MOVED — canonical home is now ashfulcra/fulcra-tools

coord2 relocated 2026-07-04 (fulcra-tools PR #287, full history preserved): engine → packages/coord-engine (releases tagged coord-engine-vX.Y.Z there; v1.3.0 was the last tag here), skills → skills/, docs → docs/coord2*. This repo is retained as provenance for the pre-move review lineage.

coord2

Agent coordination as optional layers on top of fulcra-agent-teams.

coord2 is a ground-up rebuild of fulcra-coord + fulcra-continuity as a family of sibling fulcra-agent-* skills that enhance Fulcra's official (alpha) fulcra-agent-teams skill — each a pure-prose SKILL backed by one shared tool, coord-engine, for the deterministic parts (the consistency-critical folds that prose would drift on). Not a parallel system; not pip packages to version-sync across a fleet.

Bare fulcra-agent-teams is a lightweight OKF-markdown convention over the Fulcra File Store (per-agent inboxes, progress.md continuity, consent-gated automation). coord2 keeps that as the base tier and adds, à la carte, the machinery that a convention structurally lacks: typed task lifecycle, self-healing queryable views, roles + leases, a review handshake, a VCS bridge, structured resumable continuity, and OS-level wake automation.

Why

The incumbent fulcra-tools-coord (v0.15.x) is a powerful but heavyweight parallel system — a dedicated CLI + JSON store + reconcile engine that must be kept version-synced across a fleet. coord2's thesis: make the official skill the substrate and coord the opt-in "pro tier." Bare-teams agents interoperate by reading the same markdown; power agents opt into the layers they want. OKF v0.1 explicitly permits the typed frontmatter and synthesized views this needs.

Architecture

Each capability is a skill (agent-facing interface + genuinely-conventional prose); the consistency-critical folds live in the shared coord-engine tool the skills invoke via uv tool run coord-engine … (the same way skills already invoke fulcra-api). Decision + rationale: docs/proposals/teams-convergence/03-packaging-decision.md.

Layer	Skill	Engine command(s)	Adds
L0	fulcra-agent-teams (external, base)	—	OKF-markdown team spaces, inboxes, continuity files
L1 ✅	fulcra-agent-reconcile	reconcile/status/board/needs-me/search	self-healing task/index.md+log.md, queryable views
L4 ✅	fulcra-agent-roles	roles status	roles + leases + HELD/VACANT/CONTESTED fold + SLA escalation-due
L2 ✅	fulcra-agent-tasks	task start/update/done	typed status/priority/assignee lifecycle + validated state machine
L5 ✅	fulcra-agent-review	review status	review handshake + APPROVED/CHANGES/PENDING verdict tally
L6 ✅	fulcra-agent-continuity	continuity snapshot/resume	structured resumable snapshots + deterministic resume brief
L7 ✅	fulcra-agent-automation	(schedules reconcile)	heartbeat scheduler + resume-on-wake loop

Reconcile is the linchpin (queryability + self-healing). The roles fold and reconcile share the engine's OKF parser + transport — one implementation, no drift.

Install (standalone)

coord2 installs from a checkout — the engine and skills come from the same checkout, so they're always version-matched. Matches how fulcra-agent-teams is installed (copy the skill folders into your agent's skills dir).

git clone https://github.com/ashfulcra/coord2 && cd coord2
scripts/coord2-setup.sh              # installs coord-engine (from ./engine) + copies the 6 skills into ~/.claude/skills/
# options: --symlink (dev; git pull updates in place) · --skills-dir DIR (other agents) · --engine-only/--skills-only · --uninstall

Prereqs: uv, and an authenticated fulcra-api (uv tool install fulcra-api && fulcra-api auth login) — the setup self-test checks for both. Engine-only, no checkout: uv tool install "git+https://github.com/ashfulcra/coord2.git@v0.4.0#subdirectory=engine".

Publishing to PyPI / a plugin registry is deferred to the upstream phase (see docs/proposals/teams-convergence/04-standalone-packaging.md).

Full port (A1–A8) — complete

All remaining fulcra-coord + fulcra-continuity functionality is ported as optional add-ons (engine v1.0.0, 172 tests, every add-on independently reviewed and live-verified on the real store):

Add-on	Verbs	PRs
A1 tasks-completion	task block/pause/abandon/assign/restore	#10
A2 directives	tell broadcast remind later handoff inbox respond (+acks, shard-GC)	#13,#14
A3 presence	presence beat/show, agents, roles claim/release	#11,#12
A4 retention	reconcile --retention-days, search --archived, verified-move archive	#15,#16
A5 health	doctor health digest escalate (+per-host shards)	#17,#18
A6 continuity	continuity checkpoint/park, briefing	#20
A7 forge	forge mirror (PR evidence + auto-verdict, --repo allowlist)	#22
A8 listener	install-listener.sh + listener-tick.sh (notify + consent-gated wake)	#25

Status

**coord-engine v0.4.0 + all six skills built — reconcile, roles, tasks, review, continuity, automation. 105 engine tests + a live end-to-end run of reconcile/queries against the real Fulcra File Store. Packaging decided (approach C) after independent + bus (Codex) review — both reviewers independently flagged the roles-fold-as-prose defect, now fixed; open decisions resolved in doc 03. The full stack is in place.

Foundations validated (2026-07-01): fulcra-api file is last-writer-wins + versions every upload (stat → version UUID + history; restore rolls back live files), list timestamps are minute-granular (incremental uses a conservative compare), file output is text (line-parsed), delete isn't CLI-undoable (archival = move-not-delete). The upstream skills model uses no pip/wheels — skills invoke shared tools as external commands (uv tool run fulcra-api), which is why coord-engine ships the same way.

Layout

• engine/ — coord-engine, the shared stdlib-only tool (OKF parse/render, transport, reconcile, role fold, queries). Tested. cd engine && uv run --extra dev pytest.
• skills/fulcra-agent-*/ — the pure-prose skills (SKILL.md + references/) that invoke coord-engine.
• docs/proposals/teams-convergence/ — the design set (start at its README.md).

Conventions

Python, stdlib-only in coord-engine (transport is a CLI-subprocess call to fulcra-api, so no heavy deps). Best-effort / never-raise on the reconcile path. Skills are prose + references, no bundled code. TDD; every stateful fold is engine-side + tested, never prose the agent eyeballs.

License

MIT © 2026 Fulcra Dynamics