v0.1.0 — Initial public release

First public availability of Mesedi. Ships as MIT source (Self-Hosted)
and hosted service (Cloud Hobby, Cloud Team, Cloud Production, Cloud Enterprise).

Failure-class detectors (20)

Cost + performance: context_overflow, token_waste, cost_velocity.

Reliability: crashes, tool_failures, validator_failures,
tool_schema_drift, provider_incident, infrastructure_throttled.

Multi-agent: cascading_failure, coordination_deadlock, drift.

Loops: semantic_loop, loops.

Human-in-the-loop: hitl_timeout, hitl_rejection_spike.

Security: prompt_injection, data_leakage, sandbox_escape,
grounding_failure.

Several classes carry sub-signatures rather than being separate
classes: loops folds in identical-call, similar-call, step-count, and
time-budget detection; drift covers lexical drift and new-model
drift; prompt_injection matches six named attack patterns
(jailbreak/DAN, ignore-instructions, role-override, system-prompt
inject, instruction-tag, developer-mode).

Each detector produces a stable signature that clusters recurring
failures into a single group. First-occurrence webhook notification;
per-project tunable thresholds; custom RE2 pattern support on the
security detectors.

SDKs

• Python (mesedi on PyPI). @wrap and @tool decorators,
  AsyncShipper, auto-instrumentation for Anthropic / OpenAI / Cohere /
  Gemini / Ollama (sync, async, streaming), hard-halt with local
  budgets + SSE remote channel, LangGraph and OpenAI Agents adapters,
  optional [langchain] / [crewai] extras.
• TypeScript (mesedi on npm). Feature parity with Python.
  mesedi/integrations/vercel_ai adapter for Vercel AI SDK's
  generateText.
• gzip request compression on both SDKs (backend decompresses
  transparently).
• Payload truncation with per-project tool_return_value_max_bytes
  caps.

Cross-tenant provider signal

provider_incident clusters cross-tenant errors from a single LLM
provider so a customer can tell "the provider is having problems" vs
"my code broke." Canonical error class vocabulary
(rate_limited, quota_exhausted, internal_error,
service_unavailable, timeout, invalid_api_key, client_error,
unknown) mapped uniformly across all four providers. retry_after
extracted onto the event where the provider returns it.

RAG grounding

Ragas, Promptfoo, and Vectara HHEM grounding evaluators supported via
SDK helpers. Threshold config surfaces as PerEvaluatorFloors in the
grounding_failure detector so different evaluators can use
different sensitivities.

Dashboard

Failure-group + execution surfaces with AI root-cause analysis
(Anthropic Haiku on Hobby and Team — pay-per-use on Hobby, 200
included per period on Team; a more capable model with 2,000 included
per month on Production and Enterprise). Per-project settings for detector thresholds, custom security
patterns, DLP severity policy, HITL fire modes. Audit log (Cloud Team
and above). Per-project billing cap with configurable ceiling.

Instrumentation & telemetry

OpenTelemetry parallel emission. Human-in-the-loop lifecycle with
timeout, rejection, and approval events. Multi-agent handoff tracking
with agent-name propagation.

Foundations

FOUNDATION discipline: pre-audit + post-audit + 5-line audit trail
required on every commit, enforced via a commit-msg git hook.
CI drift check (tools/check-tier-constants.sh) asserts every tier
constant + feature claim in the dashboard matches the backend
enforcement site.