fix(adapters)!: Anthropic and Bedrock input_tokens now inclusive of cache tokens

[mizcausevic-dev]

Summary

Branch 1 of the review's fix plan. Breaking, targets 0.3.0.

Anthropic and AWS both document input_tokens/inputTokens as excluding cache tokens: total_input_tokens = cache_read_input_tokens + cache_creation_input_tokens + input_tokens (Anthropic prompt-caching docs; matching formula in AWS's Bedrock Converse prompt-caching docs). This package's own types.ts and README already claimed cache tokens are "already included in inputTokens" for every provider, true for OpenAI, false for Anthropic and Bedrock, whose input_tokens was passed straight through unmodified. The OpenTelemetry GenAI semantic conventions this package targets also specify the total input count MAY be published regardless of caching.

Every 0.2.x version undercounted inputTokens on any cache-heavy Anthropic or Bedrock request. Full note in CHANGELOG.

• adapters.ts: anthropic and bedrock extract() now add cache tokens into inputTokens. openai untouched (already correct). gemini untouched, candidatesTokenCount/thoughtsTokenCount inclusion isn't stated in Google's docs, gated on a live API check not performed here, reported Blocked per the implementation brief rather than inferred from docs.
• types.ts: corrected the cacheWriteTokens doc comment, Bedrock has cache-write too, not Anthropic-only.
• test/normalize.test.ts: replaced the input_tokens: 20, cache_read: 15, cache_creation: 5 fixture, 15+5=20 made the exclusive and inclusive readings numerically indistinguishable, which is exactly why 98%+ coverage never caught this. Added a golden-invariant test per provider (inputTokens >= cacheReadTokens + cacheWriteTokens, outputTokens >= reasoningTokens) and doubled fixtures/responses.jsonl with cache/reasoning-bearing records so normalizeMany exercises the previously-untested path.
• README.md, CHANGELOG.md: documented the vendor asymmetry and the breaking change plainly.
• package.json: version bumped to 0.3.0. Not tagged, not published, per the implementation brief's boundaries.

Test plan

• npm ci
• npm run lint
• npm run typecheck
• npm run coverage — 25/25 tests
• npm run build
• npm run demo — output confirms the fix mechanically (Anthropic cache-heavy record: inputTokens: 100050 = 50 + 100000 + 0; Bedrock: 50034 = 30 + 50000 + 4)
• npm audit --audit-level=high
• Gemini live API check — not performed, Blocked. Gemini adapter intentionally untouched.

🤖 Generated with Claude Code