/start — Session Resume Skill

Invocation: /start (also called automatically from agent activation files) Works for: any agent (saga, freya, mimir)

Purpose

Loads project state and session context. Always reads the project index first — this gives the agent a complete picture of what exists before doing anything else.

Behavior When Invoked

1. Load Project Index

Always read progress/project-index.md first, regardless of whether a session state exists.

If found: parse Phase Status and Artifacts sections. Hold this as project context — it informs everything below. If not found: proceed silently. The index will be built on first wrap.

2. Detect Session State

Read ~/.claude/wds/tools/memory/SKILL.md and follow the load operation for the current agent_id.

Fallback chain: state found → show resume prompt → fresh start

3. If State Found

Parse the state file for:

Context section
Next section — extract MODEL prefix if present
Plan / Milestones section

Display:

⏸ Previous session found ([date from Wrapped field])

Project:  [N artifacts — current phase from project index, or "no index yet"]
Left off: [content from Context section]
Next:     [Next — strip MODEL prefix, show as plain task]
Model:    [Sonnet | Opus — from MODEL prefix, or inferred]

[If milestones present:]
── Session Plan ──────────────────────────────
[DONE] Milestone 1 — description
[CURRENT] Milestone 2 — description (~N sessions)
[ ] Milestone 3 — description (~N sessions)
──────────────────────────────────────────────

Resume where we left off, or start fresh?

Wait for the user's response.

Model inference (if no MODEL prefix in Next):

Any code, build, deploy, implement → Opus
High-stakes work (production, financial, compliance) → Opus
Long or complex multi-step tasks → Opus
Moderate complexity: strategy, spec, dialog, UX, config, analysis → Sonnet
Simple, low-stakes, short → Haiku
Default to lightest model that fits.

If resume:

Read the full state file
Jump straight to the Next Action — no scanning, no re-introduction
Treat context as already established

If fresh:

Proceed with the normal activation sequence
Do not delete the state file

4. If No State Found

Proceed with the normal activation sequence.

If the user describes a multi-session task at the start of a fresh session, offer to map milestones:

This looks like multi-session work. Want me to map it into milestones first?
(Adds ~2 min upfront, saves context thrashing later.)

If yes: produce milestone plan before starting work. If no: proceed directly.

If work appears single-session: proceed directly without asking.

Do not mention /start or the absence of a state file.

Notes

Always read progress/project-index.md — never skip it. It is the project's memory.
The state file lives at progress/[agent].md relative to the project root.
On resume, get back to work quickly. The user knows the context.

start.md 3.1 KB Histórico Raw