Skip to content

INSIGHTS.md

Latest commit

 

History

History
101 lines (74 loc) · 3.77 KB

INSIGHTS.md

File metadata and controls

101 lines (74 loc) · 3.77 KB
Raw

Corpus insights

Aggregate statistics from a local harvest. No transcript text or PII — counts and distributions only.

Note: Numbers below are a snapshot from one checkout. Run pnpm insights on your machine after split; repo names come from your repo_hint values, not a fixed list.

Scale

Metric Value
Manifest rows 74
Unique session_id values 36
Session files with duplicate paths (same id, multiple ingest) 29
Parent session JSONL files 57
Subagent JSONL files (…/subagents/*.jsonl) 17
Total raw JSONL lines (manifest sum) 11,282
Total raw bytes (manifest sum) ~21 MB

After pnpm normalize with default --source host (dedup by session_id within source):

Metric Value
Processed turns 675
Dropped turns (no assistant text) 30
Unique session files (deduped) 36
Duplicate files skipped 1

With --source all (host preferred over devcontainer): same ~675 turns, ~33 duplicate files skipped across sources.

Ingestion sources

source Manifest rows
host 42
devcontainer 30
manual 2

Duplicate ingestion: The same session_id often appears under manual, flat host/<date>/, workspace-scoped host/<date>/<slug>/<id>/, and again under devcontainer/. Normalizing with --source all without dedup previously inflated turn counts (~2×). Default --source host plus per-session_id file picking fixes this.

Repo breakdown (repo_hint)

Repo Sessions (manifest rows)
thepulimaangani 24
devprofile 10
premflow 8
elomaxz 8
adaptate 4
agent-prompt-tuning-lab 2
adaptate-packages-core 2
ask-grok-extension 2

Target repos for gold tagging: whichever repo_hint values dominate your eval/pool splits (see GOLD_SESSIONS.md for an example list).

Turns and tools (processed, pre-dedup baseline)

From a full raw pass (all sources, no dedup — useful as an upper bound):

Metric Approx.
User turns emitted 1,273
Turns with had_attached_skills 8 (~0.6%)
Turns with tool_call_count > 0 1,148 (~90%)
Turns with [REDACTED] in assistant text 1,141 (~90%)

Top tools (by turn presence)

Tool Turns
Shell 757
Read 731
StrReplace 533
Grep 519
Glob 270
Write 266
ReadLints 199

Redaction

Layer Rate
Raw assistant lines containing [REDACTED] ~78% of assistant JSONL lines
Processed turns with redacted assistant text ~90% of turns

Most assistant content is stored redacted in Cursor exports. Normalization keeps the last non-redacted assistant chunk per user turn; turns with only redacted assistant output go to dropped.jsonl.

Subagents

  • 17 subagent files in manifest; parent id is the folder name immediately above subagents/ (see parent_session_id in SCHEMA.md).
  • Subagents are typically shorter (5–16 lines) than parent sessions (15–1,500+ lines).

Skill attachment

Skill names are extracted from manually_attached_skills blocks in user messages (bodies not stored in processed output). Attachment rate is low in this corpus; longer devprofile sessions are more likely to include skills.

Recommendations for contributors

  1. Run harvest on the host where ~/.cursor lives; use pnpm normalize (default host) to avoid double-counting.
  2. Tag exemplar sessions with pnpm tag-manifest --tag gold --session-id <uuid>; record ids in docs/GOLD_SESSIONS.md.
  3. Treat manifest and raw/processed trees as local-only (gitignored).
  4. Before sharing splits, scan for secrets and strip paths that did not normalize to {REPO_ROOT}.