test(core): tool-execute callable-namespace test expects pretty JSON, Code Mode results are intentionally compact #65

Finding

packages/core/test/tool-execute.test.ts — "execute supports callable namespace tools" — fails deterministically on main (94276dae05).

The test's final assertion expects pretty-printed JSON for the Code Mode result:

// packages/core/test/tool-execute.test.ts:155
expect(result.content).toEqual([{ type: "text", text: '[\n  "admin",\n  "created"\n]' }])

But Code Mode result formatting is deliberately compact:

// packages/core/src/codemode/tool.ts:325-328
// Compact like native tool results: indentation costs context on every call and models read compact JSON as well.
function formatValue(value: CodeMode.DataValue) {
  if (typeof value === "string") return value
  return JSON.stringify(value) ?? String(value)
}

That compact choice matches the native tool path (packages/core/src/tool/runtime.ts:270 stringify also uses plain JSON.stringify(value)) and the sibling test in the same file, "canonical execution distinguishes declared, model-only, and raw schema outputs", which expects compact '{"value":"encoded"}'.

The pretty-printed expectation at :155 is the sole outlier — both it and formatValue landed in the same commit (1fc489d878), contradicting each other from the start. The callable-namespace behaviour itself (metadata toolCalls for slack.admin + slack.admin.create) passes; only the text formatting assertion fails.

Reproduction (verified this run, unmodified main)

cd packages/core && bun test test/tool-execute.test.ts
# 5 pass / 1 fail
# (fail) execute supports callable namespace tools
#   expected text: '[\n  "admin",\n  "created"\n]'
#   received text: '["admin","created"]'
#   at tool-execute.test.ts:155:26

This failure was first noted as an unnamed pre-existing cluster in #52's CI notes ("tool-execute 'execute supports callable namespace tools' — verified failing identically on unmodified main locally, 5 pass/1 fail"); it had no issue of its own until now.

Why it matters

A permanently red test in @opencode/core trains everyone to ignore unit (linux) failures and masks real regressions in the Code Mode callable-namespace path — the exact feature this test exists to pin (a namespace that is itself callable and also has child tools).

Acceptance criteria

• bun test test/tool-execute.test.ts in packages/core passes 6/6 on main.
• The fix aligns the outlier expectation with the intentional compact formatting (test fix), OR — if maintainers instead want pretty-printed Code Mode results — changes formatValue deliberately, updates its comment, and justifies the extra context cost against the native-tool precedent. Do not leave the two sides contradicting each other.
• No production behaviour changes unless the second route is explicitly chosen; the toolCalls metadata assertion stays as-is.