fix: compact token formatter promotes to m before rounding

[Sma1lboy]

Direction

Bugs / correctness — found by daily code review, not from an open issue.

Problem

The web chat transcript header shows a live context estimate via formatTokens (packages/kobe-web/src/lib/history.ts). The function tested its unit threshold before rounding:

if (value >= 1_000_000) return `${(value / 1_000_000).toFixed(1)}m`
if (value >= 1_000)     return `${(value / 1_000).toFixed(1)}k`

Any value in [999_950, 999_999] clears the 1_000_000 check (so it stays in the "k" branch) but then (value / 1_000).toFixed(1) rounds 999.95…999.999 up to 1000.0, printing the impossible "1000.0k". This is a real, user-visible readout — long-context sessions routinely sit just under 1,000,000 tokens, and that's exactly the range that renders ctx 1000.0k in the transcript.

The repo already had the correct precedent: preview-core.ts's formatBytes deliberately promotes at 1023.5, "not 1024… once v rounds up it would render as the next unit."

Fix

Promote to m at 999_950 — the smallest value that rounds up to 1000.0k — so the boundary the half-up rounding actually crosses is the boundary the branch switches at. One-line threshold change, mirroring the established formatBytes rule.

The existing test pinned the buggy behavior (expect(formatTokens(999_999)).toBe("1000.0k"), with a comment rationalizing it as "just under 1m → still k"). Corrected it and added boundary coverage: 999_949 → 999.9k, 999_950 → 1.0m, 999_999 → 1.0m.

Verification

• bun x vitest run test/history-usage.test.ts (kobe-web) — 6 passing.
• bun run lint and bun run typecheck from repo root — both green.
• The 3 unrelated server-session / web-state-routes test failures are pre-existing on clean main (env-dependent), confirmed by re-running with this change stashed.

Patch changeset included.

Follow-ups (deliberately not in this slice)

fmtBytes in packages/kobe/src/cli/doctor-cmd.ts:74 has the identical root cause — fmtBytes(1_048_550) prints "1024.0 KB" instead of "1.0 MB". It lives in a different surface (kobe doctor diagnostics) with no existing test harness, so I left it out to keep this slice tight and surface it here rather than silently widen scope. Happy to fix it in a follow-up.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
kobe-landing	Ready	Preview	Aug 10, 2026 4:33pm

[claude]

Claude finished @Sma1lboy's task in 3m 34s —— View job

---

Review: PR #394 — fix: compact token formatter promotes to m before rounding

Verdict: Approve — clean, correctly-derived, well-tested fix. No blocking issues.

• Read CLAUDE.md (no packages/kobe/CLAUDE.md exists)
• Gather PR diff and changed files
• Run specialized review agents (code-reviewer, silent-failure-hunter, pr-test-analyzer, type-design-analyzer, comment-analyzer, code-simplifier)
• Check kobe house rules
• Synthesize findings

House rules

• No AI/Anthropic/Claude attribution in commit or PR body. ✅
• Changeset present (.changeset/token-count-unit-boundary.md), bump is patch, single-line body (no soft-wrap). ✅
• File size cap: history.ts 104 lines, test file 78 lines — well under 500. ✅
• Regression test on bug fix: the old test pinned the bug (expect(formatTokens(999_999)).toBe("1000.0k")) and has been corrected + a new boundary test (999_949→999.9k, 999_950→1.0m, 999_999→1.0m) was added. pr-test-analyzer hand-traced that both assertions genuinely fail pre-fix and pass post-fix — this satisfies the hard rule. ✅

Correctness

Two independent agents hand-verified the IEEE-754 arithmetic (bash/node execution was sandboxed in this environment, so this was done by hand-tracing doubles): 999_950/1000 rounds up to "1000.0" under the old code, 999_949/1000 correctly stays "999.9" — so 999_950 really is the exact smallest boundary, not an off-by-one. The cross-reference to formatBytes's 1023.5 promotion rule in packages/kobe/src/tui/ops/preview-core.ts:58 is verbatim accurate. Comments are accurate and proportionate; nothing to simplify.

Should-fix

None.

Nit

• packages/kobe-web/src/lib/history.ts:96 — the promotion boundary (999_950) is a bare magic number; a named constant (M_PROMOTE_THRESHOLD or similar) would let the invariant travel with the value instead of living only in the comment. Optional, not blocking.

Note (already handled correctly by the PR, not a gap)

code-reviewer flagged that fmtBytes in packages/kobe/src/cli/doctor-cmd.ts:74 has the identical root cause (bytes in 1_048_525–1_048_575 would print "1024.0 KB"). The PR body already explicitly surfaces this as a deliberate, out-of-scope follow-up rather than silently leaving it — exactly what CLAUDE.md's scope guidance asks for. Worth a quick follow-up PR, but not a gap in this one.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
see 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.