Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 10 additions & 1 deletion CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,16 @@

All notable changes to the E2A framework documents and reference code.

## v1.8.0 — Service-to-service identity (hub design note SI-01)
## Unreleased — documents

### Removed
- **E2A Multi-Cloud Reference (PDF).** It had grown out of date: it showed the v1.1 API and the retired `BedrockLLMService`-style classes, used LangGraph as the reference, named retired models, and treated Meta Llama as if it were a cloud. Its parts now live where they are maintained:
- concurrency, admission and capacity per cloud: [Cloud Landing Zone](docs/CLOUD_LANDING_ZONE.md) §8.1.5–8.1.6;
- ports and capability packs: the e2a-managed-providers core design;
- which model family runs on which cloud, and what each cloud adds: the [models × clouds matrix](https://github.com/subhamviky/e2a-managed-providers/blob/main/docs/design/matrix/MODEL_CLOUD_MATRIX.md), whose section 7 maps each former section to its new home.
- The Strategic Overview, README §16 and the in-memory adapters now point there.

Service-to-service identity (hub design note SI-01)

Additive, plus one fix. Services in the landscape prove who they are to each other with their own cloud's workload identity, exchanged for one issuer (Entra ID); receivers grant each caller tenants and actions as roles. No secrets.

Expand Down
2 changes: 2 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -591,6 +591,8 @@ E2A separates stable architecture contracts from provider-specific realizations.

These are **reference mappings**, not claims of semantic equivalence between provider services.

Which model family runs on which cloud, through which hosting pattern, and what each cloud adds around the model (guardrails, retrieval, runtimes, gateways) is mapped in the [models × clouds matrix](https://github.com/subhamviky/e2a-managed-providers/blob/main/docs/design/matrix/MODEL_CLOUD_MATRIX.md) of e2a-managed-providers. The concurrency primitives per cloud are in [Cloud Landing Zone §8.1.6](docs/CLOUD_LANDING_ZONE.md).

### Portability claim

Agent implementations that depend exclusively on supported E2A contracts can remain largely unchanged across provider adapters. Deployment configuration, infrastructure, provider capabilities, and adapter implementations can still differ.
Expand Down
Binary file removed docs/E2A_MultiCloud_Reference.pdf
Binary file not shown.
2 changes: 1 addition & 1 deletion docs/E2A_STRATEGIC_OVERVIEW.md
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@ E2A is the governed bridge between the two. It treats a model call as a rate-lim
| **Decoupled inference service** | `BaseLLMAgentService` is the one shared owner of model concurrency: a bounded tenant-and-pod permit, then RPM/TPM quota (refunded on failure), then the upstream call under the request deadline — so one tenant's burst is rejected early instead of starving others. It also owns context budget, caching hooks and stream DLP; subclasses implement only vendor wire transport. | [Playbook §2.18](IMPLEMENTATION_PLAYBOOK.md) |
| **Concurrency & overload** | Every shared-resource risk — duplicates, lost updates, stale owners, saturation, orphaned work — has one named owner: atomic idempotency claim, conditional commit, lease with fencing, bounded admission, per-dependency bulkhead, retry budget. | [Playbook §2.19](IMPLEMENTATION_PLAYBOOK.md) · [config-reference](config_reference.md) |
| **One central configuration registry** | Every key and environment variable declared once (`E2A_CONFIG`) and resolved by one function: kwargs › config › environment › default. | [Playbook §5](IMPLEMENTATION_PLAYBOOK.md) · [config-reference](config_reference.md) |
| **Vendor and cloud portability** | Model, storage, lock, cache and queue are ports; switching provider is a binding plus a config value, not a rewrite. | [MultiCloud Reference](E2A_MultiCloud_Reference.pdf) |
| **Vendor and cloud portability** | Model, storage, lock, cache and queue are ports; switching provider is a binding plus a config value, not a rewrite. | [Cloud Landing Zone §8.1.6](CLOUD_LANDING_ZONE.md) · [Models × clouds matrix](https://github.com/subhamviky/e2a-managed-providers/blob/main/docs/design/matrix/MODEL_CLOUD_MATRIX.md) |
| **Explicit context propagation** | Six fields on every call: `correlation_id`, `io_config`, `idempotency_key`, `tenant_id` (inputs) and `message_log`, `failed_keys` (outputs). Two are additive: `trace_context` (W3C) and an optional request `deadline`. Since v1.2 the inputs travel as one frozen request context (`ctx`), and a hook declares only what it uses, including separate read and modify accessors for tools. | [Playbook §2.1](IMPLEMENTATION_PLAYBOOK.md) |
| **Enforced non-functional requirements** | Capacity, token-budget, firewall and approval checks run as ordered governance stages and refuse with a typed reason (v1.2; each refusal is still the v1.1 exception, e.g. `NFRViolationError`); low confidence falls back before anything material runs; a latency breach is recorded; CI-time checks (`python -m e2a.conformance`, run by `BasePipeline._verify_scaffold_contracts`) block merges that override framework-owned methods; an evaluation gate (RAGAS faithfulness ≥ 0.85) blocks deployment. | [Playbook §2.12](IMPLEMENTATION_PLAYBOOK.md) |
| **Human-approval hibernation** | `DehydrationInterrupt` persists workflow state and releases compute while waiting for an approval callback. | [Playbook §2.11](IMPLEMENTATION_PLAYBOOK.md) |
Expand Down
2 changes: 1 addition & 1 deletion reference/e2a_inmemory_adapters.py
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@
(atomic claim, conditional commit, fenced writes, leases, cache). They are
single-process and NOT for production: a real adapter binds the same
contracts to a managed store whose conditional-write / lease primitive
provides the atomicity (see the Multi-Cloud Reference).
provides the atomicity (see docs/CLOUD_LANDING_ZONE.md §8.1.6).
"""
import threading
import time
Expand Down