Add agent run dashboard and telemetry setup guide

[VibeSWE]

Agents previously listed individual telemetry rows, making one invocation appear as unrelated entries. This change groups related spans into observed runs with status, duration, models and available input/output tokens. Details expose the span tree, model/tool calls, errors, captured conversation and raw records.

Agents, Traces and Logs preserve captured time bounds and each view’s filters across navigation, reload and browser history. Focused Logs links retain grouped criteria and replace a previous trace focus. Missing telemetry stays visibly incomplete; event/link duplication does not inflate usage. Separate 100-trace and 1,000-record notices disclose coverage limits. A trace is the run boundary, and summaries describe loaded records.

Set up telemetry in Agents now provides Claude Code and Codex configuration, also linked from Get started and the empty state. Users can customize an existing collector URL and copy JSON/TOML or a coding-agent brief. The guide explains configuration merging, content opt-in limits, required field mappings and validation. Claude uses per-signal destinations without disabling existing metrics; Codex guidance prevents duplicate TOML tables. Opening the panel preserves filters and does not change configuration or activate exporters.

Evaluations and cost estimates remain deferred until telemetry and pricing inputs are reliable. Native Claude Code/Codex export compatibility remains unverified; receiving provider telemetry alone does not guarantee complete dashboard runs.

Validation is recorded in frontend/docs/validation-review.md:

• Original dashboard: formatting, TypeScript/build, 2,283 unit tests, 537 distinct application cases across broad/focused runs, 22 production tests and 232 Storybook tests passed. Final navigation changes passed 65 repeated executions without retries plus 16 Traces/Logs handoff cases. Two optional app screenshots and 14 opt-in/existing unit cases were skipped.
• Final setup-guide follow-up: TypeScript/build, 62 focused unit tests, 28 application tests, 8 production mount tests and 22 relevant Storybook tests pass, including clipboard failure, empty/error states, mobile accessibility, preserved query bounds and / + /next routes. Four generated JSON/TOML variants pass real parsers. The full application suite was not repeated for this follow-up.
• Three real-OTLP tests passed against disposable Parseable v3.2.4 for the original dashboard; this used the published image, not rebuilt Rust. No deployment or native CLI exporter activation was performed.

Mayfly implementation record and setup UI decision/demo document the behavior and screenshots. Sol 6.1 implementation/testing, Opus 5.5 reviews and Luna wiki discovery contributed. The final setup review confirmed all initial findings fixed with no remaining regressions or blockers.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[comment-clanker]

Reviewed 870a554197ba8a85b59030157a9de2f80d3170c3 with four GPT-6.1-Sol and four Claude Opus 5.5 reviewers. Actionable findings are in the threads below.

Validation through the Swamp Blacksmith workflow: TypeScript checking, 72 focused unit tests and 42 existing browser tests passed without retries. Targeted reproductions separately checked the reported edge cases. Native CLI exporter compatibility and live ingestion were not retested.