Add a memories table (survives index rebuilds) for agent-written
conclusions with provenance (session, message range, project). The agent
writes markdown files via Write tool (user-approved), then registers
them via a --remember CodeAct script with remember(). Recall via
memories() in --query scripts, filtered by project/session/time.
Separates query (read-only, assertReadOnlySql) from remember (write)
execution contexts in runtime.mjs.
semantics (context types, ordering, project scopes) out of pitfalls.md
into a new retrieval-semantics.md. Pitfalls.md becomes a compact debug
checklist: missing columns, FTS errors, over-large output, empty results.
SKILL.md query routing updated to point at the three reference tiers.
Teach the agent to derive second-pass filters from first-pass evidence
instead of pulling large message windows. Add the pattern and a pitfall
warning against defaulting to LIMIT 25 transcript browsing.
Teach the agent to keep intermediate retrieval inside the query script
and return compact evidence in a single turn, instead of spending
multiple conversation rounds showing raw results. Add the pattern to
query-patterns.md and the rationale to pitfalls.md.
Split the monolithic skill prompt into three tiers:
- Core API (search/context/sql) stays in the first prompt
- Structured helpers listed as one-liners with filter signatures
- Detailed patterns and pitfalls extracted to references/
Add references/query-patterns.md (copyable CodeAct recipes) and
references/pitfalls.md (scope, FTS, ordering, compactness traps).
Clarify project scope semantics (slug vs path vs cwd) throughout.
Add ORDER BY timestamp DESC to failures() for newest-first default.
lightweight workflowTree, build debounce, minor fixes
- Index phase, label, model, state, duration, tokens per workflow agent
- Index duration, total_tokens, status, name per workflow run
- workflowTree returns parsed result + agent summaries instead of
dumping all messages
- 30s debounce on buildIndex to avoid repeated directory scans
- Fix broken BASH_EXIT_PAT (SQLite LIKE has no character classes)
- Fix SKILL.md step numbering, document FTS5 hyphen limitation
Index the is_error boolean from JSONL tool_result blocks into a new
column, replacing the old ERROR_PATS text-matching approach that
produced ~90% false positives. Bash exit code pattern kept as fallback.
Add sessions() as first-class entry point for session discovery with
project/time/branch filters. Extend summaries, workflows, failures,
subagents, fileHistory with the same opts vocabulary that search()
already had — no more pulling full datasets to filter client-side.
Prepares for adding Codex JSONL support — new indexers can be added
without touching the query layer.
- db.mjs: schema, openDb, truncation helpers, shared utilities
- indexer.mjs: file discovery, JSONL parsing, SQLite writes
- query.mjs: all query API functions (search, context, raw, etc.)
- runtime.mjs: thin CLI entry point + VM script executor
- Replace readFileSync+split with chunked line reader (readLines) to
avoid loading entire JSONL files into memory at once
- Wrap each file's indexing in its own transaction so one corrupt file
doesn't roll back the entire index
- Log warnings to stderr instead of silently swallowing errors in
indexSubagentMeta, indexWorkflows, and indexHistory
- Add try/catch around readdirSync calls in discoverJsonlFiles and
indexWorkflows to survive unreadable directories
- Push error-pattern filtering in failures() into SQL LIKE clauses
instead of scanning all tool_results in JS
- Don't truncate workflow result_json and script (low-volume, must
stay parseable)
- Replace blind string truncation on tool_calls.input_json with
structure-aware truncJson that shortens individual string values
while keeping the JSON skeleton valid
- Add raw(uuid, {offset, limit}) API for windowed access to the
original JSONL line, so truncated content can be recovered on demand
Previously the session metadata accumulator started from zero on every
reindex pass, so an incremental update would overwrite started_at and
message_count with values derived only from the new lines.
Now reads the existing session row first and merges new data on top.
tommy0103