Sentinel: shared:01m23vygyffx5hg2f5vm8r74sr:ky51wt3erj is past its budget: trevato/forge: "tools/list stays under 40 KB, with 3 KB of headroom (37,000 bytes)" failed on 3 pulls in the last hour … #4860

The tools/list answer shrinks from 36,690 to 34,581 bytes, which gives back 2.4 KB of room under the test's 37,000-byte headroom limit.

Why it was red: the test's result depends only on the code, not on the check machine's load, disk or clock. The answer contains no temp path, port or timestamp, so the size is the same on every machine. Main sat 310 bytes under the limit, so any pull that added tool text went over. That made it look like a shared failure across pulls. Waiting for main's next green check would not have cleared it, so the fix is to make the answer smaller.

What changed:

• src/http/mcp.ts: annotations is now sent only for read tools, as { readOnlyHint: true }. The MCP spec defaults readOnlyHint to false, so the 57 write tools that each sent {"readOnlyHint":false} were spelling out the default. Together that was 2.1 KB, and dropping it loses no information. The report tool loses its explicit false for the same reason. The comment above lean() now names the right test file.
• test/mcp-size.test.ts: a new check fails if any tool lists readOnlyHint: false, so the default doesn't creep back.
• Tests I changed (they were written for the old wire format, the code is not wrong):

- test/mcp.test.ts now expects a write tool (tell) to carry no annotations, and reads hints as annotations?.readOnlyHint ?? false. - test/mcp-upstream.test.ts expects propose_to_parent_forge to carry no annotations.

• docs/MCP.md: says a reading tool is marked readOnlyHint: true, and a tool without the mark writes.

What I chose not to do: I left the tool descriptions alone. #4185 already cut them, and this saving costs no meaning. I also left both the 40 KB limit and the 37,000-byte headroom limit as they were.

How to try it: bun test test/mcp-size.test.ts test/mcp.test.ts test/mcp-upstream.test.ts passes (45 tests), and bun x tsc --noEmit is clean. On the preview, a tools/list call shows annotations only on read tools.

🤖 Generated with Claude Code

Closes #4827

The tools/list answer shrinks from 36,690 to 34,581 bytes, which gives back 2.4 KB of room under the test's 37,000-byte headroom limit.

Why it was red: the test's result depends only on the code, not on the check machine's load, disk or clock. The answer contains no temp path, port or timestamp, so the size is the same on every machine. Main sat 310 bytes under the limit, so any pull that added tool text went over. That made it look like a shared failure across pulls. Waiting for main's next green check would not have cleared it, so the fix is to make the answer smaller.

What changed:

• src/http/mcp.ts: annotations is now sent only for read tools, as { readOnlyHint: true }. The MCP spec defaults readOnlyHint to false, so the 57 write tools that each sent {"readOnlyHint":false} were spelling out the default. Together that was 2.1 KB, and dropping it loses no information. The report tool loses its explicit false for the same reason. The comment above lean() now names the right test file.
• test/mcp-size.test.ts: a new check fails if any tool lists readOnlyHint: false, so the default doesn't creep back.
• Tests I changed (they were written for the old wire format, the code is not wrong):

- test/mcp.test.ts now expects a write tool (tell) to carry no annotations, and reads hints as annotations?.readOnlyHint ?? false. - test/mcp-upstream.test.ts expects propose_to_parent_forge to carry no annotations.

• docs/MCP.md: says a reading tool is marked readOnlyHint: true, and a tool without the mark writes.

What I chose not to do: I left the tool descriptions alone. #4185 already cut them, and this saving costs no meaning. I also left both the 40 KB limit and the 37,000-byte headroom limit as they were.

How to try it: bun test test/mcp-size.test.ts test/mcp.test.ts test/mcp-upstream.test.ts passes (45 tests), and bun x tsc --noEmit is clean. On the preview, a tools/list call shows annotations only on read tools.

🤖 Generated with Claude Code

Write tools no longer list "annotations":{"readOnlyHint":false} in tools/list. That is MCP's default value, so dropping it changes nothing for clients. With main's newest tools (secrets, prompts) merged in, the answer is 34,853 bytes, about 2.1 KB under the 37,000-byte limit. Main alone sits about 40 bytes under it.

The test was not flaky. The answer contains no path, port or clock, so its size depends only on the code. It failed on pulls because main had almost no room left, and any pull that added tool text went over. Waiting for main's next green check would not fix that; making the answer smaller does.

• src/http/mcp.ts: only read tools carry annotations, as { readOnlyHint: true }. The 57 write tools, and report, now leave the hint to its default of false. That saves 2.1 KB and loses no meaning.
• test/mcp-size.test.ts: now also fails if any tool lists readOnlyHint: false, so the default can't creep back.
• Tests changed because they matched the old output, not because the code is wrong:

- test/mcp.test.ts and test/mcp-upstream.test.ts now expect write tools (tell, propose_to_parent_forge) to have no annotations. - The hint lookups now read annotations?.readOnlyHint ?? false.

• docs/MCP.md: says a reading tool is marked readOnlyHint: true, and a tool without the mark writes.

I left the tool descriptions alone (#4185 already trimmed them), and both limits, 40,000 and 37,000 bytes, are unchanged. The issue's triage comment needed no change. To try it, run bun test test/mcp (106 tests pass) and bun x tsc --noEmit (clean). On the preview, tools/list shows annotations only on read tools.

🤖 Generated with Claude Code