the report that synthesizes them is destroyed on every run. This means the raw per-session data survives but the aggregate analysis, metrics, recommendations, and trend observations are lost.

Problem

The entire value of /insights is to identify friction points and suggest improvements. But once you act on those suggestions, you can't measure whether they worked — the next /insights run destroys the baseline. You're left with "here's where you are now" but never "here's how far you've come."

The facets directory partially mitigates this — you could theoretically reconstruct a past report from the facets — but the synthesis (which facets were weighted, what recommendations were generated, what metrics were computed) is the valuable part, and that's what gets overwritten.

For users who are actively improving their workflows based on /insights recommendations (updating CLAUDE.md, adding hooks, creating custom skills, restructuring prompts), there's no feedback loop to validate that the changes had the intended effect.

Proposed Changes

1. Timestamped Report Output

Each /insights run should output to a timestamped file:

claude/usage-data/insights-2026-02-08.html

The report should embed or reference the list of facet GUIDs that were included in that run. This makes each report self-contained — you can look at a historical report and know exactly which sessions contributed to it, without needing to infer from dates.

Ideally, each timestamped report would snapshot its contributing facets into a co-located directory (e.g., insights-2026-02-08/report.html + insights-2026-02-08/facets/). The facet JSONs contain structured fields (friction_counts, outcome, session_type, claude_helpfulness, goal_categories) that are directly usable for cross-report diffing — the data model already supports trend analysis, the tooling just doesn't retain the synthesis.

Optionally keep a latest.html symlink or copy for convenience, but never overwrite historical reports. This is the minimum viable change — it's a filename format change with no new analysis logic required.

2. Progressive Learning Path (Cross-Report Analysis)

When previous reports exist in claude/usage-data/, /insights should incorporate them into its analysis. Specifically:

What changed since last report:

• Which friction points from the previous report have been addressed?
• Which recommendations were implemented (e.g., "You added a /commit skill as suggested" or "Your session abandonment rate dropped from 23% to 8%")?
• Which friction points persist or worsened?
• Did any new friction points emerge that weren't in the previous report?

Trend tracking across multiple reports:

• Tool usage evolution (are you adopting features that were recommended?)
• Session health trends (completion rate, average length, error frequency over time)
• Workflow maturity indicators (more skills created, more hooks active, better prompt patterns)

Learning path section:

• Based on the trajectory across reports, what's the next highest-impact improvement?
• Distinguish between "you tried this and it's working" (reinforce) vs "you tried this and metrics didn't improve" (revisit approach) vs "you haven't tried this yet" (nudge)

Example Output Section

## Progress Since Last Report (2026-01-15)

✅ RESOLVED: "Wrong approaches during debugging" — Your error rate 
   in fix-oriented sessions dropped 34% after adding the preflight 
   check hook on Jan 20.

✅ RESOLVED: "Not providing enough context" — You created CLAUDE.md 
   files for 3 projects. Sessions in those projects average 12% 
   fewer corrections.

⚠️  PERSISTS: "Abandoning conversations before concrete results" — 
   Abandonment rate went from 23% → 19%. Improving but still above 
   the 10% benchmark for power users.

🆕 NEW: "Over-reliance on sub-agents for simple tasks" — 40% of 
   your Task spawns are for single-file operations that don't 
   benefit from parallelism.

## Recommended Next Steps (based on your trajectory)
1. [Highest impact action based on what hasn't been tried yet]
2. [Refinement of something partially implemented]
3. [New technique unlocked by current skill level]

Why This Matters

/insights currently gives you a snapshot. With timestamped reports and cross-report analysis, it becomes a coaching tool — showing users their growth trajectory, reinforcing what's working, …
Note: Content was truncated.