engine: no fixed cap on a tool's result; the cut and the routing follow the model's window, read from OpenRouter on every turn

[feder-cr]

No tool result has a fixed cap any more. The only limit left is the model's own context window, read from OpenRouter on every turn, and a model's requests now go only to the providers that have that window.

Tool results come back whole

These fixed caps are gone:

Cap	Was
General cap on every tool result	12000 characters
exec output	10000, at most 50000, kept as head and tail
glob/grep pages	12000
Each matched line	2000
read_file	128000

The model's own paging stays: read_file's offset/limit and the search tools' head_limit.

A request that outgrows the model's window is compacted first, as before: old results are cleared, then the history is summarized. Only if it still does not fit are the results of its latest round cut, by ContextGovernor.fit_latest_round:

• The cut is measured with the same estimate that decides whether the request fits.
• The results share the room in their order, each leaving room for the notes of the results after it.
• The note says how much came through and how to ask for less: [Cut here: the first N of M characters. The whole result does not fit in the context window of <model> (<W> tokens) beside the rest of the request; ask for less to see the rest: a narrower selector, a pattern, an offset and a limit.]
• A result cut again keeps the whole result's length in its note.

A result is stored cut only past what no compaction could make room for: the window less the system prompt and the tool definitions.

The prompt caps stay: MEMORY.md at 25000 characters and MCP server instructions at 2048.

The window: read on every turn, and the routing follows it

• Limits: they are read at the start of every turn from GET /models/{model}/endpoints, about 32 KB. Before, the whole GET /models list (768 KB) was read once per process, so a long-lived engine never saw a model's providers change. Of the providers serving the model (a negative status means down), the model takes the widest window and the longest answer at that window.
• Routing: OpenRouter routes by the answer a request asks for, not by its prompt. A request sized for the widest window could therefore land on a narrower provider, for example Reka at 262144 among glm-5.3-flash's providers at 1048576. The model's requests now carry provider.only, set to the slugs of the providers with the widest window, from the same reading.
  • Checked live: such a request is served.
  • When none of those providers can serve, OpenRouter answers 404 "No allowed providers are available for the selected model" and the turn fails with those words.
  • A router (openrouter/auto) lists no providers, so it gets no limits and no route.

Image

invisible-playwright-mcp 0.72.1 removes its last caps:

• browser_type quotes the value the field kept in full.
• An error detail in its built-in chat and routes comes back in full.

The tools' schemas are unchanged; the recaptured interface differs only in its version.

Verification

• Engine pytest: 2492/2492, plus 272/272 in the provider and turn suites after the routing change.
• Image vitest (golden, runtime): 28/28.
• Engine smoke: 257 passed. Browser smoke: 74 passed. Nothing failed or was skipped in either.
• e2e on a real VM against the real OpenRouter: 16/16. The golden image was rebuilt from the 0.72.1 lock, and the turns read their limits from the endpoints and were routed with provider.only.

Known effect

An exec session that nobody polls keeps its whole output in memory until the next poll or the idle cleanup.