AgentAction 0.13.1

Mixed-history eval routing

• Allow refund_triage.v2 to route Hermes source-and-agent targets that also
  contain older lifecycle-only Jobs, without deleting or reclassifying that
  history.
• Keep declared evidence self-attested, while Jobs without semantic claims
  freeze the selected evaluation and report insufficient evidence instead of
  blocking the assignment or inventing a pass.
• Preserve fail-closed assignment checks for genuinely incompatible evaluator
  and traffic combinations, with matching console preview guidance.
• Present the deterministic criterion aggregate as the canonical eval verdict,
  separately from the agent-reported lifecycle outcome and the legacy intent-
  profile diagnostic, so a passing six-criterion eval is never mislabeled as
  failed by the older profile score.
• Add explicit eval_verdict and legacy_profile_result read-model fields
  while preserving existing receipt fields and rejecting mismatched criterion
  provenance instead of silently promoting it.

AgentAction 0.13.0

Explainable workspace evaluations

• Add an immutable refund-triage evaluation specification with explicit,
  criterion-level pass, fail, and insufficient-evidence results plus bounded
  explanations, evidence references, aggregate thresholds, and frozen
  evaluator provenance.
• Let the maintained Hermes integration report bounded evidence for the six
  refund-triage criteria, bound to the active Job and labeled Self-attested
  by agent · not independently verified throughout storage and the console.
• Reconcile provider usage by distinguishing uncached input, cached input,
  output, and total tokens without inventing missing counters or double
  counting provider-reported cache usage.
• Make eval assignment safer with workspace-scoped source and agent choices,
  effective-route previews, and compatibility or coverage warnings before an
  owner saves a route.
• Keep Job detail reachable from Job rows and supported deep links with a clear
  workspace-preserving return path, while removing it from permanent
  navigation.
• Preserve v1 definitions and historical finalized Jobs through additive
  fields; existing workspaces require no migration. This release does not add
  a general-purpose rubric builder or independent model judge.

AgentAction 0.12.0

Workspace eval routing

• Add tenant-scoped, immutable eval definitions and owner-managed assignments
  with exact source-and-agent, agent, source, and workspace-default precedence.
• Freeze the resolved eval version, profile digest, and assignment on Job start
  so route changes affect only later Jobs and unlike versions remain separate.
• Add an Evals workspace view plus eval identity and route provenance in
  Jobs and Job detail, while keeping sources independent from evaluation setup.
• Make the trust boundary explicit: v1 versions the existing deterministic
  observed-lifecycle or agent self-attestation evaluators and does not claim an
  arbitrary rubric, LLM judge, or independent outcome evidence.
• Use observability-console.agentaction.dev as the canonical protected
  operator-console URL while retaining the underlying Worker deployment.

AgentAction 0.11.0

Model and token usage for Hermes observability

• Show provider/model names and clearly labeled estimated or actual token
  counts on model-request Activity events.
• Aggregate provider-reported input, output, and total tokens into immutable
  Hermes Jobs, with request coverage and a bounded per-model breakdown in Job
  detail.
• Preserve the existing privacy boundary and fail-open behavior: no prompts or
  responses are captured, missing usage remains unavailable, and existing
  integrations and finalized Jobs require no migration.

AgentAction 0.10.1

Reliable Hermes activity and support-triage demo

• Give every emitted Hermes observation a unique, retry-stable event ID and
  discard legacy conflicting duplicates so one poisoned record cannot block
  the Activity retry spool.
• Add a synthetic, read-only refund-triage workflow with eligible and
  manual-review cases, bounded policy rules, expected outcomes, and fixture
  validation for meaningful intent and eval demonstrations.
• Preserve the existing Hermes observation schemas, gateway contract, privacy
  boundary, and configuration; existing installations require no migration and
  can update the AgentAction Hermes plugin in place.

AgentAction 0.10.0

Agent-declared intent for Hermes

• Add opt-in Hermes context injection and structured declaration/outcome tools
  without exporting prompts, histories, tool arguments/results, or responses.
• Evaluate bounded declarations under the server-owned
  agentaction_declared_intent.v1 profile while preserving observed-execution
  fallback and fail-open agent behavior.
• Label declarations and outcome reports as agent-generated self-attestation in
  Jobs and Job detail rather than trusted user intent or independent evidence.

AgentAction 0.9.10

Pending workspace invitations

• Show owners pending and expired workspace invitations alongside active Team
  members, including the invited role and expiration state.
• Keep invitation listings tenant-scoped and owner-only without returning
  one-time codes or secret digests.
• Backfill the tenant invitation index on first read so existing invitations
  appear without being sent again.

AgentAction 0.9.9

Workspace-stable Job detail links

• Preserve an authorized workspace across in-app Job navigation, reloads, and
  shared Job-detail URLs without treating the client preference as authority.
• Keep server-side membership authorization as the source of truth and fall
  back safely when a requested workspace is not one of the signed-in user's
  memberships.

AgentAction 0.9.8

Hermes runs in Jobs

• Finalize completed Hermes turns/runs as immutable, source-scoped Jobs under a
  server-owned observed-execution lifecycle profile.
• Keep source credentials write-only and bound to their registered tenant,
  source, and agent while making lifecycle retries deterministic.
• Label observed-execution Jobs separately from explicit semantic intent in the
  operator console and continue excluding prompts, arguments, and results.

AgentAction 0.9.7

Hermes plugin installation compatibility

• Make the AgentAction native plugin installable by the current official
  Hermes Agent release while preserving its v1 runtime hook contract.
• Keep the existing fail-open, privacy-safe shadow observability behavior and
  require no migration for existing deployments.