Go Start

Go-Start — Session Start

Loads platform and project context in a single run. One command, agent is ready.

Three phases:

1. Platform — read common platform docs
2. Project — read docs of the selected project
3. Session type — new task or resuming a stuck session

Input

The session-start signal (user invoking /go-start in a fresh terminal).

Output

Loaded platform + project context in-session; numbered project menu shown; session type resolved to new or continuation.

Hands off to

go-fast (new session), or resumes the interrupted task's per-block cycle (continuation).

---

Phase 1 — Platform

Step 1.1 — Read platform files

Read all five sequentially:

• {tasks_root}/README.md
• {tasks_root}/docs/core/WORKFLOW.md
• {tasks_root}/docs/core/DEPLOY.md
• {tasks_root}/docs/core/PLATFORM_KNOWLEDGE.md
• {tasks_root}/projects.yml

Rule: read all five — do not skip even if you think you already know the content. Each session starts from a clean slate.

Step 1.1.5 — Cleanup stale git locks

Before the summary — silently run a helper that scans active projects and clears stale .git/index.lock files (>5 min old). Idempotent. Capture stdout for use in the summary line.

Why: background git commands can hang for hours, hold .git/index.lock and block the repo. The helper kills the offending process and removes the file.

Step 1.2 — Output platform summary

A compact block (no more than 8 lines):

✅ Platform loaded.

Servers: {msk_vds} | {nl_vds}
Deploy: git push origin main → post-receive → auto-backup
Skills: {skills_project}/.claude/skills/ → sync-agents.sh → all projects
Artifacts: task_artifacts in routing.db
Git locks: {helper result}

Do not retell files in full — only operationally important information.

Step 1.3 — Project menu

Take only status: active from projects.yml. Output a numbered list with id / name / domain. Wait for explicit user selection — do not guess, do not proceed to Phase 2 without it.

---

Phase 2 — Project

Step 2.1 — Determine project path

Take path from projects.yml for the selected id. Partial name match on name or id is allowed. If the entry has claude_path — use it as {claude_root}; otherwise {claude_root}={path}.

Save to session context for subsequent skills:

• {project_id} = slug from projects.yml
• {project_path} = full path
• {project_name} = display name

These are read by the fast-track skill's Step 1 — if set, re-asking for the project is not needed.

Step 2.2 — Read project docs (by priority)

| Priority | File | What it provides | |---|---|---| | 1 | {path}/CLAUDE.md | Project rules, stack, push policy | | 2 | {claude_root}/.claude/skills/project-knowledge/SKILL.md | Architecture, patterns, DB schema | | 3 | {path}/README.md | General project overview |

If multiple found — read all. If none found — output a warning and continue without project context.

Step 2.3 — Project summary

One line: ✅ Project loaded: {name} ({domain}). [key facts].

Step 2.4 — Launch Flow Web UI (optional)

If a local Flow Web UI launcher script exists — start it and print clickable URLs scoped to {project_id}. If not installed — silently skip.

Anti-pattern: blocking go-start if the launcher errors. Print the error line and continue to Phase 3. Session start is more important than the UI.

---

Phase 3 — Session type

Step 3.1 — Ask session type

New session or continuing a task?
1. New — starting from a clean slate
2. Continuation — context was consumed, resuming a stuck task

Wait for response. Do not guess.

Step 3.2a — New session

User selected "new" → immediately call Skill('go-fast') without any intermediate text or questions. The fast-track skill's Step 1 will substitute the project from session context.

Anti-pattern: outputting "Ready. What are we doing?" before Skill('go-fast'). This breaks the chain — the session context transfer relies on going directly to go-fast.

Step 3.2b — Resuming a stuck task

User selected "continuation" → ask for a task number, title, or a paste from the interrupted session.

Step 3.3 — Find task

• Multi-line input → treat as paste from terminal. Extract task_id by pattern (first match): slug with --, "task N" reference, log folder path
• Short input → search directly (by number or by partial name match) in the routing DB

Verify the extracted task_id exists in artifacts. Not found → ask to clarify.

Step 3.4 — Read blocks

sqlite3 {routing_db} \
  "SELECT block_num, title, status, commit_hash
   FROM task_blocks WHERE task_id='{task_id}' ORDER BY block_num;"

No records → fallback: find the task file and read the ## Blocks section. Not found either → ask user to describe where they stopped.

Step 3.5 — Show blocks and review last

Status mapping:

| DB status | Symbol | |---|---| | done | ✅ | | in_progress / open (current) | 🔄 | | open (future) | ⏳ |

List the blocks, then read the last report-NN.md from the log folder via Read tool. From its ## Next section extract what's left → output "Done / Left / Continue?".

No reports → show blocks from DB and ask to describe where they stopped.

---

Anti-patterns

❌ Reading files outside the Phase 1 list

Agent decided to also read some other doc "while at it". Rule: Phase 1 reads exactly the five listed files. Everything else — only if the user explicitly asked.

❌ Guessing project without menu

User wrote /go-start and agent picked a project itself. Rule: menu is mandatory. Explicit selection every time, even if the previous session was in the same project.

❌ Skipping Phase 2 when "it's obvious"

Agent knows the project stack from past sessions and skips CLAUDE.md. Rule: every session starts from zero. CLAUDE.md may have changed. Phase 2 is mandatory always.

❌ Outputting full file summaries

Agent read WORKFLOW.md and reproduced it in full in the chat. Rule: summary is only operationally important info, no more than 8 lines per phase. Details stay in the read files.

---

Related skills

• go-fast — invoked at Step 3.2a for a new session
• project-knowledge — read at Step 2.2 for architecture/patterns

---

Step 99 — Log invocation

sqlite3 {routing_db} \
  "INSERT INTO skill_invocations (task_id, block_num, skill_name, invoked_at)
   VALUES ('{slug}', '{N}', 'go-start', datetime('now'))" 2>/dev/null || true

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bearded-illirian
• Source: bearded-illirian/trailmark
• License: MIT

Install and usage instructions live in the source repository linked above.