Session Search

Session Search

Use this when Joel asks for session/conversation history or when rag_search_sessions is missing config, stale, or too narrow.

Canonical terms:

• Run: captured agent invocation in Central/Typesense.
• Conversation: sibling Runs sharing interactive context.
• Session Search Bridge: CLI convenience that searches Central's derived Run index plus raw Pi session JSONL on the local Machine or a remote Machine over SSH.

Fast path

Search both Central Typesense and raw dark-wizard Pi sessions from Panda:

joelclaw sessions search "<query>" \
  --source both \
  --machine dark-wizard \
  --ssh-target joel@dark-wizard \
  --limit 8

Read the JSON. Prefer hits with:

• source: "typesense" when the derived index is current
• source: "local" when running on the Machine that has the raw Pi files
• source: "ssh" when Typesense is stale or you need raw remote session files from another Machine

Extract task context

When search finds a candidate but snippets are too thin, extract bounded task context without a background reader agent:

joelclaw sessions search "<query>" --source both --machine dark-wizard --limit 5 --extract
# singular alias also works:
joelclaw session search "<query>" --source both --machine dark-wizard --limit 5 --extract

For a known raw transcript path or session id:

joelclaw sessions extract <session-id-or-path> --query "<topic>" --format markdown

Use sessions inspect for deterministic evidence around exact text:

joelclaw sessions inspect <session-id-or-path> --around "<regex>" --before 20 --after 80

Use sessions chunks when you want matching chunks/snippets and neighboring raw context without full extraction:

joelclaw sessions chunks "<query>" --source local --machine dark-wizard --limit 20 --context-before 2 --context-after 4

Use sessions signals when the job is not recovery but analysis: finding high-signal user turns, friction, preferences, approvals, or decisions. This is ADR-0247-backed and deterministic in v1:

joelclaw sessions signals --kind friction --source local --machine dark-wizard --since 14d --limit 20
joelclaw sessions signals --kind any --source local --machine dark-wizard --since 14d --sample 20 --review-out ~/.joelclaw/session-signals/review.jsonl --evaluate
joelclaw sessions signals --kind mode-mismatch --source local --machine dark-wizard --since 14d --evaluate
joelclaw sessions signals --kind any --source local --machine dark-wizard --since 8w --format ndjson | jq -c 'select(.type=="hit" and .improvement.reviewPriority=="high")'
joelclaw sessions friction --source local --machine dark-wizard --since 14d

friction is an alias for signals --kind friction. The command filters to user turns first, classifies turnKind, and includes assistant/tool turns only as bounded evidence context. V1 includes operator_intent, review_feedback, and approval; it excludes task payloads, source material, and handoffs by default because those caused noisy false positives. Joel's fuck, fucking, and fuckin are strong emphasis signals, not automatic anger; classify by nearby critique/praise/correction language. ShitRat is agent identity, not friction. Use --kind mode-mismatch when hunting wrong execution shape: inline vs background, visual vs response, durable Inngest vs ad-hoc script, or feedback-blocking vs async.

Extraction returns session ID, path, dates, cwd, user prompts, decisions, commands run, files touched, outputs/receipts, verification, blockers, next actions, and transcript line evidence. It redacts likely secrets and does not dump whole transcripts.

Shape rules:

• search --extract attaches extraction directly to each emitted hit when available. .result.extractions is a compatibility convenience list. Piped | jq usage is supported for large envelopes.
• extract --format markdown still emits the canonical JSON envelope; rendered markdown is at .result.markdown.
• chunks exposes top-level .result.chunks / .result.hits, with source mirrors under .result.local.chunks and .result.typesense.chunks.
• Source metadata distinguishes rawReturned from emittedHits / emittedChunks.
• signals returns kind, clusters, hits, signals[], stable hitId, exact path, and transcript line numbers. Treat it as a signal radar, not memory truth; promote only derived reusable guidance.
• Use --sample N --review-out path.jsonl to create a small golden set for human/agent labeling before tuning the classifier. Review rows include verdict, correctedKind, and note placeholders.
• Each signal includes turnKind plus a small reducer-shaped improvement route: surface (system-prompt, skill, cli, harness, docs, memory, adr, none), target, confidence, reviewPriority, suggested next step, and reason. Use --evaluate to summarize routing coverage by surface, kind, turnKind, confidence, and review priority.
• Use --format ndjson for balls-deep scans over SSH or thin Machines. It streams one JSON object per line: meta, many hit rows, then summary, so consumers can pipe through jq without requiring a giant response envelope.

Source-specific searches

Typesense only:

joelclaw sessions search "<query>" --source typesense --machine dark-wizard --runtime pi --limit 8
joelclaw sessions search "<query>" --source typesense --machine dark-wizard --runtime codex --limit 8
joelclaw sessions search "<query>" --source typesense --machine all --runtime all --limit 8

Raw local Pi sessions only, for example on dark-wizard itself:

joelclaw sessions search "<query>" --source local --machine dark-wizard --limit 8

Raw remote Pi sessions only:

joelclaw sessions search "<query>" --source ssh --ssh-target joel@dark-wizard --limit 8

All Machines in Typesense:

joelclaw sessions search "<query>" --source typesense --machine all --runtime all --limit 8

Runtime filter:

• default: --runtime pi
• Codex: --runtime codex
• Claude Code: --runtime claude-code
• everything: --runtime all

Codex raw fallback is not scanned directly yet; Codex capture writes transcript deltas through Central /api/runs, then memory/run.captured indexes runs_dev / run_chunks_dev.

When results disagree

If SSH finds fresh hits but Typesense does not, raw capture exists and the derived index is stale. Use the recovery runbook:

read ~/Code/joelhooks/joelclaw/docs/runbooks/typesense-session-indexing.md

Then verify:

KEY=$(secrets lease typesense_api_key)
curl -fsS -G http://localhost:8108/collections/runs_dev/documents/search \
  -H "X-TYPESENSE-API-KEY: $KEY" \
  --data-urlencode 'q=*' \
  --data-urlencode 'filter_by=agent_runtime:=pi && machine_id:=dark-wizard' \
  --data-urlencode 'per_page=3' \
  --data-urlencode 'sort_by=started_at:desc' \
  --data-urlencode 'include_fields=id,machine_id,started_at,tags' \
  | jq '[.hits[].document | {id,machine_id,started_at:(.started_at/1000|todateiso8601)}]'

Local and SSH requirements

The bridge expects the Machine being scanned to have:

• python3
• raw Pi sessions under ~/.pi/agent/sessions/**/*.jsonl

SSH mode also needs:

• ssh joel@dark-wizard access

Smoke test:

ssh joel@dark-wizard 'hostname && python3 --version && find ~/.pi/agent/sessions -type f -name "*.jsonl" | head'

On dark-wizard itself:

joelclaw sessions search "<query>" --source local --machine dark-wizard --limit 8

With --source both, the raw side uses local when the current hostname matches --machine; otherwise it uses ssh. That keeps dark-wizard from SSHing to itself. If the local Machine has no Typesense credential, both skips Typesense and returns raw local hits with typesenseSkipped; use --source local when you want that explicitly. If Typesense is unavailable in both mode, the CLI reports typesenseUnavailable and continues with raw results.

Rules

• Do not print secrets.
• Do not copy raw session files unless Joel asks; return pointers and snippets first.
• Use bounded search output. No full transcript dumps.
• If you change indexing, services, config, or this skill, write a slog entry and commit.