feat(telemetry): multi-chunk response continuation cost stitching in route_attempt #8

Context & Background

In OpenCode upstream commit a3f7f42ac9 (feat(core): continue responses after output token limits (#53876)) and commit 044d8b7d2e (feat(ai): continue split Gemini responses (#55056)), OpenCode added automatic multi-turn continuations: when an LLM truncates output because it reaches its maximum generation limit (finish_reason: length or max_tokens), the core runner automatically requests a continuation to complete the generation seamlessly.

In our fork (developing-today/opencode), every LLM call produces a discrete record in route_attempt with tokens_input, tokens_output, tokens_reasoning, and cost_usd. When a response requires multiple continuation requests, logging each continuation as an isolated attempt creates fragmented records that disconnect the full cost of the assistant turn.

Feature Objective

Enhance route_attempt and our usage metrics store to stitch continuation requests together into a unified turn attempt:

1. Link continuation attempts to the primary root request via a continuation_of or parent_attempt_id reference.
2. Provide aggregated cost and token rollups for multi-segment responses so that total turn cost and duration are accurately reflected.

Technical Specifications

1. Schema Enhancement

• Add a new migration in packages/core/src/database/migration/ (e.g. 20261017000000_route_attempt_continuation.ts):

  ALTER TABLE `route_attempt` ADD `continuation_of` integer;
  ALTER TABLE `route_attempt` ADD `continuation_index` integer DEFAULT 0;
  CREATE INDEX `route_attempt_continuation_idx` ON `route_attempt` (`continuation_of`);

2. Continuation Tracking in LLM Runner

• In packages/core/src/session/runner/llm.ts:
  • When a completion finishes with finish_reason: length and initiates a continuation:
  • Pass the parent attempt ID to the subsequent continuation request.
  • Set continuation_index incrementally (0 for initial, 1 for first continuation, etc.).
  • Calculate cost_usd for the incremental tokens of each chunk.

3. Metrics Rollup Queries

• In packages/core/src/model-usage.ts and packages/core/src/model-route-log/sql.ts:
  • When querying turn attempts, provide a helper getTurnAttemptMetrics(rootAttemptId) that sums:
    • Total input tokens (excluding duplicate cached prefixes where applicable)
    • Total output tokens across all continuation chunks
    • Total cost_usd
    • Total response time (from start of chunk 0 to end of last chunk)

Key Files & References

• packages/core/src/session/runner/llm.ts: Output token limit detection, response continuation loop
• packages/core/src/database/schema.gen.ts: route_attempt table
• packages/core/src/model-usage.ts: Attempt metrics aggregation and reporting
• packages/core/src/tool/plugin/route-usage.ts: route_usage inspection tool

Acceptance Criteria

1. When an LLM response triggers automatic continuation, the follow-up requests record continuation_of pointing to the initial attempt.
2. route_usage and session cost aggregations correctly tally all continuation chunks into the total turn cost.
3. If a continuation fails, the initial partial output and failure reason are logged accurately.
4. Unit tests in packages/core/test/ verify continuation linking and aggregated token/cost calculation.