What are the paste limits people hit?

A GitHub issue or PR comment caps at 65,536 characters. Slack truncates a message past 4,000 characters into a snippet file. Jira and Linear descriptions have their own ceilings, and every one of them mangles wide tables and long code blocks well before the hard limit.

Why is truncating the agent's output a bad idea?

Because the parts that get cut are the evidence — the per-file findings, the diff context, the trace. What survives is the summary, which is the part a reader is least able to check.

What should an agent do with a long report instead?

Publish it and return a link. The agent POSTs the HTML to a report endpoint, gets a URL back, and pastes that one line into the PR or the channel where the wall of text was going to go.

Can Claude Code or Cursor do this without a custom script?

Yes, over MCP. The same scoped token works for the REST API and the MCP server, so an agent in Claude Code, Cursor, VS Code or Zed can create and update reports as a tool call.

Does publishing help the agent's own context budget?

Considerably. Handing back a URL instead of carrying tens of thousands of tokens of report text through the rest of the session leaves that budget for the work — and the report is still readable by anyone who wants it.

The agent produced 40,000 words and there is nowhere to put them

An audit across the repository. A migration plan with per-file findings. An evaluation run over two hundred cases. The agent did the work, and now the output has to reach a human.

Every destination refuses it.

The ceilings

Destination	Limit
GitHub issue / PR comment	65,536 characters — the API rejects the write
Slack message	~4,000 characters, then it becomes a snippet file
Jira / Linear description	Field-dependent, and formatting degrades first
Email	Gmail clips past ~102 KB behind "show trimmed content"
A terminal scrollback	Gone by tomorrow

The soft failures matter more than the hard ones. Long before a comment is rejected, the wide table has collapsed, the code block has lost its language, and the reviewer has scrolled past all of it.

Why splitting it up is worse

The reflexes are: truncate, or paste in parts, or attach a file.

Truncating removes the evidence and keeps the claim. What survives is the summary — the one part a reader cannot verify. That is exactly backwards for output produced by a model.

Splitting across comments destroys the ordering the moment anyone replies in the middle, and no one reads part four.

Attaching a markdown or HTML file hands the reader a download that Slack will not preview, GitHub will not render, and email clients strip. If it does get opened, it opens from file:// with whatever that breaks.

Give the agent somewhere to publish

The agent writes the report as HTML, POSTs it, and returns a URL:

curl -sS -X POST https://commareports.com/api/v1/reports \
  -H "Authorization: Bearer $COMMA_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"title": "Dependency audit — 2026-09-07", "html": "<h1>…</h1>"}'

One line goes in the PR. The report stays whole.

In Claude Code, Cursor, VS Code, Zed and anything else speaking Model Context Protocol, this is a tool call rather than a shell command — the same scoped comma_sk_… token authenticates both surfaces, so revoke, expiry and scope behave identically. Setup per client is under MCP setup; the agent-side conventions are in where should my agent post.

The two things this changes

Context budget. Carrying 40,000 words of finished report through the rest of a session is a large, permanent tax on the agent's remaining window. A URL is a dozen tokens and the content is still there for anyone who opens it. See context budgeting for AI reports.

Review. A published report is a place to disagree with the agent precisely. The reviewer selects finding #34 and writes "this one is a false positive — the call is behind a feature flag", anchored to that finding. The agent can read the comments back and revise (let an agent respond to comments), posting the next pass as a revision at the same URL rather than as a new wall of text.

That loop is the reason not to paste: a chat message is a broadcast, and a report with anchored comments is a conversation about specific claims.

Try it

Comma is free — unlimited reports, unlimited commenters, unlimited revision history.

Give your agent a place to publish →