Record every decision failure and count them at session end

docs/identity/overview.md

@@ -121,7 +121,7 @@ A concrete example that illustrates several of these challenges is Palo Alto Net
 | 1 | **Chain Integrity**<br/>*Can the full delegation chain be named and verified end-to-end?* | A single agent task can chain through five to ten identities: user → orchestrator → sub-agent → tool → downstream API. Two failures compound: the chain is not *named* (audit collapses into "the agent did it"), and the chain is not *verifiable* (downstream resources cannot confirm the chain without trusting every intermediary). In practice the full chain is often not reconstructable end-to-end, and there is a known class of attacks where the chain itself can be spliced. | **[OAuth 2.0](https://www.rfc-editor.org/rfc/rfc6749) access tokens** identify the bearer, not the chain.<br/>**[OIDC Core 1.0](https://openid.net/specs/openid-connect-core-1_0.html) ID tokens** name the human at the start but lose granularity at each hop.<br/>**[SPIFFE SVIDs](https://spiffe.io/docs/latest/spiffe-about/spiffe-concepts/)** identify the workload but not which call within it.<br/>**[OAuth Token Exchange (RFC 8693)](https://www.rfc-editor.org/rfc/rfc8693)** defines nested `act` claims that can express a chain, but per [§4.1](https://www.rfc-editor.org/rfc/rfc8693#section-4.1) *"prior actors identified by any nested `act` claims are informational only and are not to be considered in access control decisions."* The chain is recorded, not normatively enforceable.<br/>**[OAuth Identity Chaining draft](https://datatracker.ietf.org/doc/draft-ietf-oauth-identity-chaining/)** and **[Transaction Tokens draft](https://datatracker.ietf.org/doc/draft-ietf-oauth-transaction-tokens/)** address parts but specify neither chain verification rules nor revocation propagation as required runtime behavior.<br/>**[W3C Trace Context](https://www.w3.org/TR/trace-context/)** correlates but [does not authenticate](https://www.w3.org/TR/trace-context/#privacy-of-traceparent-field). | Every action must produce, at the moment of execution, a signed record that names the specific link in the chain and lets any participant (including the downstream resource) verify the entire chain end-to-end without trusting any intermediary. Cryptographic chain structure, mandatory scope attenuation at every hop, revocation propagation. | Pending |
 | 2 | **Over-Privilege**<br/>*Was this action within the agent's purpose?* | Agents inherit broad standing privileges. Service accounts. Long-lived API keys. OAuth scopes designed for human applications. One compromised agent has the union of every permission anyone has ever granted to it. The fix is not bigger scopes; the fix is no standing privilege. | **[OAuth scopes (RFC 6749 §3.3)](https://www.rfc-editor.org/rfc/rfc6749#section-3.3)** are coarse, deployment-defined strings (e.g. `read`, `write`, `repo`). Per **[RFC 9396 §1](https://www.rfc-editor.org/rfc/rfc9396#section-1)**, scope is "sufficient... for coarse-grained authorization requests" but "not sufficient to specify fine-grained authorization requirements" — it does not express per-action, per-resource, per-data-cell intent.<br/>**RBAC and ABAC** scope by role or attribute, granted once, evaluated coarsely.<br/>**AIMS** does not require **[RAR (RFC 9396)](https://www.rfc-editor.org/rfc/rfc9396)** and treats scope semantics as deployment-specific. | Credentials minted just before the call, scoped to one downstream callee, structured (per-action, per-resource), expired in minutes. Authorization evaluated per-call against the immutable intent for the run, not per-session. | Pending |
 | 3 | **Token Theft Resistance**<br/>*If a token leaks from agent memory, is it useless to the attacker?* | The agent threat model breaks OAuth's bearer-token assumption entirely. Prompt injection can exfiltrate tokens from memory. Tool outputs can leak tokens. Sub-agents can replay tokens. Logs and crash dumps capture tokens. Long-lived refresh tokens in agent memory are the worst case: a durable credential held in storage that any indirect prompt injection can read. | **[OAuth 2.0 Bearer Tokens (RFC 6750 §1.2)](https://www.rfc-editor.org/rfc/rfc6750#section-1.2)** explicitly treat possession as sufficient: *"Using a bearer token does not require a bearer to prove possession of cryptographic key material."*<br/>**API keys** are bearer secrets by definition.<br/>**[OAuth refresh tokens (RFC 6749 §1.5)](https://www.rfc-editor.org/rfc/rfc6749#section-1.5)** were designed for the same user returning later to resume a session, which does not describe how agents work.<br/>**[DPoP (RFC 9449)](https://www.rfc-editor.org/rfc/rfc9449)** and **[OAuth 2.0 Mutual-TLS Certificate-Bound Access Tokens (RFC 8705)](https://www.rfc-editor.org/rfc/rfc8705)** solve sender-constraint but are optional extensions that most deployments do not enforce.<br/>None of these are required by the OAuth core. | All non-trivial tokens sender-constrained (DPoP or mTLS) so a leaked token is useless without the corresponding private key. Aggressive lifetime ceilings measured in minutes or seconds, not hours. No long-lived secrets in model-visible context. Credentials fetched at task start, used within a tight window, discarded at task end. | Pending |
-| 4 | **Last-Mile Enforcement**<br/>*Did the resource actually enforce the decision, on every call, inside the runtime?* | The moment the action reaches a production system. All upstream identity work either holds here or fails here. Most systems today cannot distinguish an authorized action by an agent acting in good faith from the same call by an agent hijacked by indirect prompt injection. Gateway-only enforcement misses tool calls and sub-agent invocations that never cross a gateway. Cached decisions miss revocation, scope change, and freshly-discovered taint. | **API gateways and service meshes** enforce at the edge; agent tool calls never reach them.<br/>**[OAuth 2.0 resource servers (RFC 6749 §1.1)](https://www.rfc-editor.org/rfc/rfc6749#section-1.1)** validate the token but not the action's relationship to intent.<br/>**[mTLS (RFC 8705)](https://www.rfc-editor.org/rfc/rfc8705)** and **[SPIFFE](https://spiffe.io/docs/latest/spiffe-about/spiffe-concepts/)** establish workload identity but do not enforce per-action policy.<br/>None fire inside the agent runtime where the decision is made. | Enforcement fires at every tool call, inside the agent runtime, before the action executes. The runtime verifies identity, scope, taint, freshness, and intent on every call. The LLM call cost dwarfs the re-authorization cost, so caching decisions is not required for performance. Default-deny verified by a negative test in CI. | **Partially specified.** The ACS Instrument spec already defines the honoring-decisions contract in [§6.4 (Honoring Decisions, Normative)](../spec/instrument/specification.md#64-honoring-decisions-normative): the Observed Agent MUST wait for the Guardian's decision up to the negotiated timeout, MUST apply it, and MUST record every fail-open proceed as an audit event. Note the spec's current default on timeout is **fail-open with audit**, which is in tension with this row's default-deny / no-cache goal; whether identity-tier actions should tighten that to fail-closed is an open work item for this workstream. |
+| 4 | **Last-Mile Enforcement**<br/>*Did the resource actually enforce the decision, on every call, inside the runtime?* | The moment the action reaches a production system. All upstream identity work either holds here or fails here. Most systems today cannot distinguish an authorized action by an agent acting in good faith from the same call by an agent hijacked by indirect prompt injection. Gateway-only enforcement misses tool calls and sub-agent invocations that never cross a gateway. Cached decisions miss revocation, scope change, and freshly-discovered taint. | **API gateways and service meshes** enforce at the edge; agent tool calls never reach them.<br/>**[OAuth 2.0 resource servers (RFC 6749 §1.1)](https://www.rfc-editor.org/rfc/rfc6749#section-1.1)** validate the token but not the action's relationship to intent.<br/>**[mTLS (RFC 8705)](https://www.rfc-editor.org/rfc/rfc8705)** and **[SPIFFE](https://spiffe.io/docs/latest/spiffe-about/spiffe-concepts/)** establish workload identity but do not enforce per-action policy.<br/>None fire inside the agent runtime where the decision is made. | Enforcement fires at every tool call, inside the agent runtime, before the action executes. The runtime verifies identity, scope, taint, freshness, and intent on every call. The LLM call cost dwarfs the re-authorization cost, so caching decisions is not required for performance. Default-deny verified by a negative test in CI. | **Partially specified.** The ACS Instrument spec already defines the honoring-decisions contract in [§6.4 (Honoring Decisions, Normative)](../spec/instrument/specification.md#64-honoring-decisions-normative): the Observed Agent MUST wait for the Guardian's decision up to the negotiated timeout, MUST apply it, and MUST record every decision failure as an audit event. Note the spec's current default on timeout is **fail-open with audit**, which is in tension with this row's default-deny / no-cache goal; whether identity-tier actions should tighten that to fail-closed is an open work item for this workstream. |
 | 5 | **Proof of Intent**<br/>*Was the action bound to an immutable, declared purpose that survives mid-task reframing?* | OAuth scopes say what the user *can* do. They do not say what the user *is doing right now*. Without intent binding at runtime, every agent action is a guess about user authorization. The agent's reasoning can be reframed mid-task by an instruction hidden in a document or a tool output. The original intent has to survive that. This is the hardest of the five, because it is the only one OAuth cannot patch with an extension: the underlying problem is non-deterministic client behavior under adversarial inputs. | **None at the identity layer.** Intent is implicit in the OAuth consent screen and discarded after token issuance.<br/>**[RFC 9396 RAR](https://www.rfc-editor.org/rfc/rfc9396)** can carry purpose claims as `authorization_details` data but does not bind them as runtime invariants against the LLM client.<br/>**IBAC (Intent-Based Access Control)** is research-stage with no IETF/W3C standard track.<br/>No production standard treats intent as a first-class, immutable, normatively-enforced runtime invariant. | Intent treated as a first-class governance concept with a normative immutability invariant. Intent must flow across all three ACS pillars: Instrument fires decisions against it, Trace records it as a span attribute, Inspect catalogs what it authorized. Intent extension permitted only through an explicit, human-approved path. | Pending |
 
 ---

docs/spec/conformance.md

@@ -34,7 +34,7 @@ A v0.1.0-conformant deployment MUST implement ACS-Core. ACS-Core comprises:
 - **SessionContext and Intent**: `session_id`, `chain_hash` (rolling SHA-256), append-only ContextEntry chain, with the Guardian publishing the chain head (`chain_hash`) on responses for content-bearing steps ([§8](./instrument/specification.md#8-sessioncontext-and-intent)). Intent is optional but normative when IBAC is the enforcement paradigm.
 - **Replay protection** — `request_id` (UUID) and `timestamp` on every request; Guardians MUST reject replays per [§10.3](./instrument/specification.md#103-replay-protection).
 - **Baseline integrity**: every request and response carries a signature over the canonical envelope ([§10](./instrument/specification.md#10-cryptographic-signatures)). `HMAC-SHA256` with an HKDF-derived per-session key from deployment-provided key material is the baseline; asymmetric and post-quantum algorithms are the ACS-Crypto profile.
-- **Decision honoring**: the Observed Agent MUST wait for the Guardian's decision up to the negotiated timeout and apply it; on a decision failure (timeout, transport failure, or an error without a decision) it applies the `on_decision_failure` posture (default `proceed`, fail-open) and records every fail-open proceed as an audit event ([§6.4](./instrument/specification.md#64-honoring-decisions-normative)). Handshake failure follows the deployment's startup posture ([§4.1](./instrument/specification.md#41-handshake-failure-normative)).
+- **Decision honoring**: the Observed Agent MUST wait for the Guardian's decision up to the negotiated timeout and apply it; on a decision failure (timeout, transport failure, or an error without a decision) it applies the `on_decision_failure` posture (default `proceed`, fail-open) and records every decision failure as an audit event, fail-open or fail-closed ([§6.4](./instrument/specification.md#64-honoring-decisions-normative)). Handshake failure follows the deployment's startup posture ([§4.1](./instrument/specification.md#41-handshake-failure-normative)).
 - **Liveness** — `system/ping` SHOULD be implemented: the Observed Agent declares it in `methods_implemented` and sends it, and the Guardian answers it per [§13](./instrument/specification.md#13-liveness-system-methods). A deployment MAY omit it only if the Observed Agent relies on an alternative liveness mechanism that the deployment's configuration names (for example: TCP keepalive or HTTP/2 PING on HTTP transports; process supervision or an out-of-band heartbeat on stdio; or continuous observed hook traffic where the session's own cadence acts as the probe). ACS does not put that declaration on the wire in v0.1, as [§9.2](./instrument/specification.md#92-approver-incapable-clients-normative) does not for approver capability. Omitting `system/ping` without a declared alternative is non-conformant: the chain of defaults (`startup_posture: proceed`, `on_decision_failure: proceed`, no liveness) produces silent fail-open on Guardian outage, which the Core baseline cannot allow.
 - **Wrapped MCP** — `protocols/MCP/*` ([Hooks](./instrument/hooks.md#protocolsmcp)) MUST be implemented by deployments whose sessions involve MCP at any point (tools, resources, prompts, notifications), so that `protocols/MCP/resources/read` — the canonical prompt-injection ingress for attacker-controlled external content — is not bypassed. MCP `tools/call` MAY instead flow through the generic `steps/toolCallRequest` / `steps/toolCallResult` hooks when tool-level policy suffices; the requirement exists for the surfaces the generic hooks cannot see (capability negotiation, prompt fetches, resource reads, notifications). Support is declared in the handshake via `wrapped_protocols`. Deployments whose sessions never involve MCP MAY omit the namespace. **Known v0.1 limitation:** `wrapped_protocols` is declared once, at handshake, and v0.1 defines no renegotiation. A session that declared no MCP and then gains an MCP server mid-session (a mutation `agbom/changed` reports, and the `acs-inspect-dynamic` profile exists to surface) cannot begin wrapping mid-session: the generic tool hooks still govern that server's tool calls, the addition is visible through `agbom/changed`, and wrapped coverage begins at the next session's handshake. Renegotiation is v0.2 scope.
 

docs/spec/instrument/hooks.md

@@ -317,7 +317,7 @@ Schema: [`hooks/session-end.json`](https://genai-security-project.github.io/agen
 
 Session termination, audit finalization. The Guardian seals the chain at this point.
 
-**Payload:** `session_reason` (`completed`, `cancelled`, `error`, `timeout`), final aggregates.
+**Payload:** `reason` (`completed`, `cancelled`, `error`, `timeout`, `abandoned`), and an optional `summary` with the outcome, step count, errors, and `decision_failures`, the number of decision failures in each failure class (Specification §6.4).
 
 **Decision:** Audit only.
 

docs/spec/instrument/specification.md

@@ -99,7 +99,7 @@ sequenceDiagram
 
 `on_decision_failure` is negotiated in the handshake, so it cannot govern a handshake that never completes. When `handshake/hello` fails — the Guardian is unreachable, silent past the deployment's connection timeout, or errors without producing a ServerHello — the Observed Agent applies its **startup posture**, configured out of band: `proceed` starts the session unguarded, `refuse` does not start it. The default is `proceed`, matching `on_decision_failure`, so a Guardian outage does not block new sessions.
 
-A Guardian that answers with a refusal (`SESSION_REFUSED`, `UNSUPPORTED_VERSION`, `PROVENANCE_REQUIRED`) has decided; the startup posture applies only when no answer arrives. A session started unguarded MUST be recorded in the deployment's own audit log — there is no Guardian to receive the event — and SHOULD be surfaced on Trace events when the deployment claims ACS-Trace. The Observed Agent SHOULD retry the handshake for subsequent sessions. Whether a Guardian can attach mid-flight to a session that started unguarded is undefined in v0.1; renegotiation is deferred to v0.2 alongside `system/handshake_renegotiate` (§13).
+A Guardian that answers with a refusal (`SESSION_REFUSED`, `UNSUPPORTED_VERSION`, `PROVENANCE_REQUIRED`) has decided; the startup posture applies only when no answer arrives. A session started unguarded MUST be recorded in the deployment's own audit log — there is no Guardian to receive the event — and SHOULD be surfaced on Trace as an `acs.decision.failure` event with class `handshake_failure` and posture `fail_open` on the session span when the deployment claims ACS-Trace. The Observed Agent SHOULD retry the handshake for subsequent sessions. Whether a Guardian can attach mid-flight to a session that started unguarded is undefined in v0.1; renegotiation is deferred to v0.2 alongside `system/handshake_renegotiate` (§13).
 
 ## 5. Hook Taxonomy
 
@@ -199,7 +199,7 @@ A hook is a control point only if the Observed Agent waits for the verdict and a
 
 A step suffers a **decision failure** when no usable decision arrives within the negotiated timeout: the Guardian stays silent, the transport fails (connection refused, TLS failure, malformed response), or the Guardian returns an error instead of a decision. All three resolve the same way: the Observed Agent applies the deployment's failure posture, declared as `on_decision_failure` in the handshake (§4) and defaulting to `proceed` (fail-open) so that a slow, erroring, or unreachable Guardian does not halt production. A deployment MAY set `on_decision_failure: deny` (fail-closed). The negotiated timeout bounds every failure mode: an error from the §17.1 registry carries a recovery action the agent MAY attempt within the remaining budget, and an unambiguous failure (a refused connection) MAY resolve immediately rather than waiting out the clock.
 
-Every step that proceeds without a decision MUST be recorded as an audit event, so the bypass is visible rather than silent. When a decision does arrive within the timeout, the agent MUST honor it regardless of the posture. Fail-open trades enforcement for availability under disruption: an adversary who can disrupt the channel converts control into audit. Deployments for which that trade is unacceptable set `on_decision_failure: deny`.
+Every decision failure MUST be recorded as an audit event, under either posture. Under `proceed` the record makes the bypass visible rather than silent, and under `deny` it shows why the step was blocked. The Observed Agent records it in the deployment's own audit log, the rule §4.1 sets for a session started unguarded, and SHOULD also emit it as an `acs.decision.failure` event when the deployment claims ACS-Trace ([Trace Events](../trace/events.md)). At `sessionEnd`, the Observed Agent SHOULD report how many failures it recorded in each failure class, in `summary.decision_failures`. A reader treats a session that ends without that count as unmeasurable. When a decision does arrive within the timeout, the agent MUST honor it regardless of the posture. Fail-open trades enforcement for availability under disruption: an adversary who can disrupt the channel converts control into audit. Deployments for which that trade is unacceptable set `on_decision_failure: deny`.
 
 ### 6.5 MODIFY-incapable clients (normative)