ContextSearch

Finds prior PAI work — sessions, ISAs, conversations — by topic, partial words, or date phrases like "yesterday" or "last week". A deterministic Bun CLI searches five sources in parallel, scores by token-overlap times recency, applies date filters, and returns ranked results with snippets.

Work spreads across sessions, ISAs, and conversation logs, and a week later you can't remember where a project lived or what it was called. Substring search misses it because you remember the gist, not the exact title. You need to recall half-remembered work fast — by a few words, or by "the thing I did yesterday" — and pick up where you left off without rereading everything by hand.

Search prior work for: $ARGUMENTS

The skill body delegates to a deterministic Bun CLI. The tool searches five sources in parallel, scores by token-overlap × recency, applies date filters, and returns ranked results.

bun run ~/.claude/skills/ContextSearch/Tools/ContextSearch.ts "$ARGUMENTS" --pretty --limit 10

The tool auto-detects TTY for output mode; explicit --pretty keeps the human-readable block when the skill is invoked from inside a Claude Code session (where stdout is not a true TTY). Pipe to --json for structured output to jq.

# Limit results
bun run ~/.claude/skills/ContextSearch/Tools/ContextSearch.ts "$ARGUMENTS" --limit 5

# Constrain by date
bun run ~/.claude/skills/ContextSearch/Tools/ContextSearch.ts "$ARGUMENTS" --since 2026-05-01

# JSON for programmatic use
bun run ~/.claude/skills/ContextSearch/Tools/ContextSearch.ts "$ARGUMENTS" --json | jq '.results[0]'

The tool also parses date phrases inline: "yesterday markdown" becomes a bounded since/until of yesterday only, plus token search for markdown.

1. Standalone — Run the search, present the pretty block, say: "Context loaded on [topic]. Most recent: [X]. What would you like to do?"
2. Paired with request — Run the search first, then execute the request informed by found context. If a result's path looks promising, Read it for full detail.

The pretty block:

═══ CONTEXT: <query> ═══════════════════════

📋 RESULTS (N, newest+best first):
  • <slug>
    task: ... | phase: ... | progress: ... | effort: ... | <recency>d ago | score: ... | sources: ...
    ↳ "<snippet from matched source>"
    <path to ISA or jsonl>

Searched: work.json=N, names=N, dirs=N, isa=N, jsonl=N
════════════════════════════════════════════════

JSON: { query, cleaned_query, tokens, since, until, stats, results: [{ source[], slug, sessionUuid, name, task, phase, progress, effort, score, recencyDays, snippet, path }] }

• Token-overlap scoring, not substring matching. Queries like markdown extended find sessions named Research Type Markdown Extended Images even though "markdown extended" is not a substring of that name. The query is tokenized (lowercase, [^a-z0-9]+ split, stopwords removed) and any candidate that shares ≥1 non-stopword token gets scored.
• Date phrases are first-class. today, yesterday, day before yesterday, last week, last month, N days ago, and YYYY-MM-DD parse out of the query as bounded date filters; the remaining tokens still score against content. yesterday and today use single-day windows (since AND until); last week uses a 7-day rolling window.
• JSONL-source date is the first user-message timestamp, not file mtime. Conversation jsonl files get touched whenever Claude Code re-reads them, so mtime drifts. The tool parses the first type:"user" line's timestamp to recover the actual session start date — that's what date filters compare against.
• Search is restricted to ~/.claude/projects/${HARNESS_USER_DIR}/ for jsonl content. Other Claude project directories on the same machine are out of scope by design.
• No persistent index. Every query rescans from disk. ripgrep over ~9k jsonl files completes in ~1s; total query latency is typically 1–2s, well within interactive perception. If a future scale change makes that unworkable, an opt-in cached index (e.g., per-jsonl first-user-timestamp + word-frequency map) is the v2.
• Standalone vs paired. When the user invokes ContextSearch alone (/cs <topic>), present results and ask what they want to do. When invoked as a paired primer to an unrelated request, run the search silently and use the top results to anchor the answer — don't dump the search block into the response unless the user asked.
• Session-names.json date is jsonl mtime (best-available fallback). Some session UUIDs don't appear in work.json or as MEMORY/WORK/ directories — they're conversation-only sessions whose only structural trace is the auto-generated name. For those, the tool uses jsonl mtime for recency. If the file's been re-read recently the date will skew newer than the actual conversation.

Standalone: "Context loaded on [topic]. Most recent: [X]. What would you like to do?"

Paired: Proceed with the user's request, anchored by the top result's context. If deeper detail is needed, Read the specific path the tool returned.

Example 1: Resume work on a specific topic

User: "what did we do with the Telegram bot?"
→ bun run Tools/ContextSearch.ts "telegram bot"
→ Tokens [telegram, bot] hit work.json + ISA bodies + jsonl
→ Returns telegram-monitor-revival, fix-telegram-channels-plugin-broken, ranked by recency

Example 2: Find half-remembered session

User: "open up that test extended markdown document that you made yesterday"
→ bun run Tools/ContextSearch.ts "test extended markdown document yesterday"
→ Date phrase 'yesterday' parses out → since/until = 2026-05-07 only
→ Tokens [test, extended, markdown, document] hit jsonl content
→ Returns "Research Type Markdown Extended Images" session in top-3 with snippet

Example 3: Date-only browsing

User: "what did I work on last week?"
→ bun run Tools/ContextSearch.ts "last week"
→ Date phrase parses to since=7d-ago, no token filter
→ Returns all sessions from the past 7 days, newest first