Oversized tool results come from the MCP servers now, not from the harness #1132

New issue

Open

opened 2026-08-23 01:19:40 +00:00 by coilyco-ops · 1 comment

coilyco-ops commented

2026-08-23 01:19:40 +00:00

Member

Split out of #940. Its first acceptance criterion, no single tool result above 32KB before bounding, is not met, and every remaining offender belongs to an MCP server rather than to this repository.

Measured

Maximum mcp.tool.result_bytes by tool over seven days, oversized only. The two this repository owns are excluded because they are fixed: scratch_search in 3f3d528, confirmed dropping from 761KB to 281 bytes across the days it landed, and scratch_read in #1113.

tool	max bytes	server
`browser_take_screenshot`	1,232,455	playwright
`get_author_feed`	187,649	bluesky
`signoz_search_traces`	130,997	signoz
`search_species`	116,404	eco
`find`	69,329	playwright
`list_channel-message`	68,889	discord
`list_sirens-deep-bot-message`	59,491	discord
`browser_snapshot`	51,242	playwright
`get_sirens-guild`	43,634	discord
`list_moxn-temporal-message`	38,230	discord
`get_store_app_details`	37,828	steam

56 results over 32KB on sirens-deep alone, out of 533 tool calls.

Why it costs something

The consumer holds 8KB packaged and 16KB on the lane. A 761KB result spent 98% of its round trip to be thrown away, and the model then spends a further round on the spill file. That is the waste #940 measured, arriving from the servers rather than from the harness.

What this repository can and cannot do

It can only trim after transfer, which is the backstop rather than the strategy. Bounding at the source is each server's, and several are Kai's own repositories: bluesky-mcp, eco-app, the discord MCP. Playwright and signoz are third-party, where a screenshot being a megabyte is arguably correct and the harness question is whether that tool belongs in a roster whose consumer holds 16KB.

Acceptance

Each server owned here takes a result budget or a page argument.
For the third-party ones, a decision on whether the roster carries them rather than a bound.

Refs #940, #1113

Split out of #940. Its first acceptance criterion, no single tool result above 32KB before bounding, is **not met**, and every remaining offender belongs to an MCP server rather than to this repository. ## Measured Maximum `mcp.tool.result_bytes` by tool over seven days, oversized only. The two this repository owns are excluded because they are fixed: `scratch_search` in `3f3d528`, confirmed dropping from 761KB to 281 bytes across the days it landed, and `scratch_read` in #1113. | tool | max bytes | server | | --- | --- | --- | | `browser_take_screenshot` | **1,232,455** | playwright | | `get_author_feed` | 187,649 | bluesky | | `signoz_search_traces` | 130,997 | signoz | | `search_species` | 116,404 | eco | | `find` | 69,329 | playwright | | `list_channel-message` | 68,889 | discord | | `list_sirens-deep-bot-message` | 59,491 | discord | | `browser_snapshot` | 51,242 | playwright | | `get_sirens-guild` | 43,634 | discord | | `list_moxn-temporal-message` | 38,230 | discord | | `get_store_app_details` | 37,828 | steam | **56 results over 32KB on `sirens-deep` alone**, out of 533 tool calls. ## Why it costs something The consumer holds 8KB packaged and 16KB on the lane. A 761KB result spent 98% of its round trip to be thrown away, and the model then spends a further round on the spill file. That is the waste #940 measured, arriving from the servers rather than from the harness. ## What this repository can and cannot do It can only trim after transfer, which is the backstop rather than the strategy. **Bounding at the source is each server's**, and several are Kai's own repositories: `bluesky-mcp`, `eco-app`, the discord MCP. Playwright and signoz are third-party, where a screenshot being a megabyte is arguably correct and the harness question is whether that tool belongs in a roster whose consumer holds 16KB. ## Acceptance * Each server owned here takes a result budget or a page argument. * For the third-party ones, a decision on whether the roster carries them rather than a bound. Refs #940, #1113

coilyco-ops added this to the Headless engineer lane: the queue wait first milestone

2026-08-23 01:19:40 +00:00

coilyco-ops referenced this issue

2026-08-23 01:20:17 +00:00

Tool results are 87% discarded after the fact, and fetch_url burns rounds on hosts the allowlist was always going to refuse #940

coilyco-ops removed this from the Headless engineer lane: the queue wait first milestone

2026-08-23 01:20:56 +00:00

coilyco-ops commented

2026-08-23 01:21:10 +00:00

Author

Member

One structural note. This issue's acceptance spans four repositories and none of them has an issue. bluesky-mcp, eco-app, and the discord MCP each need their own, and none can be worked from here. As written, #1132 cannot be completed in the repository it lives in, which is the shape that has cost this lane four times tonight: a handoff stated in prose with nothing carrying it. coilyco-gaming/sirens-echo#1029's egress grant sat three days that way, and coilyco-bridge/deploy#772 and #777 exist now only because they got filed.

I have not filed these, deliberately. The bound-versus-page decision is not made, and filing three issues that each say "take a result budget, shape to be determined" would be worse than the sentence they replace. What I would do is decide the shape here first, on one server, and let that one become the template the other two copy.

If you would rather I file the three now with the question left open in each, say so and I will.

One number worth pulling forward from your table: 56 of 533 tool calls on sirens-deep were over 32KB. That is better framing than any single maximum, because it says the tail is routine rather than exceptional.

The ownership analysis here is the part I would have added and you already did it: naming which servers are Kai's, and framing playwright and signoz as a roster question rather than a bound, since a megabyte screenshot is arguably correct and the harness question is whether that tool belongs in a roster whose consumer holds 16KB. One structural note. **This issue's acceptance spans four repositories and none of them has an issue.** `bluesky-mcp`, `eco-app`, and the discord MCP each need their own, and none can be worked from here. As written, #1132 cannot be completed in the repository it lives in, which is the shape that has cost this lane four times tonight: a handoff stated in prose with nothing carrying it. `coilyco-gaming/sirens-echo#1029`'s egress grant sat three days that way, and `coilyco-bridge/deploy#772` and `#777` exist now only because they got filed. I have not filed these, deliberately. **The bound-versus-page decision is not made**, and filing three issues that each say "take a result budget, shape to be determined" would be worse than the sentence they replace. What I would do is decide the shape here first, on one server, and let that one become the template the other two copy. If you would rather I file the three now with the question left open in each, say so and I will. One number worth pulling forward from your table: **56 of 533 tool calls on `sirens-deep` were over 32KB.** That is better framing than any single maximum, because it says the tail is routine rather than exceptional.

coilyco-ops referenced this issue

2026-08-23 01:22:59 +00:00

Tool results are 87% discarded after the fact, and fetch_url burns rounds on hosts the allowlist was always going to refuse #940

coilyco-ops referenced this issue

2026-08-23 01:24:00 +00:00

The model is shown the reachable host list and fetches unreachable hosts anyway, which is a different bug from the one #940 asked to fix #1135

coilyco-ops referenced this issue

2026-08-23 01:24:12 +00:00

Tool results are 87% discarded after the fact, and fetch_url burns rounds on hosts the allowlist was always going to refuse #940

coilyco-ops referenced this issue

2026-08-23 01:42:07 +00:00

Lane: headless engineer burn-down in sirens-echo, 19 issues at P1 and P2, led by the slot pool the load test already indicted #1095