Cass

cass Session Search

Table of Contents

• The Goldmine Principle
• THE EXACT PROMPT — Discovery Workflow
• Version Pinning Caveat
• Two-Step Bootstrap (Replaces "ALWAYS first")
• Stuck-Index & Recovery Decision Tree
• Quick Reference
• When to Use What
• Critical Rules
• Agent Harness Exclusion
• Search Modes
• Cross-Machine Search (Multi-Workstation Corpus)
• Anti-Patterns (Don't Do These)
• Resume a Past Session in Its Native Harness
• The Heuristics
• jq Essentials
• Hidden Power: Capabilities the Old Skill Missed
• Token & Cost Analytics (Bonus Use Case)
• Recovery Cheat Sheet (No-Permission Moves)
• Reference Index
• Quick Search (Grep Recipes for References)
• Scripts
• Validation

Core Insight: Your repeated prompts are your best prompts. If you typed it 10+ times, it works. Mine your history.

The Goldmine Principle

Your conversation history contains:

• Refined prompts — Every rephrase that worked better was captured
• Working rituals — Prompts repeated 10+ times ARE your methodology
• Scope decisions — "When did we decide NOT to do X?"
• Recovery moments — What you searched for after context loss = what mattered

The insight: Mining your past beats inventing new approaches.

THE EXACT PROMPT — Discovery Workflow

1. Bootstrap: Check health, refresh index, get project overview
   cass status --json && cass index --json
   cass search "*" --workspace /data/projects/PROJECT --aggregate agent,date --limit 1 --json

2. Find prompts: Search for keywords, filter to user prompts (lines 1-3)
   cass search "KEYWORD" --workspace /data/projects/PROJECT --json --fields minimal --limit 50 \
     | jq '[.hits[] | select(.line_number <= 3)]'

3. Follow hits: View the actual content
   cass view /path/from/source_path.jsonl -n LINE -C 20

4. Expand context: See the full conversation flow
   cass expand /path/from/source_path.jsonl --line LINE --context 3

5. Discover related: Find the whole work cluster
   cass context /path/from/source_path.jsonl --json

Why This Workflow Works

• Aggregations first — Know the terrain before diving in
• --fields minimal — 5x smaller output, preserves context window
• line_number <= 3 — User prompts live at the top of sessions
• Context clustering — Work happens in clusters; one good hit → many related sessions

Version Pinning Caveat

cass evolves quickly. The skill describes HEAD behavior (latest source in /dp/coding_agent_session_search). The released v0.3.6 binary lacks several features added since:

• cass sources agents {list,exclude,include} — added 2026-04-20 (commit 82d8d70e)
• Tail-end writer-race tolerance for full rebuilds — fixed 2026-04-22 (commit e06342f2, bead zz8ni)
• Lexical generation manifests for federated installs (commits 2b7b86a1, 683ccd03, cf76fe15)
• Rebuild producer stall telemetry (commit 73a86604)

When a flag/subcommand returns "unrecognized" or behaves differently than documented, run cass --version and check git log -- src/lib.rs for the relevant commit. Each affected section calls out which commit/version it depends on.

Probe what your installed binary actually supports:

cass capabilities --json | jq '{version: .crate_version, features, connectors}'
cass introspect --json   | jq '.commands[].name'

Two-Step Bootstrap (Replaces "ALWAYS first")

Three states matter — never conflate them.

The trap: Treating a stale index as broken triggers an unneeded full rebuild (8–25s cost) when an incremental refresh (1–3s) or even a stale-but-correct query would have worked.

# Robust two-step bootstrap that never blocks the user.
# IMPORTANT: every cass index call gets a wall-clock cap. cass index has been
# observed to hang indefinitely under contention — without `timeout`, the
# bootstrap itself becomes the symptom.
cass status --json | jq '{healthy, fresh: .index.fresh, stale: .index.stale, db: .database.exists, sem: .semantic.available}'

# Refresh policy: stale → bg refresh (capped); never block search
if [ "$(cass status --json | jq -r '.index.stale')" = "true" ]; then
  ( timeout 600 cass index --json >"/tmp/cass-index.$$.log" 2>&1 </dev/null & ) 2>/dev/null
fi
# Search even with stale index — results are still useful
cass search "KEYWORD" --workspace /path --json --fields minimal --limit 10

For the production-quality version of this logic (cap-on-every-call, broken-state escalation, exit-code semantics for hooks), use scripts/recover.sh — it implements the full decision tree with timeouts and per-PID logs.

cass health returning exit 1 on stale is a deliberate preflight signal for cron/CI. In an interactive agent loop, prefer cass status --json and decide.

Stuck-Index & Recovery Decision Tree

Real-world bugs we've hit (all observed in mined sessions). Walk top-down — first match wins.

Never paper over by skipping cass and falling back to manual find/grep. The autonomous fixes above resolve every real failure without user intervention. Asking the user to "rebuild your index" is the worst outcome — they have agents waiting.

Quick Reference

# Health + refresh (ALWAYS first)
cass status --json && cass index --json

# Project overview: who did what, when?
cass search "*" --workspace /path --aggregate agent,date --limit 1 --json

# Find keyword, minimal output
cass search "KEYWORD" --workspace /path --json --fields minimal --limit 50

# Follow a hit
cass view /path.jsonl -n LINE -C 20        # Line-oriented
cass expand /path.jsonl --line LINE --context 3  # Message-oriented

# Find related sessions
cass context /path.jsonl --json

# Export for parsing
cass export /path.jsonl --format json --include-tools -o /tmp/out.json

# Inspect or change persistent agent-harness exclusions
cass sources agents list --json
cass sources agents exclude openclaw
cass sources agents exclude openclaw --keep-indexed-data
cass sources agents include openclaw

When to Use What

Critical Rules

Agent Harness Exclusion

When a user tells you one agent harness is producing garbage, loops, or too much disk usage, handle that directly in cass instead of telling them it cannot be excluded.

# See current state
cass sources agents list --json

# Persistently stop indexing this harness in future scans/syncs/watch mode
cass sources agents exclude openclaw

# Keep already indexed data but block future indexing
cass sources agents exclude openclaw --keep-indexed-data

# Re-enable later
cass sources agents include openclaw

What exclude actually does

• Writes the preference to sources.toml, so the setting survives future runs
• Prevents future indexing even if the source files still exist on disk
• By default, purges already archived local data for that harness and rebuilds lexical search so the exclusion also reclaims space

When to use it

• A harness is spamming looped or low-value output
• A user wants cass to remember "ignore this source going forward"
• You need a reversible, agent-friendly way to reduce archive bloat without manually deleting source files

Search Modes

Default to lexical. Only use semantic when you don't know exact wording.

Enabling Semantic / Hybrid (one-time)

cass models status --json    # state: not_installed | installed | partial
cass models install          # downloads ~90MB MiniLM bundle from HuggingFace
cass index --semantic --build-hnsw --json   # builds vector + HNSW
cass search "QUERY" --mode hybrid --json    # then queries fall back to lexical if semantic missing

If the model is not_installed, --mode hybrid and --mode semantic silently fall back to lexical — no panic, no degraded experience. See SEMANTIC_AND_HYBRID.md.

Cross-Machine Search (Multi-Workstation Corpus)

When the user has agents running on css, csd, ts1, ts2, etc., the cass corpus on each machine is disjoint. Three ways to reach across:

# Option A: One-shot remote query (no setup, slow per call)
ssh css 'cass search "KEYWORD" --json --fields minimal --limit 20' | jq '.hits'

# Option B: Configured sources (preferred — caches the remote sessions locally)
cass sources setup                                      # interactive wizard, auto-discovers from ~/.ssh/config
cass sources add ssh://user@css --name css --preset linux-defaults
cass sources sync --source css --json                   # rsyncs new sessions, then re-indexes
cass search "KEYWORD" --json                            # results now span all configured sources
cass sources list --json                                # see what's wired up

# Option C: Parallel fan-out (when speed matters more than dedup)
for h in css csd ts1 ts2; do
  ssh "$h" 'cass search "KEYWORD" --json --fields minimal --limit 10' > "/tmp/cass-$h.json" &
done
wait
jq -s '[.[] | .hits[]] | unique_by(.source_path + (.line_number|tostring))' /tmp/cass-*.json

cass sources doctor diagnoses connectivity. Configured-source results carry origin_host in their hit metadata — preserve it when reporting back to the user. Full reference: REMOTE_SOURCES.md.

Anti-Patterns (Don't Do These)

Resume a Past Session in Its Native Harness

cass resume resolves a session path into the exact command its native CLI uses to continue the conversation — Claude Code, Codex, Gemini, OpenCode, pi_agent.

# Find a relevant past session
cass search "KEYWORD" --json --fields minimal --limit 5 \
  | jq -r '.hits[0].source_path' > /tmp/sess.path

# Print the resume command without executing
cass resume "$(cat /tmp/sess.path)" --shell

# Or replace the current process with the resumed agent
cass resume "$(cat /tmp/sess.path)" --exec

Pitfall: Subagent files (subagents/agent-*.jsonl) are not resumable by design — they're orchestrated by a parent. You'll get session_id_not_found with a hint to pass --agent claude. Resolve to the parent session via cass context <path> --json first. See RESUME.md.

The Heuristics

jq Essentials

# User prompts only
| jq '[.hits[] | select(.line_number <= 3)]'

# Source paths for follow-up
| jq '.hits[].source_path' -r

# Aggregation buckets
| jq '.aggregations.agent.buckets'

# Count matches
| jq '.total_matches'

# Find repeated prompts (ritual detection)
| jq '[.hits[] | select(.line_number <= 3) | .title[0:80]] | group_by(.) | map({prompt: .[0], count: length}) | sort_by(-.count) | .[0:20]'

Hidden Power: Capabilities the Old Skill Missed

If a flag listed here errors as "unrecognized" the installed cass binary is older than the source. Check cass --version against git log -- src/lib.rs upstream — most missing commands were added in v0.3.x.

Token & Cost Analytics (Bonus Use Case)

# Daily token usage across your whole agent fleet
cass analytics tokens --days 30 --group-by day --json | jq '.data.buckets[].api_tokens'

# Per-agent tool-use intensity (rows are keyed by agent slug, not tool name)
cass analytics tools --days 30 --json | jq '.data.rows[0:10]'

# Top models by API tokens
cass analytics models --json | jq '.data.by_api_tokens.rows[0:10]'

# After import / sync, refresh rollups
cass analytics rebuild --json
cass analytics validate --json   # confirm no drift between raw and rollup

cass analytics status --json shows coverage_pct: anything below ~90% means you should analytics rebuild to backfill. See ANALYTICS.md.

Recovery Cheat Sheet (No-Permission Moves)

You are pre-authorized to take any of these actions to keep the corpus searchable. None destroy source session data; they only rebuild derived index data.

cass doctor --json --verbose         # Diagnose without changing anything
cass doctor --fix --json             # Apply safe rebuilds; backs up bad DB to .corrupt.<ts>
cass doctor --fix --force-rebuild --json   # Same, but rebuild even when healthy
cass index --full --force-rebuild --json   # Workaround for OPEN issue #196 (incremental hang)
cass sources doctor --json           # Probe remote sources
cass sources sync --source NAME --json     # Re-fetch and re-index a single source
cass models install                  # Restore missing semantic model
cass models verify                   # Validate model file checksums

What you must NOT do without explicit permission: delete core.NNNNN files, delete .beads/, git reset --hard, edit anything under the user's .config/cass/sources.toml by hand. The CLI commands above already do everything safely.

Full disaster recovery for encrypted Pages archives: RECOVERY.md.

Reference Index

Quick Search (Grep Recipes for References)

When the right reference isn't obvious from titles, grep the references directory directly — cheaper than loading whole files into context.

REFS=.claude/skills/cass/references

# Find any anti-pattern by symptom keyword
grep -ni "limit 0\|broken pipe\|workspace path\|stale" "$REFS"/ANTI_PATTERNS.md "$REFS"/PITFALLS.md

# Find the recipe / jq for a task
grep -niE "ritual|user prompt|aggregate|subagent|cluster|timeline" "$REFS"/RECIPES.md "$REFS"/PATTERNS.md

# Find a recovery recipe (issue numbers, error strings, fix names)
grep -niE "doctor|--force-rebuild|issue #196|last_indexed_at|database is busy|core\." "$REFS"/RECOVERY.md "$REFS"/OBSERVABILITY.md

# Find a flag, command, or response field
grep -niE "robot-format|--mode|--cursor|_meta|fallback_mode|hits_clamped" "$REFS"/COMMANDS.md "$REFS"/INTROSPECTION.md

# Find a real "what did I ask" prompt template you've used before
grep -ni "use cass\|find that\|session history\|what worked" "$REFS"/PROMPTS.md

# Find harness/connector slugs and exclusion behavior
grep -niE "openclaw|disabled_agents|sources agents" "$REFS"/HARNESS_EXCLUSION.md "$REFS"/REMOTE_SOURCES.md

# Find resume / cross-harness behavior
grep -niE "resume|--shell|--exec|subagent.*not resumable" "$REFS"/RESUME.md

These grep across the whole references directory in <50ms and surface a line+filename that you can then open with the Read tool — far cheaper than reading the whole reference.

Scripts

Scripts live under scripts/. They contribute zero context tokens — they execute, never load. None of them mutate state without explicit confirmation.

Validation

# Quick health check
cass status --json | jq '.index.fresh'

# Should return: true

If false, run: cass index --json