Tool Defs total is a never-pruned union across sessions, then multiplied by toolsRatio #797

A session with 34 tool definitions showed 62K for Tool Defs. Those 34 tools measure 13,062 with MC's own estimator. The rest of the number is tools that were not in the request.

Two effects stack.

1. The measurement set is a lifetime ledger. recordToolDefinition() (packages/plugin/src/features/magic-context/tool-definition-tokens.ts:170) stores each tool under ${providerID}/${modelID}/${agentName}, keyed by toolID. There is no session in the key, and the module docstring (:4-8) says so on purpose. getMeasuredToolDefinitionTokens() (:255) sums the whole map, and nothing removes a toolID when that tool stops being registered. loadToolDefinitionMeasurements() (:128) reloads every row from tool_definition_measurements at startup, so the stale set also survives restarts. Rows come from the tool.definition hook in index.ts:671-689.

Both the per-flight measurement and the hook attribution look correct. The problem is that the total is a union over every session, directory and OpenCode version that ever ran on that model and agent.

2. A calibration ratio multiplies the union. estimateFinalWireInputTokens() (packages/plugin/src/hooks/magic-context/final-wire-token-estimate.ts:159-162) returns Math.round(measuredToolDefinitions * calibration.toolsRatio), and the sidebar uses that value (packages/plugin/src/plugin/rpc-handlers.ts:464-465). For the 62K session toolsRatio is 1.54, so 40,271 measured becomes 62,000 displayed.

Numbers from my store, agent orchestrator:

The 74 extra rows in model A include lsp, task and todowrite (built-ins OpenCode 2 no longer ships, last measured 2026-09-23), bash (AFT's earlier name for shell), openchamber, and tools from unrelated repos: rdepot_*, k8s_math_api, regress_*, pr_*, k6_*, splunk_*, positron_*. Rows keep being added today (a k6_* row at 23:59 last night, from a work repo), and bucket size tracks age: the oldest pair holds 171 rows, a pair created a week ago holds 74.

To reproduce:

select provider_id, model_id, agent_name, count(*), sum(token_count)
from tool_definition_measurements
group by 1, 2, 3;

Then compare one bucket against the tool ids the sidebar session actually sends (OpenCode 2 with Code Mode on: execute plus the tools registered with codemode: false).

Suggested fix. Measure the set that is serialized for the flight. OpenCode 2 hands plugins the exact post-split tools record on the session.context hook, so the payload set is available without a second mechanism. If keeping tool.definition, prune ids not seen in the last N flights of that key and total the pruned set, so removing a tool also shrinks the number.

Environment: plugin 0.49.23, OpenCode 2.1.4, darwin arm64, doctor PASS 18 / WARN 1 / FAIL 0.

One unrelated thing seen while running doctor there: it reports Local embedding runtime unverified: no installed plugin tree found to inspect on OpenCode 2 because getOpenCodePluginCacheRoots() looks under ~/.cache/opencode/packages/..., which is the OpenCode 1 layout. OpenCode 2 installs to ~/.cache/opencode/npm/<name>@<spec>/<timestamp>/node_modules/..., so the probe never finds a tree.