Reflect Tree

/reflect-tree — session quest-tree

Produce a single-file interactive HTML quest-tree of a Claude Code session. The vertical spine is the linear sequence of turns; back-edges and off-spine nodes encode wrong turns (corrections, retries, reversals, dead-ends, backtracks) so the user can see the whole journey at a glance — including detours and where the agent had to redo work.

This is a sibling to /reflect. Same input contract, different output: /reflect is a two-pane incidents/recommendations dashboard; /reflect-tree is a graph.

When to use

• User invokes /reflect-tree (no args) → analyze the current session from the in-context conversation. Do not re-read the session JSONL — work from the agent's own memory.
• User invokes /reflect-tree <session-id> or /reflect-tree <path-to-jsonl> → run scripts/analyze_session.py <path> directly (this skill runs in a forked context — context: fork — so the compressed transcript is safe to ingest). The script outputs a compact markdown transcript: system reminders stripped, tool calls/results collapsed to one-liners, compaction blocks expanded with embedded user quotes. Then you classify each turn — the script does NOT classify, it only compresses.

Session JSONLs live under: ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl

Node classifications

Each turn becomes a node. Pick the strongest applicable label:

• normal — routine forward progress (user request or assistant tool call/result with no problem)
• correction — user pushed back on an approach ("no", "don't", "stop", "actually")
• retry — agent ran the same tool/intent 2+ times with variations before it worked
• reversal — agent edited then unwound (Edit → revert, Write → delete)
• dead-end — tool failed because of the environment (missing binary, wrong path, OS mismatch)
• backtrack — agent abandoned a path and resumed from an earlier state
• self-correction — agent caught its own mistake mid-stream

Edge types

• flow (default) — solid thin line between consecutive turns
• back-edge — curved dashed line from a retry/reversal/correction node back to the ancestor turn it relates to (color matches the classification)
• backtrack — curved solid arrow showing where the agent jumped back to
• wasted-segment — thicker grey edge with a T12–T18 · 7 turns label when many turns of fruitless searching collapse into one capsule

Each non-normal node carries a refs array listing the ancestor turn IDs it relates to.

Knowledge mining (per-node insight)

Every node must carry an insight block — the single most important, reusable lesson from that turn, framed so a future agent could drop it into a prompt or rule. This is the value of the tree: each node becomes copyable knowledge.

For each turn, extract:

• highlight (1 line, ≤120 chars) — the single most important fact/decision/lesson. The "if you only read one thing from this turn" sentence.
• insight (2–5 lines) — the why behind the highlight. What the agent learned, what was non-obvious, what a future agent should do differently or keep doing. Phrase it as durable advice, not a play-by-play.
• prompt_snippet (copyable text) — a self-contained chunk a user could paste into CLAUDE.md, a memory file, or a future prompt. Must read independently of the tree (no "see T5" references). Format depends on the classification:
  • correction / dead-end → a rule: - Always X, because Y or a <rule> block
  • retry / waste → a recipe: the canonical command/path/tool that worked
  • reversal / backtrack → a guard: "Don't edit X — edit Y instead" with the reason
  • self-correction → a heuristic: "Before doing X, check Y"
  • normal (only the load-bearing ones — request, key decision, final answer) → a fact: the goal statement, the chosen approach, the outcome

Skip prompt_snippet for filler normal turns (routine Reads/Greps with no insight). Better to have 6 strong insights than 50 weak ones.

The HTML drawer renders these per-node with a Copy insight button that copies prompt_snippet to the clipboard. A header-level Copy all insights button assembles every node's prompt_snippet into one markdown document grouped by classification.

Visual language

X-position: normal = center spine; correction/reversal nudged left; retry/dead-end nudged right. This keeps the spine readable while making detours visually distinct.

A legend is rendered inline in the header.

Collapsed-run capsule

When grouping consecutive normal turns into a capsule, render the entire label inside the rounded rect — turn-range, count, and topic together. Do not place the turn-range outside the box; it looks broken when the rect background only covers part of the label.

Pattern (SVG):

<g class="capsule" transform="translate(300 380)">
  <rect x="-150" y="-13" width="300" height="26"/>
  <text x="0" y="0">
    <tspan class="turn-range">T120–T139</tspan> · 11 turns · refactor + INFRA_GUIDE.md
  </text>
</g>

The rect width must accommodate the full text. Center the <text> (anchor middle, dominant-baseline middle) so it sits inside the rect. Use rx/ry ≥ 8 for the pill shape. Style the turn-range inline with <tspan class="turn-range"> so it stays bold/foreground while the rest of the label is muted.