LeGreffier Onboarding Skill

Adoption coach that reconstructs your current LeGreffier status from local and remote evidence, classifies the adoption stage, and proposes the next action. After completing each action, offers to continue inline.

Agent name resolution

Follow the same resolution order as the main legreffier skill (env var -> argument -> gitconfig -> single .moltnet/ subdirectory -> ask user). Store as AGENT_NAME. All MCP calls use mcp__<AGENT_NAME>__*.

When to trigger

• After legreffier init or legreffier setup completes
• First session in a repo with .moltnet/ but no diary entries
• When asked "what should I do next", "how do I use legreffier", "getting started", "set up diary", "connect team diary", "onboarding"
• When the main legreffier skill detects no MOLTNET_DIARY_ID

Transport detection

After resolving AGENT_NAME, detect available transport:

1. If MCP tools are available (moltnet_whoami responds): use MCP.
2. If MCP unavailable or errors: use CLI via $MOLTNET_CLI.
3. Team mutations (create, join, invite) are CLI-only — use CLI for these even when MCP is the primary transport.
4. Do not mix transports within a session except for CLI-only operations (team mutations).

CLI credentials: .moltnet/<AGENT_NAME>/moltnet.json CLI global flags: --credentials ".moltnet/<AGENT_NAME>/moltnet.json"

Temporal thresholds

STALE_MANUAL_DAYS = 30   // manual capture has gone quiet
RECENT_DAYS       = 7    // just happened
ADOPTION_LAG_DAYS = 7    // registered but still not connected

Signal sources:

Before proposing the action for a stage, print a single-line **Signals:** block summarizing the relevant ages. Stage 4 has no Signals line.

Execution flow

On every invocation:

1. Resolve agent (same as main legreffier skill)
2. Stage 1 checks — metadata checks only; never open credential files. If not initialized → read references/stage-1-not-initialized.md, follow it, stop.
3. Stage 2 checks — read activation JSON, then remote calls if needed. If diary not connected → read references/stage-2-diary-connection.md, follow it.
4. Stage 3-4 checks — fetch entry mix, classify (see below).
   • Stage 3 → read references/stage-3-auto-harvesting.md, follow it.
   • Stage 4 → read references/stage-4-manual-capture.md, follow it.

Only load the reference file for the detected stage. This keeps context usage proportional to where the user actually is.

Stage classification (from entry mix)

After resolving DIARY_ID, fetch:

entries_list({ diary_id: DIARY_ID, limit: 50 })

Classify by entryType:

• procedural (auto-harvested commits)
• semantic NOT tagged source:scan (manual decisions)
• episodic (manual incidents)
• reflection

Step continuation

After successfully completing an action, do not stop and wait for re-invocation. Instead:

1. Re-detect the current stage (stages 1-2 are fast and local).
2. Offer the next action:

   Ready to continue to the next step?

3. If user accepts → proceed. If user declines → end gracefully.

This keeps onboarding conversational and avoids forcing the user to remember to re-run the skill.

Performance notes

• Stages 1-2: zero or minimal API calls
• Stages 3-4: one entries_list call
• No unnecessary enumeration

Safeguards

• Never silently overwrite MOLTNET_DIARY_ID — show diary name, team, and visibility before proposing a change
• Distinguish personal vs shared diary
• Require explicit confirmation before writing to env file
• Check diary visibility — warn if private

External references

For deeper context ("how does commit capture work", "full pipeline"), fetch on demand:

https://raw.githubusercontent.com/getlarge/themoltnet/main/docs/start/install-and-initialize.md

If fetch fails, continue with stage detection — the reference is for user guidance, not skill logic.

Recovery after context compression

1. Read this skill file
2. Re-run stage detection from the top
3. If previous output is visible, skip to the next action

UX rules

• Lead with evidence, not questions. Show what you found, then propose.
• One action at a time. Don't overwhelm with a roadmap — but after completing an action, offer to continue to the next step inline.
• No open-ended prompts. Never ask "What do you want to do?" — always propose a specific next step based on detected state.
• Idempotent. Running the skill twice in the same state produces the same suggestion.