v0.7.0 — Middleware Chat Neutralization

v0.7.0 — Middleware Chat Neutralization

Same tag also includes the already-landed v0.8 runtime ops contract. Crate version stays 0.6.0.

Middleware chat

• Generic Agent chat stays domain-neutral by default.
• EV-repair prompt and FaultCode RAG load only when an Agent mounts the builtin ev-repair-fault-kb pack or its ev-repair alias.
• Ordinary business packs pass through unchanged; same-instance coexistence is covered by tests; missing or invalid claims fail closed.
• #207 / #229, #210 / #230, #208 / #231. (#206 already shipped in v0.6.1.)

Runtime operations on this tag

• Claims-scoped GET /api/v1/tasks, redacted Guard audit/stats, claims-scoped blackboard list/nodes, and the Admin control-plane API matrix.
• #221–#224 / PR #225–#228.

HEAD: 7a4ff19

v0.6.2

v0.6.2 — Multimodal chat wiring (reuse existing)

• Docs (#215): docs/19-multimodal-reuse-existing-capabilities(.zh).md — reuse DESIGN_DETAIL; no parallel multimodal platform.
• Vision fail-closed (#219 / #217/#218): Imaged chat requires vision-capable model_mounts.vision; default 422 vision_mount_unavailable. Opt-in AGENTOS_VISION_FALLBACK=degrade only. Image count/byte limits (AGENTOS_MAX_IMAGES / AGENTOS_MAX_IMAGE_BYTES) → 413. CI chat vision/isolation gate.

HEAD: a8385b4 (or later main tip)

v0.6.1

v0.6.1 — StructCapture minimal cut

• Generic chat neutrality (#206 / #212): Agent HTTP chat and OpenAI-compatible completions no longer inject EV-repair system prompt or FaultCode RAG by default.
• Workload OIDC contract (#211 / #213): Short-lived JWTs mint IsolationClaims; BFFs must not share AGENTOS_JWT_SECRET. See docs/17-isolation-contract.md.

HEAD: 05da010

v0.6.0 — Online Corpus Job + Watcher

Wild AgentOS v0.6.0 — Online Corpus Job + Watcher

[0.6.0] — 2026-09-08

Online Corpus Job + Watcher

• Added authenticated, claims-scoped online corpus jobs for configured corpus
  changes and incremental deltas, with idempotent create, list, get, cancel,
  and runner paths.
• Added enabled-by-default corpus watchers that enqueue work through the same
  idempotent path. Deployments can explicitly disable watcher polling and
  enqueueing without deleting jobs, cursors, reviews, or audit records.
• The runner now requires entity-resolution suggestions and retains them for
  explicit approval; it stages evidence and never auto-merges entities,
  promotes ontology, or materializes production data.
• Added bounded provenance and observability for source versions and content
  digests, canonicalization, quality/review, ER suggestions, queue saturation,
  retries, and job state. Transient sidecar failures retry at most three times;
  validation, authentication, and policy failures are terminal.
• Added fail-closed isolation and production-write CI coverage for online jobs,
  runners, and watchers. Unauthenticated, invalid, cross-scope, and failed
  paths cannot write production data.
• The companion Admin online-corpus job list is delivered separately and does
  not expand this repository's scope.

v0.5.0 — Ontology Knowledge Engineering + Graph Engineering

Wild AgentOS v0.5.0 — Ontology Knowledge Engineering + Graph Engineering

[0.5.0] — 2026-09-07

Ontology Knowledge Engineering + Graph Engineering

• Completed the bounded, claims-scoped ontology Knowledge Engineering pipeline:
  constrained extraction and optional Morph-KGC/RML materialization write only
  to staging; canonicalization uses promoted ontology definitions and retains
  provenance and mapping decisions.
• Added the KgQualityGate supervisory loop with deterministic SPARQL ASK
  anchors, an opt-in pySHACL sidecar, a compliance-first
  quality-versus-coverage policy, and a claims-scoped review queue. Failed
  anchors cannot be overridden by a Judge.
• Added anchored staging-to-production materialization. It requires verified
  claims, explicit confirmation, and a passed gate or recorded human approval;
  the server re-reads production before reporting success and retains staging
  evidence for audit.
• Added conservative entity-resolution suggestions using a process-isolated
  GLinker-style sidecar. Suggestions write owl:sameAs and provenance only to
  staging and require approval; they never auto-merge.
• Froze ontology KE golden evaluations behind a SHA gate and documented the
  measurement-decay audit policy. Added the read-only, claims-scoped ontology
  health report for slow-loop goal review.
• Added golden-SHA, named rule-review, and audit-evidence requirements to
  tenant Skill and emergent-candidate promotion.
• Completed the pre-kernel ontology design path: OpenAPI and SQL DDL
  type-drafts, draft-only schema induction, a read-only readiness report, and
  compatibility-gated promotion with explicit force_breaking audit evidence.
  The companion Admin ontology design studio is delivered separately and does
  not expand this repository's scope.
• Extended the explicit, offline isolation-migrate tool to safely migrate
  historical local vectors, L0 data, and blobs in addition to named graphs;
  it validates targets and records audit evidence while retaining sources by
  default.
• Production deployments now require OIDC/JWKS authentication and refuse to
  boot with HS256 or incomplete OIDC configuration.

Companion

• Admin ontology design studio: Wild_AgentOS_Admin#26
• Admin docs-site sync: Wild_AgentOS_Admin#27

Notes

• Milestone v0.5.0 closed (13/13 issues)
• Still pilot-oriented; no auto-promote of schema; claims-scoped throughout

v0.3.0

Wild AgentOS v0.3.0 — Markets + IdP + Emergent

Kernel

• #97 / PR #127 — Versioned Logic/Skill package market (claims-scoped install/upgrade/rollback)
• #98 / PR #128 — OIDC/JWKS auth mode beside local HS256; fails closed before minting IsolationClaims
• #99 / PR #125 — Gated emergent-tool promotion pipeline (no direct publish)
• #100 / PR #126 — Optional default-off limited RDFS inference (query-time only)

Also in CHANGELOG

• v0.2.2 — claims-scoped coding artifacts, external sandbox adapter, private-deploy benches

Notes

• Docs + crate metadata bumped to 0.3.0
• Still not production-hardened (historical keys, HA/backup, Admin ops) — see Isolation Contract
• No deploy in this release step

v0.2.2

Wild AgentOS v0.2.2 — Artifacts + Sandbox + Bench

Kernel

• #94 / PR #123 — Claims-scoped coding artifact store
• #95 / PR #122 — External SandboxProvider adapter (feature-flagged)
• #96 / PR #124 — Reproducible private-deploy benches (Oxigraph / redb / Hyperspace)

Notes

• Tag points at the last v0.2.2 feature commit on main before v0.3.0 merges.
• Crate metadata on later tip is 0.3.0; see CHANGELOG [0.2.2] on main.

v0.2.1

Wild AgentOS v0.2.1 — Ontology Data + Protocols

Kernel

• #90 / PR #119 — ObjectType/LinkType drafts from CSV / JSON Schema; JWT claims + explicit promote
• #91 / PR #120 — MCP inbound tool catalog filtered by IsolationClaims
• #92 / PR #118 — Published Skills as MCP tool units (claims + gate)
• #93 / PR #117 — Outbound A2A adapter (feature flag default OFF)

Notes

• Docs + crate metadata bumped to 0.2.1
• A2A is outbound-only; inbound platform is follow-up

v0.2.0

Wild AgentOS v0.2.0 — Control Plane + Skill CI

Kernel

• #88 / PR #114 — Skill package format, CI gate (verify + golden; Judge hook default off), gated tenant publish channel
• #89 / PR #115 — Golden eval suite under evals/ + scripts/test_golden.sh in Rust CI

Companion Admin (separate repo)

• Wild_AgentOS_Admin #16 / PR #20 — five-screen control-plane skeleton (Runs / Skills / KB·Ontology / Keys·Models / Isolation)

Notes

• Docs + crate metadata bumped to 0.2.0 in this release commit
• No silent skill promotion; production skill writes remain gated

v0.1.8 — Ontology Action HITL (+ Isolation Proof v0.1.7)

[0.1.8] — 2026-09-04

Ontology Action HITL

• Added Action staging with configurable commit_strategy: actions can commit
  automatically or remain pending explicit approval, with merge, discard, and
  TTL-expiry handling.
• Added configurable ontology guardrails, including SPARQL ASK assertions,
  and a high_risk hook for approval-sensitive actions.
• Published ACTION_AUDIT EventBus events for committed, pending, approved,
  rejected, and violated Action outcomes. See the
  Ontology Action Data Sandbox for the
  current data-sandbox boundary.

[0.1.7] — 2026-09-04

Isolation proof and operations

• Added the read-only isolation-diagnose CLI to distinguish
  claims-minted targets from historical keys.
• Added a customer-readable isolation matrix and fail-closed
  isolation_contract golden test coverage in CI. See the
  Isolation Contract and
  Isolation Matrix.
• Added the optional, explicit isolation-migrate CLI for named-graph
  migration. It never performs a silent UNION; diagnosis remains the default
  operational path.

---

Full notes: CHANGELOG.md. Isolation: docs/17. Action sandbox: docs/15.