Changelog

Version history for agentic-orchestration. Source: CHANGELOG.md at repo root.

All notable changes to agentic-orchestration are documented here.

Format based on Keep a Changelog.
Versioning follows Semantic Versioning (VERSION at repo root).

[Unreleased]

User harness Phase 4 — subprocess/Kubernetes backend parity in shared run_harness_kickoff; scenario inputs.matrix variant expansion; planner context via user_harness_performance_summary (AGENTIC_USER_HARNESS_FEED_PLANNER).

User agent harnesses — domain scenario packs per catalog agent_provider_id via orchestration/user_agent_harness.py.
CLI: --harness-dir, --user-harness-run-all (with shared --harness-agent, --harness-json, --harness-fail-fast).
Env: AGENTIC_EXTRA_AGENT_HARNESS_DIRS, AGENTIC_USER_HARNESS_RECORD_STATS.
Healthcare reference pack: examples/verticals/healthcare/harnesses/gpt_research/ (three scenarios).
Vertical --example overlay merges harnesses/ into harness discovery.
Scripts: scripts/run-user-harness.ps1, scripts/run-user-harness.sh.
@pytest.mark.user_harness tests; included in harness CI job alongside platform harness tests.
User harness stats in learning_store (user_harness_stats).
Public run_assertions() and forbids_regex assertion type shared with platform harness.

Platform agent harness — tiered per-catalog-agent verification (static, connectivity, smoke, capability) via orchestration/agent_harness.py.
Harness profile templates under config/agent_harnesses/ (general, research, write, reason, coding, vision).
CLI: --harness-agent, --harness-batch, --harness-tier, --harness-filter, --harness-json, and related flags.
Scripts: scripts/run-agent-harness.ps1, scripts/run-agent-harness.sh, scripts/harness-report.py.
CI jobs: agent-harness-static (L0 full catalog), agent-harness-connectivity (L1 + unit tests); nightly workflow agent-harness-smoke-nightly.yml (L2).
harness_profile / harness.skip_live on reference cloud agents (gpt_*, claude_*, ollama_llama3).
Harness stats in learning_store and planner context via harness_performance_summary.
Optional AgentProvider.run_harness_probe() hook for custom provider classes.
Subprocess backend support for smoke/capability tiers (--harness-backend subprocess).

Agent skills catalog (config/agent_skills/, --agent-skills-catalog) — YAML procedural playbooks injected into task descriptions or agent backstory; composes with MCP attachments.
Shipped skills: echo_skill, echo_backstory_skill, release_process, pr_review; smoke workflow workflow_agent_skills_smoke.yaml.
Dynamic planner skill_ids (workflow default + per-step), keyword goal-match augmentation, and relevance pruning (workflow and per-task).
Worker skills re-resolve — --execute-step and distributed backends reload skills from StepSpec.skills + paths.agent_skills_catalog.
Kind CI skills e2e — stub worker verifies skills spec handoff (k8s_stub_skills, test_agent_skills_smoke_kind_kubernetes_workflow).
Catalog features: content.summary, required_files gating, SKILL.md frontmatter strip, bundle YAML, AGENTIC_EXTRA_AGENT_SKILLS_PATH.
Combined attachment fingerprint (MCP + skills) for learning stats, KB writes, web UI ratings, and planner traces (attachment_fingerprint with legacy mcp_fingerprint alias).
Web UI parses (agentic) run_rating_meta stderr for thumbs up/down envelope.

learning_store / knowledge_base APIs accept attachment_fingerprint; KB dynamic runs store per-task fingerprint on the final task.
Stub worker Docker image embeds config/agent_skills for skills verification in kind e2e.

Web ratings now record combined MCP+skill attachment digest instead of always none.

Kubernetes execution (AGENTIC_EXECUTION_BACKEND=kubernetes) — coordinator dispatches each workflow step to workers via shared run-store PVC; one-shot Jobs or warm pool.
In-cluster coordinator (K3.7) — Deployment + RBAC; serves web UI and creates worker workloads.
Warm pool (K5.1) — long-running worker pods dequeue steps from the PVC (AGENTIC_K8S_WARM_POOL_ENABLED=1).
Delegation RPC (K5.5) — k8s_delegate_task CrewAI tool + delegation broker for child Jobs.
Structured K8s logging — AGENTIC_LOG_FORMAT=json with run_id / step_id / component (see deploy/k8s/LOGGING.md).
MCP on K8s — sidecar/gateway manifests (fetch, filesystem); pod-sidecar compatibility layer.
Docker images — Dockerfile.coordinator, Dockerfile.worker; kind full-stack and Jetson k3s deploy scripts.
Web UI prose mode — AGENTIC_WEB_PROSE_DELIVERABLE for chat runs; client unwraps JSON-shaped stdout.
Step recovery, kind/kubernetes CI e2e, load-test scripts, GitHub Pages product site + docs hub.

Dual execution framework completed through F3: StepCoordinator, run store, --execute-step worker entrypoint, subprocess and kubernetes runners.
Dynamic planner and synthesis steer toward readable prose when the web UI sets prose delivery.

Warm pool run-store paths on Windows host dispatch.
Coordinator image: correct public/ path for web UI static assets.
Jetson deploy: Traefik vs hostPort 80 conflict; NodePort web exposure and rollout reliability.

Mocked @pytest.mark.backend_inprocess regression tests for F1.4 (static, dynamic, and CrewAIExecutionBackend kickoff path) in default CI.

Dual execution framework (F0–F3): pluggable execution backends with AGENTIC_EXECUTION_BACKEND (inprocess default, subprocess, kubernetes stub).
orchestration/backends/ — CrewAIExecutionBackend, factory, step contracts (StepSpec, StepResult).
workflow_materializer, step_coordinator, run_store, and --execute-step worker entrypoint for distributed step execution.
Subprocess backend (AGENTIC_SUBPROCESS_WORKERS=1) spawns per-step workers via python main.py --execute-step.
Unit tests for step context, workflow materializer, run store, and execution backend factory.
Opt-in Live LLM GitHub Actions workflow (.github/workflows/live-llm.yml) and tests/test_live_llm_smoke.py.

main.py delegates kickoff to CrewAIExecutionBackend via factory; lifecycle hooks moved to orchestration/backends/crewai.py.
.env.example documents AGENTIC_EXECUTION_BACKEND and AGENTIC_SUBPROCESS_WORKERS.

Unit test suite (agentic-orchestration-tool/tests/) and pytest configuration.
GitHub Actions CI (.github/workflows/ci.yml) and GitLab CI (.gitlab-ci.yml).
Release process: VERSION, CHANGELOG.md, RELEASING.md, tag-triggered GitHub Release workflow.
Wiki roadmaps: dual execution framework and Kubernetes execution upgrade.

Test runner scripts (run-tests.ps1 / run-tests.sh) use correct tool root and unit-only deps.