Ecosystem mode

Coordinate work across multiple related projects from a single agent session. Opt-in, additive, and built so single-project usage stays identical.

Ecosystem mode is the state layer for multi-project work — the durable record that Orchestration primitives (the action layer) read and write. Together they form the canonical multi-project pattern. Ecosystem mode without orchestration is just bookkeeping; orchestration without ecosystem mode is action with no record. You’ll typically reach for both.

Single-project momentum is the default. When you have multiple related projects that need to coordinate — a platform + SDK + CLI, or a frontend + backend + infra — ecosystem mode lets one agent session work across all of them while preserving every project’s per-project tracking discipline.

Hard invariant: a project running momentum init without --ecosystem sees zero difference from before ecosystem mode existed. Ecosystem mode is purely additive.

What ecosystem mode adds

When you opt in, an ecosystem root appears as a sibling git repo containing shared state. Each member project keeps its own specs/ exactly as in single-project mode, plus a small pointer block in its primary instruction file that tells the agent it’s a member.

parent-dir/
├── my-eco/                    ← ecosystem root (sibling git repo)
│   ├── ecosystem.json         ← members, roles, dependency edges
│   ├── initiatives/           ← cross-project features
│   │   └── 0001-memory.md
│   ├── sessions/              ← daily activity log (auto-appended)
│   │   └── 2026-06-08.md
│   ├── .state/                ← state-machine artifacts
│   └── README.md
├── sapience/                  ← member: own specs/ + pointer block
├── sdk/                       ← member: own specs/ + pointer block
└── cli/                       ← member: own specs/ + pointer block

Each member’s CLAUDE.md / AGENTS.md gets a fenced pointer block injected by momentum join:

<!-- momentum:ecosystem-pointer:start -->
> Member of `my-eco` ecosystem at `../my-eco/`.
> Cross-repo work → `../my-eco/initiatives/`. Don't create
> coordination files here.
> See `ecosystem.json` for siblings and `momentum ecosystem
> status` for live state.
<!-- momentum:ecosystem-pointer:end -->

The pointer is the agent’s discovery surface — “you’re in a coordination layer; cross-project coordination lives next door.”

When to use ecosystem mode

Reach for it when:

You have two or more repos that share a single product or platform.
You frequently change two or more repos for one feature.
You want one agent session to scan all of them, draft initiatives that span them, and keep a unified daily log.

Skip it when:

You only have one repo. (Single-project mode is the default for a reason.)
Your “related” repos are loosely coupled and rarely change together.
You’re not yet on a sustained build cadence — ecosystem mode pays off most when there’s enough cross-project activity for the shared log + initiatives to be useful.

Quickstart

Create from your first project

# From inside the first project you want to anchor:
npx @limina-labs/momentum@latest init --ecosystem my-eco

This creates ../my-eco/ as a sibling git repo, scaffolds the manifest + initiatives + sessions dirs, and registers the current project as the first member.

The ecosystem root also gets its own managed CLAUDE.md + AGENTS.md (written by init) telling agents at session start that they’re in a coordination layer, not a project — so they reach for the orchestration primitives instead of accidentally hand-authoring planning files in the wrong layer.

Add subsequent members

# From inside each additional project:
npx @limina-labs/momentum@latest join ../my-eco

join is idempotent — running it twice does the right thing. It registers the project in ecosystem.json and injects the pointer block into the instruction file (skipped if already present).

Diagnose state

momentum doctor

Reports whether the current directory is a standalone project, an ecosystem root, or an ecosystem member, plus detected issues with the state files. Walk-and-check covers parent dirs AND sibling dirs (the ecosystem root is typically a sibling of members, not an ancestor).

Upgrade every member at once

# From the ecosystem root (or any member):
npm install -g @limina-labs/momentum@latest   # update the CLI first
momentum ecosystem upgrade

Sweeps momentum upgrade across every member in one pass — no more cd-ing into each repo. It’s a pull sweep over your local checkouts (no bot, no forge integration), so it stays vendor-neutral. Per member it:

skips a dirty working tree by default, so an upgrade never lands on top of unrelated uncommitted work;
detects the adapter from that repo’s .momentum/installed.json lock file (override with --agent <name>);
runs the upgrade and reports the momentum version before → after;
tolerates partial failure — one bad repo never aborts the fleet.

Got members with uncommitted work in flight? Don’t clean them by hand — --autostash does it for you:

momentum ecosystem upgrade --autostash   # stash → upgrade → restore, per repo
momentum ecosystem upgrade --dry-run     # or preview the whole sweep first

--autostash stashes each dirty repo’s changes, upgrades the now-clean tree, then restores your work exactly as it was. Because momentum upgrade only writes momentum-owned paths (.claude/, .agent/, scripts/, the managed top of your instruction file), your application changes and the upgrade don’t collide — and on the rare occasion they would, your work is kept safe in git stash rather than overwritten. --force is the blunt alternative (upgrade in place, no stashing).

Leave an ecosystem

# From inside the member project:
npx @limina-labs/momentum@latest leave

Removes the pointer block and de-registers the project from ecosystem.json. Per-project work continues unchanged.

Initiatives — cross-project features

When a feature genuinely spans multiple repos, create an initiative file at <eco-root>/initiatives/NNNN-<slug>.md. The initiative captures cross-project context: which members are involved, what each member’s role is, dependency order, links to per-member phases or PRs.

Create one via CLI (non-interactive — works from any agent context):

momentum ecosystem initiative create memory-v1 \
  --why "ship persistent memory across backend + SDK + CLI" \
  --repos sapience,sdk,cli \
  --owner limina

Or via the slash door (same code path, agent-friendly):

/initiative create memory-v1

The created file scaffolds frontmatter (id, slug, status, started, owner, repos) plus the fixed sections below. The new initiative is also set as the active initiative so the SessionStart hook surfaces it on every new agent session in the ecosystem.

# Initiative 0001 — Memory module v1

## Summary
Ship persistent memory for the platform — sapience exposes the API,
SDK consumes it, CLI documents it.

## Members involved
- sapience/ — backend implementation
- sdk/ — client SDK update
- cli/ — docs + examples

## Dependency order
1. sapience/ (Phase 5 of that repo)
2. sdk/ (consumer phase — depends on sapience@v1.0)
3. cli/ (docs phase — depends on sdk@v0.4)

## References
- sapience: specs/phases/phase-5-memory/
- ADR: ../sapience/specs/decisions/0042-memory-api.md

Initiatives are written by the user or by an agent invoking /initiative — both work the same way. They live in the ecosystem repo, not in any member. Per-member phases reference the initiative; the initiative references the phases.

Sessions — auto-recorded activity log

Every git commit and gh pr create / merge in any member repo auto-appends a line to today’s session file at <eco-root>/sessions/YYYY-MM-DD.md. The auto-append runs from a PostToolUse hook on Claude Code (other adapters via equivalent hooks).

# Session — 2026-06-08

## sapience
- 09:42 commit  feat(memory): kv store + JSON schema (abc1234)
- 11:15 pr      merge #128 — memory write-path
## sdk
- 12:03 commit  feat(memory): typed client + tests (def5678)
## cli
- 14:30 commit  docs(memory): worked example for /init (789abcd)

One file per day. The log is best-effort and never blocks user work — a session-append failure logs to stderr but doesn’t error out the commit.

Concurrent-commit safety

Two members committing simultaneously could race on the same session file. The auto-append script uses a per-session-file mkdir-based lock (portable across macOS / Linux without flock). Verified via 10-process stress test in Phase 10.

What ecosystem mode doesn’t change

Each member’s specs/ is untouched — phases, backlog, history all live in the member repo exactly as in single-project mode.
Every per-project slash command (/start-phase, /complete-phase, /track, etc.) works the same way.
The npm package surface is unchanged — @limina-labs/momentum ships one CLI that handles both modes.
The 13 autonomous agent rules apply identically; Rule 9 (multi-project doc sync) adds the “don’t modify docs you don’t own” guard for cross-project agent sessions.

A worked example

Imagine you maintain three related repos:

~/work/
├── platform/      ← Node.js backend, monorepo, owns the API
├── sdk/           ← Python client SDK that consumes the API
└── cli/           ← Go CLI that ships with the platform docs

You’re already running momentum in each individually. Now you want to ship “memory module v1” — a feature that touches all three.

Step 1 — Create the ecosystem

cd ~/work/platform
npx @limina-labs/momentum@latest init --ecosystem my-stack

../my-stack/ appears as a new sibling. The platform repo gains a pointer block. momentum ecosystem status from inside platform now reports a 1-member ecosystem.

Step 2 — Join the other members

cd ~/work/sdk
npx @limina-labs/momentum@latest join ../my-stack

cd ~/work/cli
npx @limina-labs/momentum@latest join ../my-stack

Now ecosystem status reports 3 members. The shared ecosystem.json lists all three.

Step 3 — Write an initiative

From any member, run /initiative (or hand-write the file at ../my-stack/initiatives/0001-memory.md). The initiative captures cross-project context, dependency order, and link-outs to per-member phases.

Step 4 — Work normally in each member

In platform/, you brainstorm and start a phase: /brainstorm-phase → /start-phase. The phase plan references the initiative. Every commit auto-records to today’s session log. The ecosystem layer adds no friction to the per-project workflow.

Step 5 — Cross-repo questions

When the agent needs to know “where is the SDK on consuming the memory API?” — instead of opening a session in sdk/, the agent can scout sdk or dispatch platform sdk cli for parallel state checks. See the Orchestration page for the deep dive.