Thanks everyone who sent in comments and suggestions. They were immensely helpful.

I’ve gone ahead and added the following to my PPC to-do list:

1. Application Stream: This will be an optional time stream (enable in the preferences) that can be turned on to track which applications are active at any given slice of time.

2. URL Stream: This will be an optional time stream that tracks the current URL of the Safari and Firefox web browsers. I’ll probably break this into two time streams, one for each browser.

3. Rename the “Statistics” tab to “Reports”: I plan on implementing this as an en export option that spits out your time streams as an XML file. Scripts and other programs can be run on this file to generate specific reports of how time has been spent. I’ll include some options for dropping such scripts in the application support folder for running the reports straight from the interface.

4. AppleScript support: Initially this will be implemented as two main classes of commands: reading and writing. Reading will allow calling scripts to get a full lists of slices in a given time slice (“tell app ‘PPC’ to return every slice of the timestream named ‘Foo’). I’m also planning on implementing a feature where calling AppleScripts can log slices to streams (“tell app ‘PPC’ to log ‘Eating’ to timestream named ‘Foo’).

Fortunately (with the exception of the reporting functionality), I’ve already written most of this code for other projects. I’m going to keep plugging away and see if I can have an early beta version released sometime next week.

One of the simultaneously exciting and frustrating things working on Pennyworth is the system’s potential. On one hand, I think that the Pennyworth system has within it the right stuff to really make computing more convenient, effective, and less frustrating. On the other hand, I can’t really demonstrate how this vision “works” without some sort of obvious application. Pennyworth by itself is able to take action on context changes using the built-in AppleScript functionality, but regular people don’t seem to really get this.

One of my major tasks over the summer is to conduct a directed search for the killer app that really lets Pennyworth shine. I’m following the Growl model of releasing the system as a service that is accessible to other applications, and I need to find Pennyworth’s Adium.

While I wait for a project proposal to snake through the appropriate review boards (a cost of doing business in academia), I’ve been thinking about applications that I could release to the general public that helps me explore this space to begin to understand what people want and need. I am happy to say that I’ve been working on a fairly simple application that readers of this weblog might find useful.

Pennyworth Punch Clock (PPC) is a time tracking application that uses Pennyworth to help you recall and analyze how, where, and with who you spend your time with. PPC does this by using the idea of time streams and time slices. A time stream is a timeline that describes the temporal properties of one aspect of your person. For example, a location stream might look like “home, train, school, park, school, train, home”. The elements of the stream are called time slices. In other words, a stream is an ordered collection of slices.

By default Pennyworth exposes three time streams: Activity, Location, and Social Context. PPC listens for changes in these streams and logs them automatically. However, you will likely prefer to define other streams that you populate automatically: projects, official roles, billing rates, phone calls, and so on. To be honest, I expect that my notion of useful time streams differs greatly from the next person, and PPC is designed to accommodate these differences by being as general as possible.

The application is still in its early stages of development, but I wanted to publish a few screenshots of the work in progress to spark a conversation about who might consider using this software and how.

This first window shows the main time log. Time streams are in the upper half of the split pane and slices logged to each stream are visible in the lower half. In this screenshot, I have the “Activity” stream selected. The logs may be filtered by time as shown with the pulldown containing various options for narrowing or expanding the scope of the log.

In user-created streams (“Activity” is not one of them), you can log new slices to a stream by clicking the “Update Selected Stream” button. This invokes a small dialog where you can select a previously-used slice name or enter a new one.

In the preferences tab, you can create or remove new user-defined streams in addition to setting the standard update checking options. In this screen, I’ve created a “Project” stream that I can use to tag time periods when I’m working on different projects. If you click the magnifying glass button, the “Projects” time stream details appear, including an associated color and a list of saved values:

This is some of the more mundane aspects of the interface, and I’d be interested in learning more from you about what you might find useful in an application such as this. Personally, I plan on using this application to accomplish the following tasks:

• Create a finer-grained log of how I spend my time. In particular, I am interested in tracking how much time I spend on different projects for different “clients”.

• Better understand how I spend my time throughout the day. I’d like to have some hard data about how much time I spend dealing with e-mail, idly surfing the web, and so on in order to be able to identify and fix particular problem areas. I’d also like to better understand when I am doing these things (during lunch, late afternoon, etc.).

• Generate better time estimates for getting things done. At the moment, I don’t have enough information about how long it takes to do something like set up a web site, write a 10 page paper, etc. to come up with anything better than off-the-cuff predictions about how long something similar will take in the future.

Once I get the statistical and reporting features of the application in decent shape, I plan on thinking about how to make this application proactive in the sense of providing prompts to the user. For example, I may set up a built in prompt that nags me when the system detects I’m playing games, but I haven’t allocated time for gaming. Another prompt would be one that encourages me to do a little exercise and stretch after programming for a long stretch to reduce the chances of RSI.

I am interested in hearing from any interested readers out there. If you had a convenient and flexible tool for tracking and tagging how you spend your time, what would you do with it?

It took me long enough, but I’m happy to report that the next version of my context sensing toolkit will include some new functionality provided via AppleScript.

First of all, you can query the system for current predictions and observations:

tell application "Pennyworth"
  value of prediction named "Location"
end tell

might return something like Home. Likewise,

tell application "Pennyworth"
  value of observation named "Desktop Observer (Running)"
end tell

might return something like Safari.

Secondly, if you use other apps that call AppleScripts in response to events, you can now instruct the system to log the observation using

tell application "Pennyworth"
  log "My Observation" from observer "My Custom Observer" with duration 10
end tell

This is equivalent to an internal observer logging observations. The “from observer” and “with duration” parameters are optional. The system will choose reasonable defaults if left empty.

This functionality has been requested from a few users, so I’m glad that it is now working. If you have any comments or questions, please post them in the comments below.

The recent Mac Developer Roundtable podcast featured a good deal of information and discussion of context-sensitive interfaces and applications on the Mac platform (start listening around 50:00). I was surprised to hear one of the panel members, David Symonds, talk about a product on the Mac that sounded extremely similar to my own work. I feared that I’d reinvented the wheel and someone had build what I wanted before I even realized it.

The product in question is MarcoPolo, a context-sensing application that uses sensors to detect the user’s “context” and take actions accordingly. The interface design is eerily similar to my own and there is quite a deal of overlap of functionality between the two apps. However, there are a few significant differences:

1. MarcoPolo uses a rule-based system to make sense of observations. MarcoPolo does not learn from user-provided labels, but relies upon the user to define rules that define a given context. Genkanban uses an ID3 decision tree (or a naive Bayes classifier) to accomplish the same goal.

2. MarcoPolo seems to have a more system-oriented action functionality than Genkanban. Recall that Gankanban can initiate an arbitrary action using either distributed notifications (and listening apps) or executing AppleScripts on context changes. MarcoPolo includes a number of specific actions that are oriented toward configuring the local system for the new context. This includes setting up the network, configuring the default printer, etc. MarcoPolo includes a ShellScript action type that allows shell scripts to be executed on context changes. This can conceivably encompass Genkanban’s AppleScript functionality. (Update: ActionScripts can be directly executed using the “Open” action.)

3. MarcoPolo seems oriented toward the idea that location is largely the same as context. Genkanban takes a bigger picture view and explicitly looks at location, activity, and social contexts as discrete elements of context. This finer-grained formulation allows Genkanban to take specific actions based on changes in the specific components. For example, I only care about activity when updating my Adium status message. I probably only care about location when deciding whether to activate the system screensaver.

That said, both Genkanban and MarcoPolo are more similar than different. The way I see things, components like Genkanban’s machine learner can be replaced with MarcoPolo’s rule-based system without much fuss (and vice versa). Genkanban’s sensors work much like MarcoPolo’s, and there’s probably no good reason that a common sensor interface couldn’t be be written so both systems’ sensors are interchangeable. I couldn’t tell if MarcoPolo is issuing distributed notifications on context changes, but this may be another element that can be standardized.