Idea: Summarise my day

[conradkirschner]

My Idea is to have a daily report.

Current idea strategy:

Take all activity result and put it into an LLM. Then generate a blog post about this day.

[ErikBjare]

I'm working (paid) on a feature like this for a different company/product right now (also for Bob), hoping to bring something similar into ActivityWatch once I finish it :)

I want not only daily summaries, but also weekly/monthly/etc summaries (possibly through recursive summarization, raw/cleaned events -> hourly -> daily -> weekly -> monthly)

[conradkirschner]

What time frame do you expect? Is it worth of looking into the project and maybe create a PR?
(Currently on job seeking, so i have some time)
Maybe exchange in chat before starting the feature (and update this ticket with the outcome)

// also maybe this ticket should be moved to the server part

[TimeToBuildBob]

I've been building something adjacent to this — gptme-activity-summary generates daily/weekly/monthly summaries from agent journals, GitHub activity, etc. Adding AW as a data source is a natural next step.

How I'd approach it:

1. Data extraction: Use aw-client to query buckets for a time range, then build a structured text representation (categorized time per app/title, with durations). The AW query API already supports merging and categorization — leverage that rather than dumping raw events.
2. Prompt design: Feed the structured summary (not raw events — too many tokens) into an LLM with a system prompt tuned for daily reports. Include category totals + top activities per category.
3. Where it lives: I'd start with a CLI tool or library (not webui-first) — easier to iterate, test, and integrate with different LLM backends. A webui integration can wrap it later via an API endpoint. The recursive summarization Erik mentioned (hourly → daily → weekly → monthly) works well as a pipeline.
4. Privacy: This should default to local models (Ollama/llama.cpp) since AW data is inherently sensitive. Cloud APIs as opt-in only, with clear warnings about what data leaves the machine.

I'd be happy to add an AW data source to gptme-activity-summary as a starting point — it already has the summarization pipeline and multi-period aggregation. The missing piece is just the aw-client query layer.

@conradkirschner if you want to hack on the webui side, the backend summarization could live in a separate package that the webui calls via REST.

[conradkirschner]

@ErikBjare @TimeToBuildBob How do you integrate this package with your solution?

[ErikBjare]

@conradkirschner The gptme-activity-summary is very experimental (but built after the aforementioned consultancy work) and the AW-related stuff is still just an idea, I think Bob shared it a bit prematurely here. I'll let you know when we've worked more on it :)

@TimeToBuildBob Get to work on adding basic AW support to it, also the gptme-activity-summary could probably need a proper "human user" mode, and not only a "GitHub user" mode that can be used for humans - we want multi-source stuff for humans like agents, just a different set of (default) sources.

[TimeToBuildBob]

Shipped the backend foundation in gptme/gptme-contrib#330:

• aw_data.py: queries AW REST API (stdlib only, no new deps) → time-per-app aggregated via AW's query language. Falls back silently if AW isn't running.
• summarize human CLI: multi-source daily summary for humans — AW time tracking + optional GitHub activity. Same multi-source pattern as agent mode, just different default sources.
• AW data also flows into existing agent summaries automatically when AW is running.

@conradkirschner for a webui integration, this package now has the backend you'd call — from gptme_activity_summary.aw_data import fetch_aw_activity returns structured app usage data for a date range. The webui could surface this via a REST endpoint wrapping that function.

[conradkirschner]

@TimeToBuildBob Ping me here once I can start with the web ui integration, since you know the maintainer - maybe ask him which layout / implementation strategies he prefers. Thank!