AgentSync

AgentSync is a rule-governed skill acquisition loop for AI agents: it checks task rules, finds reusable skills, forces skill authoring when coverage is missing, and promotes validated skills back into the shared skill catalog.

One MCP service, registered once with ContextForge.

Repo status: this is the reference implementation for a private deployment. It is not yet clone-and-run for outside operators — the promotion substrate (stele) resolves from a local path and the MCP server imports it at startup. See Stele substrate and Local deployment before attempting to run it.

Stele substrate

The promotion step commits validated SKILL.md artifacts to a Stele ledger — the real EVECOR Stele LedgerStore, driven in-process as a library dependency, not a vendored copy and not a separate service.

• EvecorStelePort (src/agentsync/kanon/evecor_stele_port.py) implements the StelePromotionPort seam against stele.ledger.store.LedgerStore.
• It bypasses Stele's ledger_transaction(store, SandboxResult) write path — a promoted SKILL.md is a pre-trusted blob with no sandbox run — and calls the store primitives directly: create_pending(...) then commit(...).
• The storage/stele/ledger.db schema (artifact_records) and the artifact files under storage/stele/artifacts/ are owned by Stele, not by AgentSync.
• Stele is an optional dependency ([stele] extra). Tests run without it via FakeStelePromotionPort; the live MCP server requires it (it imports stele.ledger.store at module load and fails loud if absent).

Stele is currently sourced from a local path (pyproject.toml):

stele = ["stele @ file:///mnt/jarvis-data/projects/Stele"]

Outside operators must repoint this at their own Stele checkout, index, or wheel before --extra stele or the MCP server will work.

Setup

git clone https://github.com/Deep-Sixed/AgentSync.git
cd AgentSync

# core + tests run without Stele (promotion is exercised via the fake port)
uv sync --extra mcp --extra dev
uv run pytest -q

# to run the live MCP service you also need a Stele install (see above)
uv sync --extra mcp --extra stele --extra dev

Runtime service

AgentSync runs as a user systemd unit exposing streamable-http MCP on port 8585. ContextForge federates the service; agents call ContextForge only.

Install

mkdir -p ~/.config/agentsync
cp ops/agentsync.env.example ~/.config/agentsync/env
# edit ~/.config/agentsync/env — set paths for your deployment

cp ops/systemd/agentsync-mcp.service ~/.config/systemd/user/
# edit the copied unit — WorkingDirectory / ExecStart paths are deployment-specific
systemctl --user daemon-reload
systemctl --user enable --now agentsync-mcp.service

The shipped ops/systemd/agentsync-mcp.service and ops/agentsync.env.example contain absolute paths for the reference deployment. Edit them for yours.

Logs

journalctl --user -u agentsync-mcp.service -f
journalctl --user -u agentsync-mcp.service --since today

Set AGENTSYNC_LOG_LEVEL=DEBUG in ~/.config/agentsync/env for verbose output.

Restart behavior

• Restart=always with RestartSec=10
• ExecStartPre=scripts/bootstrap-storage.sh ensures storage dirs exist
• main() also creates storage parent dirs before serving

systemctl --user restart agentsync-mcp.service
systemctl --user status agentsync-mcp.service

Mutable storage

Path	Owner	Purpose
storage/obligations/obligations.jsonl	AgentSync	Authoritative obligation lifecycle log
storage/obligations/pre_obligations.jsonl	AgentSync	Non-authoritative rule evaluation records
storage/skills/approved/	AgentSync	Approved SKILL.md registry
storage/stele/ledger.db	Stele	Stele LedgerStore artifact ledger (artifact_records)
storage/stele/artifacts/	Stele	Committed SKILL.md artifact files

All runtime paths overridable via AGENTSYNC_* env vars (see ops/agentsync.env.example). Live *.db, *.jsonl, and artifact files are git-ignored — the tree ships empty dirs (.gitkeep) plus the rules.yaml / capabilities.yaml fixtures only.

Backup and restore

# backup (writes storage/backups/agentsync-YYYYMMDD-HHMMSS.tar.gz)
./scripts/backup-storage.sh

# restore (interactive confirmation; stops service if running)
./scripts/restore-storage.sh storage/backups/agentsync-YYYYMMDD-HHMMSS.tar.gz

Backups include obligations, the Stele ledger, artifacts, and approved skills.

ContextForge registration

Register AgentSync as a ContextForge gateway named agentsync-mcp pointing at http://127.0.0.1:8585/mcp with transport STREAMABLEHTTP.

Agent-facing auth uses a bearer token supplied via the ROUTERCORE_MCP_BEARER_TOKEN environment variable. The token is consumed by the smoke tests and MCP clients — it is not read by the AgentSync server itself. Mint it with whatever bearer tooling fronts your ContextForge / RouterCore deployment, then export it before running the agent-facing tests.

Tests

uv run pytest -q                          # full gate
uv run pytest tests/test_operational_hardening.py -v
uv run pytest tests/test_agent_facing_a2a_smoke.py -v  # requires ContextForge

The test_agent_facing_a2a_smoke.py and test_contextforge_smoke.py suites require a running ContextForge and a valid ROUTERCORE_MCP_BEARER_TOKEN; the rest run standalone.

Frozen checkpoints

Current validated gate: v1.1-kanon-operational-hardening — 98/98 (HEAD is this tag plus docs-only commits).

Tag	Gate	Note
v1.1-kanon-operational-hardening	98/98	current
v1.1-kanon-agent-facing-a2a-smoke	95/95	historical

Local deployment (Deep-Sixed)

Concrete paths for the reference deployment on JARVIS:

# project root
cd /mnt/jarvis-data/projects/AgentSync

# Stele source (pyproject [stele] extra)
#   stele @ file:///mnt/jarvis-data/projects/Stele

# mint the ContextForge agent bearer
ROUTERCORE_MCP_BEARER_TOKEN="$(/mnt/jarvis-data/projects/EVECOR/bin/evecor-routercore-mcp-bearer)"