feat(opencode): show round usage metadata

[junmo-kim]

Problem

OpenCode previously exposed only the final ACP usage row to the UI. Tool-loop calls, compaction summaries, and synthetic continuations could therefore be omitted from the response metadata popover.

Solution

This change snapshots OpenCode persisted message IDs before each prompt, fetches the post-prompt delta, and aggregates every valid assistant row by provider/model. It sends an HAPI-owned round-summary carrier before ready and reuses the existing response-group metadata popover. Malformed data, duplicate IDs, unavailable APIs, parse failures, timeouts, and aborted requests fail closed by omitting only the summary.

Screenshots

Desktop (1440×1000)	Mobile (390×844)

Tests

• Added CLI coverage for snapshot Abort/Switch, the fetch-plus-JSON deadline, duplicate IDs, strict aggregation, and carrier ordering.
• Added Web coverage for strict carrier normalization, persisted wire grouping, tool-only responses, late metadata reconciliation, and legacy compatibility.
• bun typecheck
• bun run test
• bun run build
• git diff --check
• Manually verified actual OpenCode 1.18.21 ACP with opencode/big-pickle end-to-end on desktop and mobile viewports, including a two-step tool loop and summary-before-ready ordering.

[github-actions]

Findings

• [Major] Keep telemetry off the prompt critical path — captureOpencodeRoundSnapshot() performs a full /message fetch and JSON parse before session.onThinkingChange(true) and backend.prompt(). Its deadline is 1 second, so a slow/unavailable OpenCode HTTP API delays every user prompt by up to a second; because the endpoint returns the complete history, the work also grows with session length. Evidence: cli/src/opencode/opencodeRemoteLauncher.ts:627, cli/src/opencode/utils/opencodeRoundSummary.ts:81, cli/src/opencode/utils/opencodeRoundSummary.ts:90.
  Suggested fix:

  private nextRoundSnapshot: OpencodeRoundSnapshot | null = null
  
  const roundSnapshot = this.nextRoundSnapshot
      ?? await captureOpencodeRoundSnapshot({ baseUrl, sessionId: acpSessionId, signal })
  this.nextRoundSnapshot = null
  
  // Return the validated post-fetch ID snapshot alongside the optional summary.
  const result = await fetchOpencodeRoundSummary({ ...options, snapshot: roundSnapshot })
  if (result) {
      this.nextRoundSnapshot = result.snapshot
      if (result.summary) this.handleAgentMessage({ type: 'round_summary', summary: result.summary })
  }

Summary

• Review mode: initial
• One major performance regression found in the OpenCode prompt path.

Testing

• Not run (automation). GitHub test, integration, and drift-gate checks pass.
• Add a launcher regression test proving a second prompt reuses the prior post-fetch snapshot and does not wait on another pre-prompt full-history request.

HAPI Bot

[github-actions]

Findings

• No high-confidence issues found in the latest diff.

Summary
Review mode: follow-up after new commits

The prior Major finding is addressed: the launcher now reuses the validated post-fetch snapshot for subsequent prompts and invalidates it before explicit compaction. Residual risk remains around compatibility with OpenCode's internal message API; the implementation fails closed by omitting metadata.

Testing

• Local tests not run: Bun is unavailable in the review environment.
• GitHub test, integration, and fixture drift-gate checks pass.
• git diff --check passes.

HAPI Bot