shiplog
The captain's log for your codebase. Every decision, discovery, and trade-off — captured in GitHub as you ship.
Quick Install
npx skills add devallibus/shiplog --skill shiplog
Works with Claude Code, Codex, and Cursor. Only requires gh CLI.
Your AI assistant forgets everything between sessions
You brainstormed a design with Claude yesterday. Today, a new session starts from scratch. The reasoning behind that auth middleware? Gone. The three alternatives you rejected? Nowhere to be found. Last week's architectural decision? Buried in a chat log you can't search.
shiplog makes your git history remember everything. Not just what changed — but why it changed, what you considered, and what you discovered along the way.
Brainstorm → Issue → Branch → Commits → PR
↑ ↑
| Discoveries → New Issues / Stacked PRs
|
└── Search it all later: issues, PRs, commits, memory
What you get
Every brainstorm becomes a GitHub Issue. Design decisions, alternatives considered, and a task list — all searchable forever.
Every commit carries context. Not just "add JWT validation" — but why JWT over sessions, what you tried first, and what to watch out for.
Every PR tells the full story. A timeline of the entire journey: what was planned, what was discovered mid-work, what changed and why.
Nothing falls through the cracks. Find a sub-problem while coding? shiplog routes it — fix inline, stack a prerequisite PR, or spin off a new issue. Your discovery is never lost.
Any model can pick up where another left off. Every artifact is signed with who wrote it (which AI model, which tool). Context handoffs between models are first-class — not copy-paste.
How it works
You say "let's brainstorm auth middleware" and shiplog captures the result as a GitHub issue. When you say "work on #42", it creates an isolated branch in a git worktree. As you commit, it logs the reasoning. When you open the PR, it writes a timeline of the whole journey.
Your existing workflow, with a knowledge trail that persists across sessions, models, and tools.
| Step | What happens |
|---|
| Brainstorm | Design discussion becomes a GitHub Issue with tasks |
| Branch | Isolated worktree created, timeline started, plan loaded |
| Discover | Sub-problems routed: inline fix, stacked PR, or new issue |
| Commit | Conventional commits with context comments on significant changes |
| Ship | PR with full journey timeline, decisions, and lessons learned |
| Search | Find any past decision across issues, PRs, and commits |
Two modes
Full Mode (default) — Knowledge goes directly into issues and PRs. Perfect for personal projects and open source.
Quiet Mode — Your team sees clean PRs. The full reasoning lives in a separate --log branch, one click away.
main
└── feature/auth-middleware ← Clean PR (team sees this)
└── feature/auth-middleware--log ← Knowledge trail (one click away)
Features
Cross-model review
No PR merges without review from a different AI model or a human. A single model authoring, reviewing, and merging its own work is the anti-pattern shiplog prevents. Reviews carry signed Reviewed-by: lines, support four dispositions (approve, approve-with-follow-ups, request-changes, comment), and generate self-contained review contracts when spawning a reviewer isn't possible.
Agent identity signing
Every artifact carries a provenance signature — Authored-by: claude/opus-4.6 (claude-code). The signing system auto-detects the current model from the platform and makes everything searchable:
gh issue list --search "Authored-by: claude/" # all Claude artifacts
gh pr list --search "Reviewed-by:" # all reviews
Model-tier routing
Use your best model for brainstorming, a fast one for implementation. shiplog prompts you to switch at phase transitions and writes context handoffs so the receiving model can execute without guessing.
| Tier | Best for | Example models |
|---|
| tier-1 (reasoning) | Architecture, trade-offs, PR synthesis | Claude Opus, o3 |
| tier-2 (capable) | Context loading, structured docs | Claude Sonnet |
| tier-3 (fast) | Implementation, routine commits | Claude Haiku, GPT-4o-mini |
Delegation contracts
When a reasoning model hands work to a faster model, the handoff is structured: allowed files, forbidden changes, stop conditions, verification requirements, and decision budget. If a tier-3 model reading the handoff would need to make a judgment call, the handoff is not specific enough.
Runtime-aware orchestration
