Public-readiness review¶

Review date: 2026-08-29. Scope: package/runtime architecture, provider boundary, official A2A integration, MCP and gateway seams, security/release controls, testing, examples, and the public documentation path.

Executive decision¶

The repository and documentation site are public and suitable for technical review. The core library is substantially implemented and heavily gated; it is not a planning-only repository. Branch, tag, security, Pages, and CODEOWNERS settings remain hosted controls that maintainers must periodically audit for drift.

It should be presented as a pre-1.0 production-oriented library, not as:

a hosted agent platform;
a drop-in replacement for another product named ADK;
a turnkey, process-owning official A2A deployment;
a universal adapter for every model API;
a substitute for gateway authentication, infrastructure isolation, or operations.

Production adoption should be approved per subpackage and integration. The official A2A adapter now includes an AgentRunner executor bridge and Google ADK interoperability, while the embedding service still owns HTTP hosting, authentication, durable task state, crash recovery, subscriptions, and any push delivery it advertises.

What is strong today¶

Area	Assessment	Evidence
Core runtime	Ready with pre-1.0 stability limits	Typed agent/run contracts, bounded loop, cancellation, deadlines, retry policy, budgets, structured output, tools, approvals, idempotency, checkpoints
Provider abstraction	Ready	Runtime-facing `ModelProvider`, explicit capabilities, normalized errors/usage/streams, conformance suite
Hosted portability	Ready for supported wire contracts	Native OpenAI/Anthropic/Gemini plus Groq, xAI/Grok, and OpenRouter compatible presets
Self-hosting	Ready with deployment-owned operations	vLLM, Ollama, TGI, llama.cpp, explicit capabilities and timeouts
Tool safety	Ready with correct deployment wiring	Typed schemas, default-deny allowlists, scopes, approvals, concurrency, timeouts, result boundary, idempotency
Tenant boundary	Strong library controls; infrastructure still matters	Required tenant per run, ambient scoped context, propagation contract, store/isolation tests
MCP	Broad alpha integration	Clients, servers, transports, credential context, resilience, surface pinning, AgentGateway routing
Tesserix peer protocol	Broad alpha integration	Typed discovery, delegation, invocation, trust containment, registry caching
Official A2A	Server bridge ready with deployment-owned operations	Official 1.x cards, clients, registry/custom transport seams, bounded runner executor, cancellation, HTTP round trip, and Google ADK 2.8 card compatibility
Durability	Broad primitives and adapters	State, queues, checkpoints, leases, outbox, events, replay-safe workflow primitives
Testing	Strong	Strict fake provider, network blocking, HTTP replay, conformance suites, isolation, evals, release gates
Supply chain	Strong controls, subject to configuration	Frozen lock, dependency admissions, advisory/licence/secret scans, build provenance and trusted publishing workflows
Public onboarding	Current in source	Lean README, quickstart, capability map, command-line guide, runnable cookbook, migration and interoperability guides, generated API reference, strict docs build

At audit start, make check passed 8,518 tests with 99.16% coverage, strict Ruff and mypy, import boundaries, API/event/release gates, and dependency policy. The offline getting-started example and package build also passed. Final verification after the review's changes is recorded at the end of this page.

Gaps that must stay explicit¶

Priority 0: required before claiming a turnkey official A2A deployment¶

Inject a durable, tenant-scoped official TaskStore and reconcile tasks left working after a process crash.
Authenticate every route and authorize task, context, message, artifact, subscription, and cancellation objects in server/gateway context.
Add signed push notifications, callback validation, retries, and dead letters only if the deployment advertises push support.
Run deployment-specific conformance and fault tests against its selected gateway, task store, and at least one additional external A2A implementation.

The bridge now maps bounded text requests, final artifacts, terminal failures, and cancellation. Card security metadata remains descriptive and cannot replace enforcement.

Priority 0: required for the public GitHub launch¶

Keep the public repository description, homepage, topics, social preview, and default branch protection current.
Keep GitHub Pages on GitHub Actions and monitor documentation deployments.
Keep private vulnerability reporting, secret scanning, push protection, Dependabot alerts, and code scanning enabled where the organization plan permits them.
Audit required checks and reviewed-pull-request enforcement for main after workflow or ownership changes.
Verify the PyPI project, trusted publisher, release environments, and package ownership are suitable for public contributors.
Keep the organization-owned support@tesserix.app conduct-reporting rota private, monitored, and current.

Priority 1: reliability and adaptability¶

Add dedicated adapters for Azure OpenAI, Amazon Bedrock, Vertex AI, and any other non-compatible API that is a supported product requirement.
Add scheduled credentialed integration tests against real provider sandboxes. Keep the default suite offline and replay-first.
Automate model-catalogue freshness and require a reviewed effective date for price, context, and capability changes.
Add fault-injection and soak tests against real Redis, PostgreSQL, NATS, Temporal, MCP, gateway, and A2A deployments, including restarts and partial network failure.
Define supported production SLO profiles for provider latency, queue lag, registry staleness, checkpoint recovery, and trace loss.
Add a standard signed-card verifier or document the one trust service every deployment must use.
Keep the project-qualified tesserix-adk dispatcher limited to self-contained commands; application-wired CLIs continue to accept explicit storage/build callbacks.
Reduce the very broad pre-1.0 public API before 1.0. Keep new experimental seams out of stable re-export modules until consumers need them.
Keep the capability map, task guides, generated API reference, runnable recipes, and release-tagged Pages site synchronized as the public surface moves.

Priority 2: ecosystem usability¶

Publish framework-specific examples for FastAPI/ASGI lifecycle, Kubernetes shutdown, background workers, and serverless constraints.
Publish a complete sample combining a model gateway, MCP, official A2A discovery, durable state, telemetry, and policy without embedding real infrastructure credentials.
Consider TypeScript/Java/Go clients only for cross-language protocol needs; do not copy the Python runtime merely for ecosystem parity.

Failure scenarios reviewed¶

Scenario	Existing defense	Remaining deployment responsibility
Provider is slow or unavailable	Phase/run deadlines, normalized transient errors, opt-in jittered retry, fallback rules	Provider SLO, quota, regional strategy
Provider claims success with an error body	Compatible adapter refuses it	Vendor-specific regression recordings
Provider omits usage or tool IDs	Preset reconciliation and estimated-count attribution	Validate every deployed model/server version
Model lacks a required feature	Explicit capability gate before request	Keep capability records current
Tool call is duplicated	Idempotency declaration/store and stable keys	Durable shared store and downstream idempotency
Process dies mid-effect	Checkpoint frontier and indeterminate disposition	Recovery runbook and status/read-back APIs
Event is delivered twice	Idempotent consumer and transactional outbox	Broker retention and dead-letter operations
Registry is unavailable	Tesserix peer cache/stale policy; official registry is injectable	Official registry caching/availability policy in application
Registry substitutes an agent	Exact card-name check plus optional verifier	Signature, endpoint, issuer, tenant policy
Gateway is bypassed	No automatic direct fallback in AgentGateway route	Network policy and service identity
A2A card says Bearer	Official security metadata emitted	Actual token validation and per-object authorization
One tenant guesses another object ID	Tenant-scoped runtime/store contracts and isolation suite	Database RLS/partitioning and gateway enforcement
Tool or peer returns instructions	Untrusted result/peer boundaries and guardrails	Domain-specific allowlists and human approval
Credential rotates	Secret resolved at use time; pools can retire old key	Secret-manager availability and rotation procedure
Telemetry backend fails	Observation path designed to fail open and redact	Buffering, sampling, alerting, retention
Docs drift	Link test and strict MkDocs build	Executable snippet and owner review expansion

Comparison position¶

The public site linked during this review represents the broader “Agent Development Kit” category. Tesserix ADK should not compete on the generic acronym alone. Its credible differentiation is:

explicit production policy and tenant boundaries;
infrastructure and protocol substitutability;
conservative provider capability handling;
deterministic tests and compatibility gates;
clear separation of MCP, official A2A, and the richer Tesserix peer protocol;
a lean base dependency graph.

It is not a fork of Google's ADK and does not promise source compatibility with it. Users should choose Tesserix ADK when those production controls and replaceable boundaries are the requirement, not because both projects share three initials.

Public launch checklist¶

[x] Accurate root README and five-minute path
[x] Step-by-step custom agent guide
[x] Groq, xAI/Grok, OpenRouter, native, self-hosted, and gateway recipes
[x] Official A2A support/limitation matrix
[x] Integration and registry guidance
[x] Apache-2.0 licence file
[x] Root contribution guide and pull-request template
[x] Strict local documentation build and link test
[x] Complete capability map and installed-command guide
[x] Generated API reference and public-symbol recipe coverage
[x] Staged migration and framework-interoperability guides
[x] GitHub Pages workflow in source
[x] GitHub Actions pinned to reviewed commit SHAs
[x] Repository made public
[x] Pages source enabled and deployed
[x] Public security settings verified
[x] Organization-owned conduct-report contact chosen and published privately
[ ] PyPI/trusted-publisher ownership verified
[x] First external clean-room onboarding completed

Verification record¶

Google ADK/A2A bridge — 2026-08-28¶

Final verification used Python 3.13.15:

make check: passed lint, formatting, import-boundary, strict typing, dependency, admission, disclosure, API, event, replay, deprecation, release-note, documentation, and complete test gates; 9,148 tests passed, 2 were skipped, 125 were deselected by the coverage profile, and total coverage was 98.97%.
make audit, make secrets, make licences, make deps, and make admissions-check passed with no locked advisories, credential-shaped values, licence-policy violations, dependency-policy violations, or unrecorded consumer dependencies.
uv lock --check passed for the 153-package all-extras development graph. The inventory now follows dependency-selected extras, so the official A2A HTTP server and psycopg binary/pool packages carry their true profiles and package-specific licence decisions.
uv build produced the source archive and universal Python wheel. The wheel contains the typed-package marker, A2A executor, and Google ADK helper; the source archive also contains the Code of Conduct.
A fresh public clone of commit 0ae26f0 passed uv sync --frozen --extra google-adk, the offline Google ADK/A2A example, all 11 focused bridge tests, and uv build without relying on the development worktree.
The offline official HTTP round trip passed through the official client, Starlette route, request handler, task store, Tesserix executor, runner, and final artifact. Google ADK 2.8.0 accepted the same official card through its current non-legacy A2A path.

Documentation and Pages synchronization — 2026-08-29¶

make docs-check, the API-surface and generated-reference checks, and public-symbol recipe coverage passed. Regression tests additionally proved that every Markdown page is navigable, the command-line guide covers all eight installed commands, and the release runbook names every release workflow job.
The rebuilt site rendered the capability map and command-line guide correctly. All 64 distinct external links exposed by the site and README resolved successfully.
Ruff, formatting, import-boundary checks, strict mypy across 557 source files, strict Pyright, the repository policy/contract gates, and uv build passed. The complete all-extras test run passed 9,428 tests with 2 platform/runtime cases skipped, 125 intentionally deselected, and 96.55% total coverage.
The public Pages root redirects to stable/, whose 0.53.0 sitemap contained 189 URLs; every URL returned HTTP 200. The tagged release workflow completed both documentation jobs and published the exact tesserix_adk-0.53.0-py3-none-any.whl release asset.
Pages is deliberately release-tagged rather than deployed from the moving main branch. At review time the live canonical documentation remained 0.53.0; the reviewed 0.53.1 release publishes these source updates. PyPI publication remains disabled, so GitHub Release assets are the supported distribution channel.

These source and artifact checks do not replace the unchecked PyPI trusted-publisher and package-ownership actions in the launch checklist.