Mine past agent sessions for working prompts, decisions, and patterns (session archaeology). Triggers: "cass", "mine past agent sessions for", "cass skill".
Mine past agent sessions for working prompts, decisions, and patterns (session archaeology). Triggers: "cass", "mine past agent sessions for", "cass skill".
practices
["pragmatic-programmer"]
skill_api_version
1
user-invocable
false
hexagonal_role
supporting
consumes
[]
produces
[]
context_rel
[]
output_contract
stdout: cited session hits (source_path + line) and index-state facts; no artifact directory unless an explicit -o export path is given
metadata
{"dependencies":[],"capabilities":["cass"],"effects":["rebuild_local_index","sync_remote_sources","download_semantic_model"],"canonical_status":"canonical","disposition":"keep_specialist","tier":"execution","external_dependencies":["cass binary (>=0.3.6 recommended; some commands require HEAD — see Version Pinning)","jq (required for parsing --json output)","GNU coreutils 'timeout' (recommended; cass index can hang under contention)","ssh + rsync (optional; only for cross-machine `cass sources` workflows)","fastembed model bundle ~90MB (optional; only for --mode semantic / hybrid; install via `cass models install`)"]}
cass Session Search
Core Insight: Your repeated prompts are your best prompts. If you typed it 10+ times, it works. Mine your history.
cass is an upstream (Dicklesworthstone) tool and is self-describing — do not re-learn its surface from this skill. Discover it live:
cass capabilities --json # features/connectors/limits of the installed binary
cass introspect --json # full schema of every command + response
cass robot-docs guide|commands|examples|schemas|contracts # machine-targeted docs
This skill carries only the AgentOps operating doctrine: when to reach for cass, the discovery workflow, the recovery posture, and the anti-patterns we have actually hit.
Constraints
Never run bare cass because it launches a blocking TUI; use a JSON, robot, or explicit file-output command.
Treat a stale index as searchable and refresh it with a bounded background command because stale is not broken and an unbounded rebuild can stall the lane.
Preserve source sessions and require explicit permission for destructive cleanup; recovery may rebuild only derived index state.
When to Use
"What did I ask last time?" / "find that prompt that worked" — session archaeology
Prior-art check before inventing a new approach, plan, or prompt
Scope archaeology: "when did we decide NOT to do X?"
Post-context-loss recovery: what was searched for after a crash = what mattered
Folded triggers (ag-s43tg wave 1): casr + cass-memory route here
casr → cross-harness resume. Cross Agent Session Resumer: convert and resume sessions across Claude Code, Codex, Gemini, and other providers — cass resume plus RESUME.md own this lane (resolve subagent logs to their parent via cass context first; subagent files are not resumable).
cass-memory → cm procedural memory. Use when starting non-trivial work, mining lessons, or preventing repeated mistakes with cm procedural memory — mine past sessions here first, then promote the durable lessons through cm instead of re-deriving them each session.
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. CASS is a context source in the federated graph: it supplies cited episodic evidence on demand — mine as a research-phase move before writing a fresh plan or prompt. What it returns is evidence with source identity and freshness, never policy, and AgentOps maintains no merged corpus of its own around it.
History-First Routing
Before deriving a plan, prompt, or approach from scratch, run one bounded
search of past sessions (one query family, --fields minimal, a real
--limit, under a minute of wall clock). Three outcomes, each with its own
routing:
Direct hit — a prior session solved this. Reuse its prompt or decision;
cite source_path and line in whatever you build on it.
Adjacent hit — prior work borders the problem. Extract the working
fragments, then derive only the missing part fresh.
Verified absence — zero hits after retrying against discovered workspace
keys (--aggregate workspace). Now derivation is justified, and the absence
itself is worth noting: you are in new territory, so budget accordingly.
The named failure mode is re-derivation drift: solving the same problem
slightly differently each session, so the corpus accumulates near-duplicate
approaches and no single one ever hardens into a ritual. Stop condition for
the history pass itself: one query family exhausted or a direct hit found —
history search is a bounded pre-step, not an open-ended excavation that
displaces the actual task.
Lesson Weighting: Decay and Failure Overweight
Mined lessons are evidence with a shelf life, not doctrine:
Confidence decays with corpus drift. Weight a mined lesson by what has
changed since it was captured, not by calendar age alone. A lesson about a
tool surface or repository that has since moved is a hypothesis to re-verify
— one probe against the current surface — before it steers a fresh plan. A
lesson about durable method (how to decompose, how to verify) decays far
more slowly. Never carry a stale-surface lesson forward at its original
confidence; the named failure mode is fossil doctrine — a dead workaround
reapplied for months because it once worked and nobody re-checked.
Overweight failures. A session where an approach failed is worth more
than a session where one worked: successes are overrepresented in what gets
polished and remembered, while failures encode the boundary of validity.
When mining prior art for an approach, explicitly search for its failures
("didn't work", "reverted", "gave up", error strings) before adopting it. A
hit showing the approach failing in circumstances like yours outranks three
hits showing it succeeding elsewhere.
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 it works: aggregations first (know the terrain), --fields minimal (5x smaller output), line_number <= 3 (user prompts live at the top), context clustering (one good hit → many related sessions). >10 matches for a prompt = a ritual; document and reuse it.
Operating Doctrine: Stale ≠ Broken
Three index states matter — never conflate them:
State
Meaning
Do
cass health exit 0
Healthy
Search immediately
stale (index.stale=true)
Usable but old
Search NOW; refresh in background with a wall-clock cap: ( timeout 600 cass index --json &>/tmp/cass-bg.log </dev/null & ) — NEVER a bare &, cass index can hang
broken (database.exists=false or documents=0)
Truly uninitialized
cass doctor --fix --json, then cass index --full --json
The trap: treating stale as broken triggers an unneeded 8–25s full rebuild when a 1–3s incremental (or a stale-but-correct query) would do. scripts/recover.sh implements the full decision tree with timeouts. Detailed symptom→fix tables (issue #196 hang, stale locks, database is busy race, etc.): RECOVERY.md, OBSERVABILITY.md, PITFALLS.md.
Incremental refresh can become authoritative
An invocation requested as cass index --json may discover that incremental
state cannot be reconciled and expand into an authoritative rebuild over the
full conversation corpus. A large total or a longer run is evidence of recovery
mode, not evidence that source sessions were lost. Do not start a second
indexer or report a zero-result search while the first call is still converging.
Concurrent status and search reads can exceed their normal latency during that
rebuild. Bound observations with a wall-clock timeout, retain the exit status,
and distinguish timeout from an empty result:
status_rc=0
timeout 15 cass status --json > /tmp/cass-status.json || status_rc=$?
# Exit 124 means status was not observed within the cap.
search_rc=0
timeout 30 cass search "QUERY" --json --fields minimal --limit 20 \
> /tmp/cass-search.json || search_rc=$?
# Exit 124 is NOT zero hits; retry after the rebuild settles.
If status returns, inspect .rebuild.active, .rebuild.phase,
.rebuild.processed_conversations, .rebuild.total_conversations, and
.rebuild.updated_at. When updated_at or processed count advances, wait for
that bounded run rather than stacking recovery. If a bounded read times out,
report only that the read was not observed within the cap and keep the last
known freshness; do not infer corruption or data loss. See
OBSERVABILITY.md.
Version Pinning
cass evolves quickly; the released binary may lack HEAD features. When a flag returns "unrecognized", do not guess — probe: cass capabilities --json and cass introspect --json | jq '.commands[].name', and check cass --version.
Anti-Patterns (Don't Do These)
Anti-pattern
Why it's wrong
Do instead
Asking the user "should I rebuild the index?"
They have agents waiting; rebuild is safe and idempotent
Just run cass doctor --fix --json (preserves source data)
Running cass index --full whenever status says unhealthy
A 25s rebuild for a 30-min stale index is wasteful
Check index.stale separately from database.exists; prefer incremental
Running bare cass to "see what's there"
Launches blocking TUI in the agent's session
Always --json or --robot; never bare
Piping cass export into head/jq
Broken-pipe panic on large sessions
cass export ... -o /tmp/x.json first, then operate on the file
Treating subagent files as parent sessions
Subagents are separate logs with their own line-2 prompt; also NOT resumable
Filter by select(.source_path | contains("subagent")); resolve to parent via cass context before cass resume
Using --limit 0 for "no limit"
Earlier cass panics
Use a real limit (--limit 50); --limit 1 minimum for aggregations
Trusting 0 hits with --workspace /X
Workspace strings are case- and trailing-slash-sensitive
Re-run with --aggregate workspace --limit 1 to discover the canonical key
Skipping --fields minimal on wide scans
~3KB per hit × 100 hits = 300KB context burn
--fields minimal for wide passes; upgrade to summary/full for keepers
Reading session files with cat
Loads the full conversation into context
cass view PATH -n LINE -C 5 or cass expand PATH --line LINE --context 3
Re-indexing on every search
Index is shared across processes
Refresh only when status says stale
Treating a timed-out search during rebuild as 0 hits
Concurrent reads may exceed normal latency while an authoritative rebuild holds shared resources
Preserve exit 124 as "not observed" and retry once the active rebuild settles
Falling back to manual find/grep when cass misbehaves
Recovery is autonomous; skipping cass loses the corpus
Walk the recovery tree in RECOVERY.md. One real exception: terms inside tool stdout/stderr are skipped at index time — there rg -n "TERM" /path.jsonl is correct
Pre-authorized (rebuilds derived index data only, never destroys source sessions): cass doctor --fix --json, cass index --full --force-rebuild --json, cass sources doctor/sync, cass models install/verify.
Do NOT without explicit permission: delete core.NNNNN coredumps, delete .beads/, git reset --hard, or hand-edit ~/.config/cass/sources.toml — the CLI commands above already do everything safely. Never run bare cass (blocking TUI) inside an agent loop.
When the right reference isn't obvious from titles, grep -ni "SYMPTOM" references/*.md — cheaper than loading whole files into context.
Scripts
Scripts live under scripts/. They execute, never load — zero context tokens. Consistent with the Safety Boundaries above, recover.sh and quick_analysis.sh may rebuild derived index state autonomously (pre-authorized: doctor --fix, index --full) and multi_machine_search.sh reads remote sources over ssh; none destroy source sessions, and nothing destructive (coredump/.beads deletion, git reset --hard, source edits) runs without explicit confirmation.
Script
Usage
./scripts/quick_analysis.sh /path
One-command project overview (status → aggregate agent/date → top prompts)
./scripts/prompt_miner.py --workspace /path
Find repeated prompts (ritual detection)
./scripts/validate.sh
Validate cass install + skill structure
./scripts/recover.sh
Autonomous recovery decision tree (READY → STALE_BUT_USABLE → BROKEN); wraps every cass index in timeout
./scripts/multi_machine_search.sh "QUERY" [host…]
Parallel fan-out across the fleet; merges + dedups hits
Validation
# Quick health check
cass status --json | jq '.index.fresh'# Should return: true
If false, run: cass index --json
Output Specification
Path: stdout for search, status, capability, and introspection results; this skill creates no artifact directory by default.
Filename: none unless the caller explicitly requests an export path such as /tmp/cass-export.json with -o.
Format: use the installed command's JSON schema from cass introspect --json; wide searches should keep --fields minimal and downstream narrowing must preserve source_path and line location.
Exit code: validate with cass status --json | jq -e . and parse every selected result with jq; a nonzero command, malformed JSON, or unresolved source path blocks the handoff.
Downstream handoff: consumed by research, planning, recovery, or postmortem work with the exact query, canonical workspace, selected source paths/lines, and index freshness noted.
Quality Checklist
Results come from a canonical workspace key and include enough source location to reopen the session context.
A zero-hit result was retried against discovered workspace keys before concluding that no prior art exists.
Index recovery stayed bounded and preserved source sessions; no stale state was misreported as broken.
