-
Notifications
You must be signed in to change notification settings - Fork 92
Roadmap
Living page. The long-form plan is in the design dossier: 10-mvp-and-roadmap.md (phases P0–P5, build-vs-steal ledger) and the "What's next" section of guides/OVERVIEW.md. Month-by-month progress: Project Updates.
- Observe — OTLP ingest from any source, blob-first durability, first-class agent / conversation / turn / step columns, hierarchical trace explorer, conversation replay timeline + Fleet view, state deltas, declared agent catalog.
- Detect — DB-backed evaluators as table columns (structural + multi-level LLM judge, basic/advanced templates, batch/sequential, targeting + sampling, advisory), live SSE scores, rolling summary, judge calibration against human labels.
- Triage — structural + semantic failure clustering with analysis, suggested fix and severity; near-duplicate merging.
- Test — one-click promotion of a trace or cluster into a hermetic case with a fail-to-pass contract.
-
Ship —
tracely simulate/replay/gate, the GitHub Action, scenarios (scripted + adversarial), honest verdicts (UNGRADED,NO_COVERAGE). - Insights — Trends, per-agent meta-analysis, ops metrics, alerts.
-
Platform — three auth modes, organizations + workspaces, hosted-cloud billing (20k traces/month free, BYO LLM key), MCP server at
/mcp, the Tracely agent skill, an in-app assistant that operates the product as you, one-click Railway deploy, self-watching deployment.
- Alerts flow builder — drag-and-drop rule flows (condition → Slack / email / webhook / LLM prompt / Python expression), a rule assistant that drafts the flow from a sentence, and "Run test" against a real recent failure.
- Security hardening pass — invite takeover, SSRF via customer URLs, root ingest keys and auth rate-limiting closed 2026-08-21; remaining items tracked in guides/SECURITY_AUDIT.md.
- Hermetic replay for every provider — OpenAI, Anthropic, Gemini, Mistral and LiteLLM are patched; Bedrock and the remaining drop-ins still fall back to live.
-
Agent-version attribution surfaced —
agent_versionsis content-hashed and cases record the version that first failed; show "v12 introduced cluster #4" on case and cluster pages, and let the gate pin to a hash. - SDK auth-aware ingest — bootstrap ingest keys from the hosted org so SaaS users never provision keys by hand.
- TypeScript / JS SDK parity — plan in 13-ts-js-parity-plan.md.
Designed in the dossier, partially built: full 7-signal failure detection + RCA + automatic test generation, the canary-as-GateRun loop, multi-agent edges + impact analysis, a zero-config GitHub App, and codebase-aware fixes.
The bet: the left half (ingest → store) borrows Langfuse's proven substrate; the right half (promote → replay → gate, judge-in-gate, calibration) is the trace-native CI layer nobody else has.
- Latency / cost ≈ 0 in hermetic replay, so the soft delta gates are informational there.
- Backups are an operator action in your provider; Tracely doesn't take them.
- Python SDK only today (non-Python services use raw OTLP + the HTTP-endpoint gate, which needs no SDK).
Have a request? Open a feature request.
Tracely — trace-native CI/CD for AI agents · MIT · Contributing · Security · Found a mistake on this page? Edit it, or open an issue.