Sidewatch

Swift Agent Session — AI skill for Claude Code

AI community

How to install Swift Agent Session

This entry records only its repository, not the path inside it, so there is no exact command to give. Open Sidewatch/swift-agent-session and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Swift Agent Session does

A tiny, dependency-free reader for terminal AI coding-agent session transcripts — maps an agent's on-disk session onto one agent-agnostic model (activity timeline, token/cost telemetry, edited-files/to-dos roll-up) so review surfaces stay identical across agents. Read-only, no network, no telemetry.

Alternatives in AI

Cartographer — Claude Code plugin that maps and documents codebases of any size using parallel AI subagents 518 ★
NeoRead — NeoRead: a local-first EPUB/PDF reader with annotations, side-by-side translation, and AI assistance through C 294 ★
Wiki Timeline — Generate a chronological timeline for an entity or concept in the LLM Wiki 79 ★

README

Swift Agent Session

A tiny, dependency-free reader for terminal AI coding-agent session transcripts. It maps an agent's on-disk session onto one agent-agnostic model — an activity timeline, token/cost telemetry, and an edited-files/to-dos roll-up — so review surfaces stay identical across agents. Read-only: it never talks to a model, keeps no account, and sends no telemetry.

Features

🧭 Agent-agnostic model — TimelineEvent, AgentUsage, AgentSummary
🔌 Adapter protocol — implement AgentAdapter once per agent
📐 Turn boundaries — TurnBoundary splits the flat timeline into agent turns (one user prompt to just before the next), with a content-derived stable id so a checkpoint can be pinned to a turn
🤖 Claude Code adapter — ClaudeCodeAdapter parses ~/.claude/projects/…/*.jsonl transcripts. The only adapter shipped. Codex, Gemini CLI and OpenCode adapters existed until 11 Sep 2026 and were removed: their fixtures were written by hand from each tool's published schema and never run against a real session, and an adapter nobody can run reports "no session" forever without telling anyone. Recover them from git history, and only restore one with fixtures captured from a REAL run.
🕘 Activity timeline — prompts, assistant prose, tool calls, and file edits, with local-clock HH:MM timestamps
💰 Telemetry — context fill % (AgentUsage.contextPercent), output tokens, and estimated USD cost, deduplicated per API response
✅ Plan vs actual — the files edited (Edit / Write / MultiEdit / NotebookEdit). summary.todos is currently always empty: it parses Claude Code's TodoWrite tool, which was renamed to TaskCreate/TaskUpdate; across 63 real transcripts TodoWrite appears zero times. Fixable, and a good example of why anything keyed on another tool's vocabulary needs a test that fails when that vocabulary moves.
🔎 Auto-detection — Agents.active(for:) picks the agent that owns a project
🧪 **Fu