fix(opencode): Support incremental step usage, reasoning tokens, and model pricing

[4ndreello]

Summary

This PR addresses the issue where OpenCode sessions reported Tokens: 0 and unpriced $0.00? costs in CodeDeck metrics, and fixes associated pricing and parser regressions across the runtime.

Root Causes

1. Parser dropped step_finish: src/drivers/opencode/parser.ts explicitly discarded step_finish events where OpenCode reports step tokens and costs.
2. Harness dummy cost: OpenCode reports cost: 0 for models missing from its local models.json (such as alibaba-token-plan), which suppressed CodeDeck's fallback pricing.
3. Overwrite vs Accumulation: OpenCode emits tokens per-step (incremental deltas), whereas the daemon previously assumed cumulative session totals and overwrote tokens on each update.
4. Model ID prefix mismatch: Model identifiers with provider prefixes (e.g., alibaba-token-plan/qwen3.8-max or opencode/claude-sonnet-4-6) failed static lookup in MODEL_PRICES.
5. Claude cached token drop: src/drivers/claude/parser.ts used ?? between read and creation tokens, dropping cache creation tokens when read tokens were present.

Key Changes

• Core Events: Added optional incremental?: boolean to UsageUpdatedEvent.
• OpenCode Parser:
  • Captures step_finish events.
  • Folds tokens.reasoning into outputTokens (standard provider billing for o-series/Qwen/DeepSeek).
  • Maps tokens.cache.read to cachedTokens.
  • Converts cost: 0 to undefined to allow static table fallback while preserving positive reported costs.
  • Emits incremental: true.
• Daemon (updateSessionFromEvent):
  • For incremental: true, atomically adds inputTokens, outputTokens, cachedTokens, and cost to existing session usage in SQLite.
  • Non-incremental events continue to replace usage (preserving compatibility with Codex, Antigravity, and Claude).
  • Updates Session.model if reported in the usage event.
  • Deduplication via sourceKey in session-runtime.ts prevents double-counting during daemon restarts and process reattach.
• Pricing & Models:
  • Added resolveModelPrice() to normalize model IDs by stripping provider prefixes, display labels, and providing case-insensitive resolution.
  • Added Alibaba Qwen models (qwen3.8-max, qwen3.8-flash, qwen-plus, qwen-turbo) to MODEL_PRICES.
  • Added explicit cached rates for Claude models (0.1x of input rate).
• Claude Parser: Sums cache_read_input_tokens + cache_creation_input_tokens.
• Backfill Script: Added scripts/backfill-opencode-usage.ts to retroactively recover tokens and costs from historical .ndjson session logs.

Testing & Verification

• 83 test files passed, 1,130 tests total (0 failures, 0 warnings).
• Added test coverage for:
  • step_finish parsing, reasoning folding, incremental flag, and root/part fallbacks in tests/opencode-parser.test.ts.
  • Model normalization, provider prefixes, display labels, Qwen pricing, and negative tokens in tests/pricing.test.ts.
  • Incremental usage accumulation across steps and pricing in tests/usage-daemon.test.ts.
  • Cache token summation in tests/events.test.ts.
• TypeScript build (npm run build) completed with 0 errors.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud