Spiderling

See what your AI coding agents actually did.
A little spider crawls your code while Claude Code, Codex and Antigravity read, search and edit,
and every session becomes a trace you can replay, review and share.

Why

An agent can read forty files and make a dozen edits before you look up. Was it heading the right way? Why did it change that line? What did it do between the plan and the pull request? The diff shows what changed, not how it got there.

Spiderling reads the logs your agents already write and turns them into something you can watch live and review afterwards. Nothing to install in the agents, no proxy, no terminal wrapper, no network.

What you get

• Live view (Code Crawl). One lane per agent. The spider walks the exact lines an agent reads, weaves a web over the lines it edits, types its searches into a search box and walks the hits, and when the agent opens one of those hits it crawls from the hit into the file. An agent that reads twenty files in a burst is followed file by file, and the lane catches up instead of falling behind.
• Overview (Crawl Map). Files as nodes, a spider per agent crawling between them, and a full recap of every step.
• Replay. Every session is written to a trace file you can scrub like a video:
  • a timeline with one lane per agent (and idle gaps marked),
  • phases (explore → plan → implement → verify → ship) as a band you can click,
  • the agent's own words next to each step: the "why",
  • what led to each step: "led here by ⌕ query" and "context for this edit",
  • worth a look: a file re-read nine times, the same lines rewritten, a run of searches that found nothing, failed tool calls, long idle time,
  • a Flow view: searches → files read → files edited,
  • the function around an edit ("edited calculateTotal") for live steps.
• From code to trace. Select code, right-click → Show Agent Trace for This Code: pick the step that wrote those lines and the replay opens right there, with the agent's reason. The answer to "why is this line like this?".
• Share it. Export a trace as one self-contained HTML file anyone can open, or copy a short markdown summary for a pull request.

Supported agents

Only sessions that worked in the folder you have open, or in another git worktree of the same repository, are read; other projects' logs are never opened past their first few lines. Details and per-agent limits: docs/AGENT_SUPPORT.md.

Install

From a release (until it is on the Marketplace):

1. Download spiderling-<version>.vsix from the Releases page.
2. code --install-extension spiderling-<version>.vsix (or Extensions → … → Install from VSIX…), then reload the window.

From source:

git clone https://github.com/EliasHridoy/spiderling.git
cd spiderling
npm install
npm run package            # builds spiderling-<version>.vsix
code --install-extension spiderling-*.vsix

Needs VS Code 1.84 or newer. Function names on edits need Node 22+ in VS Code (current VS Code has it) and a CodeGraph index in the repo; without them everything else works and edits just show their lines.

Quick start

1. Open the folder you will run an agent in (File → Open Folder).
2. Run your agent as usual, in any terminal.
3. A bug icon with the number of agents and files appears in the status bar, and the Spiderling view appears in the Explorer.
4. Command Palette (Ctrl/Cmd+Shift+P):
   • Spiderling: Open Code Crawl: the live view; keep it beside your editor.
   • Spiderling: Open Trace Replay: the review page; it follows the trace live while the agent is still running.
   • Spiderling: Open Crawl Map: the overview and recap.

On start it also replays the last hour of sessions (spiderling.sessionLookbackMinutes), so a session you started a few minutes ago is already there.

Commands

Settings

Reload the window after changing the capture and trace settings.

Traces and privacy

• Everything stays on your machine. Spiderling only reads local log files and writes local trace files. It makes no network requests.
• A trace holds paths, line ranges, search queries, the paths a search found, the agent's one-line explanations, failed-call counts and token/cost metrics. It never stores file contents or tool output.
• Credentials are masked before anything is written: well-known token formats (GitHub, AWS, Slack, Anthropic/OpenAI, Google, npm, GitLab, Stripe, JWTs), private keys, user:password@host URLs, Authorization/Bearer values, --password-style flags and name=value assignments of secret-looking names. This is pattern-based: a secret with no telltale format or name can slip through, so skim a trace before sharing it. Paths under your home directory are written as ~.
• Trace files are one JSON file per session in .spiderling/traces/, a folder that ignores itself in git. Format: docs/TRACE_FORMAT.md.

Good to know

• A step appears when a tool call finishes, so a lane is quiet while an agent is only thinking or waiting for you to approve something. Agents' thinking is not in their logs (Claude Code stores it redacted), so explanations come from what the agent wrote between tool calls.
• Phases are a guide, not a log. They are inferred from agent names, what was edited and which commands ran, so unusual setups get less precise labels.
• The agents' log formats are undocumented and change between versions. Spiderling reads them defensively and skips what it does not understand; if a new version breaks something, please open an issue.
• Antigravity logs carry no tool-call ids, so exit codes and hits are attached only when a turn made a single call; its grep_search hit lists are not in the log.

Documentation

• Architecture: how the pieces fit together
• Trace format: the JSON file, field by field
• Agent support: what is read from each agent, and how to add one
• Roadmap: what is planned and what is not
• Contributing · Changelog · Security

License

MIT © Elias Hridoy