Prototype the eight v2 tool schemas and measure the tool list #311

Gate for #310. Write the eight tools from #309 as schemas with their final descriptions and stub bodies, advertise only them in a test host, and measure with the method of ToolListStaysWithinItsContextBudget.

Baseline today: 27 tools, 19,014 characters, about 4,750 tokens. The four write tools (memory_store 2,663, page_store 1,610, task_patch 1,274, task_create 1,068) are 6,615 of them.

Two measurement series

1. v2: the eight tools with descriptions as short as they will ship, including the one-line field list per kind for item_store and the action table for item_action. Per tool: characters and o200k tokens.
2. Control: today's 27 tools with only the consistency fixes and no consolidation. Same numbers per tool.
3. Growth: for both series, the cost of one more kind.

Client context

Round-trip size for on-demand loading and keyword findability, recorded without weight in the decision.

Real session

On the stubs, one agent session forms six calls; count refused calls against today's tools.

Result (2026-09-18)

All numbers are on #309 under "Measurements"; stubs and measuring test in tests/Conderis.Tests/ToolSurface (da0842d), session harness in tests/ToolSurfaceSession (ae81f71).

• List: v2 6,386 characters / 1,431 tokens against 19,014 / 4,043 today and 12,614 / 2,872 for the consistent 28 tools.
• Growth by one kind: v2 +97 characters, consistent series +1,038.
• Retries: final v2 one refused call in three headless runs, today two in three; every refusal was corrected in the next call.
• Findings that changed the design: agents always turned "open tasks" into status: open and missed blocked work, so the task list has no status filter, only includeClosed; agents left out the version on approve, and since proposals are immutable, proposal actions need none. Both are in #309 (Decision 8), #313 and #314.

Decision

Continue with v2 as in #309. Both gating criteria are met. The consistency fixes alone would leave the list at twice the size and add three to four tools per future kind.