JsonlView

JsonlView is a read-only VS Code event data studio. Its first format adapter is a bounded JSONL/NDJSON engine; semantic profiles turn records into useful views for traditional software logs, telemetry and traces, and Agent trajectories.

It opens .jsonl and .ndjson as an optional read-only custom editor with:

• bounded byte indexing and lazy record hydration;
• virtualized table, event timeline, schema, problems, and record detail views;
• typed filtering, full-text search, forward/backward cursor pagination, and cancellation;
• direct physical-page input for large files, with BigInt-safe ordinal anchors;
• automatic semantic profile detection without changing the source record;
• collapsible JSON Tree, highlighted Pretty JSON, exact Source, on-demand full event text, derived semantics, and Agent content containers with Auto/Text/Markdown/JSON modes, and byte-boundary detail views;
• source generation tracking, stale-message rejection, manual rebuild, and follow mode;
• tail follow that keeps the prior page visible until a verified append generation can replace it atomically;
• on-demand category and time-bucket Insights with cancellation and hard scan limits.

The source file remains authoritative. JsonlView never edits it, executes its contents, or sends records over the network. File-backed resources are required in this release.

Visual examples

These screenshots use synthetic fixtures maintained in the companion JsonlView-harness checkout. They show the workbench across agent events, telemetry, and aggregate analysis.

Codex Rollout - Table

Generic Agent Events - Table and Tree

OpenTelemetry - Schema and Tree

Claude Code Session - Insights

Product boundary

JSONL is the current physical format, while profiles provide the semantic view. A profile may interpret one JSONL record as an application log, OTel log/span, request event, Codex/Claude event, or software-engineering Agent step. Profiles do not own file I/O and cannot rewrite, drop, or split the authoritative record.

Formats that are not line-delimited JSON need separate adapters. Plain or multiline text logs, a regular JSON array/object, Chrome/Jaeger trace JSON, OTLP protobuf, and compressed .jsonl.gz are not claimed as supported in v0.1. See docs/format-and-profile-boundaries.md and the measured docs/acceleration-roadmap.md.

Built-in semantic profiles currently cover Codex rollout, Claude Code, vendor-neutral Agent events, software-engineering trajectories, OpenTelemetry, traditional structured application logs, and Generic JSONL fallback. Domain events keep their identity in the timeline (log, span, task, action, observation, patch, test, and result); severity remains a separate dimension.

Follow mode is opt-in. It opens the current tail page, keeps the visible rows stable while an appended snapshot is rebuilt, and swaps to the new tail page in one state update. Replacements, truncation, and invalid prefix changes still advance the source generation rather than mutating old record references.

Insights is also opt-in: no aggregate scan runs during open. Entering the tab or pressing its refresh button examines at most 100,000 physical records, 64 MiB of logical record content, or 5 seconds per request, whichever comes first. Bounded category/time accumulators receive only matching rows; requests are cancellable and report examined versus matched records, truncation reason, and group overflow.

Agent message, reasoning, and tool-result text keeps its original value for copying and can be viewed as plain text, conservative Markdown, or embedded object/array JSON. Known Codex item_completed -> AgentMessage -> Text content selects Markdown by default; tool output and ordinary log text retain Auto mode. Auto recognizes complete JSON, parseable JSON embedded in prose, headings/lists/quotes/tables, and fenced code; malformed or ambiguous content stays text. Markdown HTML is never executed. Fenced Python, Shell, JS/TS, Rust, and SQL snippets receive bounded lexical highlighting, plus per-block copy and Wrap controls; unknown languages stay exact source. Rich parsing is bounded to 64 KiB and displays its boundary; the Text mode remains available for the full hydrated field after an explicit Show full action.

Experimental native scanner

jsonlView.native.newlineScanner defaults to off. auto probes the packaged win32-x64 Rust/napi-rs scanner and keeps it only when a bounded calibration is at least 10% faster than the adaptive Node scanner; missing, incompatible, or invalid native output fuses back to Node. on prefers native but retains the same validation and fallback. A native process crash cannot be caught by a JavaScript fallback, which is why native is not the default in v0.1.

pnpm benchmark:engine --file <stable.jsonl> --scanner off --no-query
pnpm benchmark:engine --file <stable.jsonl> --scanner auto --no-query
pnpm benchmark:engine --file <stable.jsonl> --scanner on --no-query

Development

Generated benchmark and CDP acceptance artifacts are kept outside the source workspace by default, under the operating system temporary directory. Set JSONLVIEW_ARTIFACT_DIR to retain them in a specific external directory.

pnpm install --frozen-lockfile
pnpm typecheck
pnpm test
pnpm build
pnpm native:verify
pnpm package:vsix

The architecture and format boundary notes live in docs/format-and-profile-boundaries.md and docs/acceleration-roadmap.md.

Package distribution

The VSIX is the normal installation path for VS Code. The same runtime can be prepared as @sumi-labs/jsonl-view for the Sumi Labs package namespace with pnpm package:npm. The generated npm package is source-available under LICENSE.txt and is not an open-source license grant.