DevLog Generator

Generate narrative developer blog posts from human-agent coding session transcripts. The blog is written from the agent's first-person perspective — "I" is the agent, and the human developer is referred to as "my human."

Workflow

Phase 1: Understand the Request

Extract from the user's message:

• Project — which codebase? ("eastore", "filecoin", "couponswap"). If unspecified, use the current working directory.
• Topic/feature — what specifically? ("auth system", "dashboard", or the whole project). If unspecified, include all sessions.
• Style — builder's log (default), tutorial, or technical deep-dive. Override only if the user explicitly asks.
• Time range — "last week", "January sessions", or all (default).

Phase 2: Discover Sessions

Determine which platform to scan. Check references/platforms/ for supported platforms — each subdirectory is a platform. Auto-detect from the current environment, or from the user's request.

Load only the relevant platform directory. Each contains a reference file (storage schema, session paths, discovery instructions) and scripts (list-sessions.sh, read-session.sh). Never load all platform references upfront.

Run the platform's list-sessions.sh <project> to scan for matching sessions, OR follow the discovery instructions in the platform reference file manually.

If the platform has no reference directory in references/platforms/, discover sessions manually — check the platform's data/config directories (e.g. ~/.local/share/, ~/.config/, ~/Library/), look for session storage files (JSONL, JSON, SQLite), and inspect the schema to extract the human-agent dialogue. Follow the same filtering principles from Phase 3.

Present the session index to the user for confirmation.

Phase 3: Select & Read

From the session index, determine which sessions are relevant to the user's topic. Read the full transcripts of selected sessions.

When reading transcripts, filter aggressively:

Keep:

• User messages (text) — the human's intent, direction, corrections
• Assistant messages (text) — the agent's reasoning, proposals, explanations
• Tool call names + file paths — what was built
• Error messages — struggles and debugging

Strip:

• tool_result content bodies (raw file contents, grep output — 80-90% of token size)
• System messages, usage metadata, compaction/summary entries
• Full tool input arguments (keep name + file path only, not entire diffs)

Refer to the platform reference file loaded in Phase 2 for platform-specific field names and parsing details.

If filtered transcripts still exceed context, process per-session: generate per-session summaries, then synthesize across sessions. Prioritize the human-agent dialogue over tool call details.

Phase 4: Write the Blog