Writing knitr hooks

Customising knitr output with hook functions is a handy skill, and I can never quite remember it when I need to, so here’s a blog post to remind myself
Literate Programming

December 30, 2023

A very common situation I encounter when writing a blog post or writing a book chapter using R markdown or quarto arises when the command I want to use generates a lot of output, and I don’t want all of the output to be displayed in the output. Every time I run into this problem, I have this vague recollection that “oh yeah, I need to write a knit hook for this”, but I can never quite remember how to do that and have to search online for the answer. In my last post I wrote a jokey footnote grumbling about this and saying I was thinking of writing a short blog post on it just so that I’d know where to look next time.

So, uh, yeah… that’s exactly what I did.

The answer to that specific question, incidentally, is described explicitly in the R Markdown Cookbook, and – to set expectations appropriately – there’s nothing in this post that isn’t already covered in the documentation and books. I don’t actually need to write a blog post about this. But I’m going to anyway, because every time I actually do need to write a knit hook, I find myself realising that I don’t understand them as well as I ought to. So here goes.

Chunk options

This is a post about knit hooks, but it helps to start with a refresher on knitr chunk options. I’m assuming, for the purposes of this post, that anyone reading is already pretty familiar with R markdown and quarto, and knows that when I write a document like this and want to execute some R code, I include an appropriately annotated code chunk in the source like so:

1 + 1
[1] 2

When the document is rendered to HTML, it’s the job of the knitr package to parse this chunk, execute the code, and append the output to the document as necessary. You can customise the manner in which knitr does this via chunk options, but the code chunk above doesn’t specify any options, so default values are used.

So what are the defaults, and where are they stored?

The knitr::opts_chunk object is used to control the options for code chunks. The object is a list of several functions. The two we use most often are $get() and $set().

options <- knitr::opts_chunk$get()

This options variable is a list containing all the default values that are applied when knitting the code chunks in the markdown document. If the user doesn’t specify a value for a specific chunk option, these are the default values that are applied. There’s a lot of these options:

 [1] "eval"          "echo"          "results"       "tidy"         
 [5] "tidy.opts"     "collapse"      "prompt"        "comment"      
 [9] "highlight"     "size"          "background"    "strip.white"  
[13] "cache"         "cache.path"    "cache.vars"    "cache.lazy"   
[17] "dependson"     "autodep"       "cache.rebuild" "fig.keep"     
[21] "fig.show"      "fig.align"     "fig.path"      "dev"          
[25] "dev.args"      "dpi"           "fig.ext"       "fig.width"    
[29] "fig.height"    "fig.env"       "fig.cap"       "fig.scap"     
[33] "fig.lp"        "fig.subcap"    "fig.pos"       "out.width"    
[37] "out.height"    "out.extra"     "fig.retina"    "external"     
[41] "sanitize"      "interval"      "aniopts"       "warning"      
[45] "error"         "message"       "render"        "ref.label"    
[49] "child"         "engine"        "split"         "include"      
[53] "purl"          "fenced.echo"   "ft.shadow"    

The fig.path option, for example, is used to specify where generated output images should be written. It’s a nice one to illustrate the customisability of knitr because you get a different output depending on context. The blog post is a quarto document, and has different knitr defaults to what you’d see if the same code were run at the console:

[1] "index_files/figure-html/"

If I’d run the same output at the console, I would get a different answer. When called from the console the default option for fig.path is "figures". For example, when I constructed my “knitr + eleventy” blog this is how I was calling knitr, and accordingly the images were written to a “figures” folder. The defaults, when knitr is used in the context of this on this quarto blog, are different.

Knit hooks

So now we turn to knitr hooks. Hooks are user-customisable functions that you can use to control how the knitr options are interpreted, and modify the output that knitr creates. In the same fashion that the knitr::opts_chunk object is used to control the chunk options, there’s a knitr::knit_hooks object used to control hooks. Again, this object is a list of functions, and the two we use most often are $get() and $set().1 We can retrieve the hooks by calling the $get() function:

hooks <- knitr::knit_hooks$get()

There are 12 default knit hooks in this list:

 [1] "source"          "output"          "warning"         "message"        
 [5] "error"           "plot"            "inline"          "chunk"          
 [9] "text"            "evaluate.inline" "evaluate"        "document"       

The documentation for output hooks gives a nice summary for most of these. Seven of the hooks are quite specific, and are applied to only one type of output:

  • source: Handles how knitr processes the source code inside a chunk
  • output: Handles how knitr processes ordinary R output (i.e., not warnings, messages, or errors)
  • warning: Handles how knitr processes warning output (e.g., from warning())
  • message: Handles how knitr processes message output (e.g., from message())
  • error: Handles how knitr processes error output (e.g., from stop())
  • plot: Handles how knitr processes graphics output
  • inline: Handles how knitr processes output from inline R code

There are two output hooks that are broader in scope:

  • chunk: Applied to all output from a code chunk
  • document: Applied to all output within the document

The other three (evaluate, evaluate.inline, and text) aren’t discussed as much, and while I did get a little curious and started going down a rabbit hole looking at them, for once in my life I’ll be smart and not get sucked all the way in.

Custom output hooks

The general advice when writing custom output hooks is that you shouldn’t try to write the whole thing yourself. By design, knitr will create default hooks that are appropriate to the specific context, and your safest bet is to first retrieve the default hook by calling the $get() function, like this:

default_hook_output <- knitr::knit_hooks$get("output")

Then you can write your own hook that does some pre-processing to the inputs, before passing the modified inputs to the default hook. So, having already saved the default hook as default_hook_output I’d write my custom output hook like this:

custom_hook_output <- function(x, options) {
  n <- options$out.lines
  if(!is.null(n)) {
    x <- xfun::split_lines(x)
    if (length(x) > n) x <- c(head(x, n), "....\n")
    x <- paste(x, collapse = "\n")
  default_hook_output(x, options)

There’s a few things going on here that are worth highlighting. First, notice that output hooks take two arguments x and options. The x argument is the raw text string that needs to be rendered: in this case, the string would correspond to the output that would normally be printed to the R console. The options argument is the list of knitr chunk options. The value of options that gets passed to the hook includes any values that were specified by the user in the chunk options, and also any default values that were not specified by the user. In this instance, out.lines is intended to indicate the maximum number of lines of R output to write to the rendered output document. It’s not one of the default chunk options (i.e., it wasn’t one of the options we saw in the previous section), and so if the user doesn’t specify a value for out.lines in the chunk options, options$out.lines will return a value of NULL in our custom hook, and so our custom_hook_output() will skip all the pre-processing in that case. However, if the user does specify a value for out.lines, it does a little text manipulation to alter the value of x before it is passed onto the default output hook.

Having written our custom hook, we apply it by using the $set() function:

knitr::knit_hooks$set(output = custom_hook_output)

Now that we have a knit hook that knows how to interpret out.lines as a chunk option, I can incorporate it into a knitr code chunk just like any other one:

```{r, out.lines = 4}
  [1] 0.26550866 0.37212390 0.57285336 0.90820779 0.20168193 0.89838968
  [7] 0.94467527 0.66079779 0.62911404 0.06178627 0.20597457 0.17655675
 [13] 0.68702285 0.38410372 0.76984142 0.49769924 0.71761851 0.99190609
 [19] 0.38003518 0.77744522 0.93470523 0.21214252 0.65167377 0.12555510

The output here would normally be considerably longer than 4 lines, but we’ve applied a custom hook that enforces the truncation, so we get nicer output. Notice also that, in the same way that standard chunk options like fig.width and fig.height become fig-width and fig-height when you’re setting them via custom code comments, our new out.lines option becomes out-lines when used in that context:

#| out-lines: 4
  [1] 0.26750821 0.21864528 0.51679684 0.26895059 0.18116833 0.51857614
  [7] 0.56278294 0.12915685 0.25636760 0.71793528 0.96140994 0.10014085
 [13] 0.76322269 0.94796635 0.81863469 0.30829233 0.64957946 0.95335545
 [19] 0.95373265 0.33997920 0.26247411 0.16545393 0.32216806 0.51012521

Custom chunk hooks

In the previous section, we effectively created a new chunk option called out.lines simply by virtue of modifying one of the standard output hooks that is able to interpret it and modify the output accordingly. That approach doesn’t always work, particularly if the new option that you want to create requires that code be executed before and after knitr processes the chunk. In those situations we may need to write a “chunk hook” that is triggered whenever the new chunk option has a non-null value. Chunk hooks have a different structure than output hooks. The R Markdown Cookbook has some nice examples of this, including one for timing how long it takes the chunk to execute. I’ll adapt that one here.

To understand how to write a chunk hook, the key thing to realise is that it gets called twice: once before knitr executes the code in the chunk, and once again afterwards. The function can take up to four arguments, all of which are optional:

  • before is a logical value indicating whether the function is being called before or after the code chunk is executed
  • options is the list of chunk options
  • envir is the environment in which the code chunk is executed
  • name is the name of the code chunk option that triggered the hook function

As a general rule, the chunk hook is called for its side effects not the return value. However, if it returns a character output, knitr will add that output to the document output as-is.

Designing a chunk hook that records the amount of time taken to execute takes a little thought. When the hook is triggered the first time (with before = TRUE) we want to record the system time somewhere (e.g., in a variable called start_time). Then, when the hook is triggered the second time (with before = FALSE) we want to record the system time again (e.g., as stop_time), and compute the difference in time. We can do this using a function factory to create stateful functions. Here’s what that looks like:

create_timer_hook <- function() {
  start_time <- NULL
  function(before, options) {
    if (before) {
      start_time <<- Sys.time()
    } else {
      stop_time <- Sys.time()
      elapsed <- difftime(stop_time, start_time, units = "secs")
        "<div style='font-size: 70%; text-align: right'>",
        "Elapsed time:", 
        round(elapsed, 2), 

When create_timer_hook() is called it returns a function that will become our custom hook. Or – to be more precise, because in this instance the distinction matters – it returns a closure. When called with before == TRUE, it records the system time and uses the super assignment operator <<- to store that value as start_time. Normally, an assignment that takes place during the function execution isn’t persistent and can’t be reused on later calls to that function. But we’ve structured things differently here: in this case, the start_time variable is defined in the enclosing environment (the one in which the function was defined) rather than the execution environment (in which the function body code executes). That changes things: the execution environment is inherently ephemeral and lasts as long as a single function call is in progress. The enclosing environment, however, is persistent, and will survive for (at least) as long as the function itself exists. As a consequence, the value assigned to start_time is persistent also, and still exists when the hook is triggered a second time with before == FALSE. That makes it possible to compute the difference between start_time and stop_time with difftime().

Having computed the elapsed time, all that remains is to format it a little bit and then return a nice character string with some HTML that will be printed in the final document. To put this into action, we set the custom hook like this:

knitr::knit_hooks$set(timer = create_timer_hook())

By doing this timer become the code chunk option that triggers the hook, and we can now use it in the document:

#| timer: true
#| out-lines: 4
    [1] 0.6588776091 0.1850699645 0.9543781369 0.8978484920 0.9436970544
    [6] 0.7236907512 0.3703570659 0.7810175403 0.0111495086 0.9403087122
   [11] 0.9937492262 0.3574057452 0.7476350635 0.7929090238 0.7058590064
   [16] 0.4758250387 0.4946545260 0.3080524488 0.6950122463 0.8227933056
Elapsed time: 0.03 secs

And with that, we are done!

Yes, there are other kinds of hooks that you can write for knitr,2 but the only two kinds of hooks I’ve ever actually had the need for myself are output hooks and chunk hooks, so in the interests of brevity I’ll leave it at that.


  1. In addition to knit_hooks and opt_chunks, knitr has several other objects that can be used to control the behaviour of the package. These are knit_patterns, knit_patterns, opts_current, and opts_knit. They all have the same basic structure, including $get() and $set() functions. These objects are documented here.↩︎

  2. There are also option hooks that you can use to modify the value of some options based on the values of other options, and those are managed by opts_hooks in the same way that knit_hooks manages output hooks and chunk hooks.↩︎



BibTeX citation:
  author = {Navarro, Danielle},
  title = {Writing Knitr Hooks},
  date = {2023-12-30},
  url = {https://blog.djnavarro.net/posts/2023-12-30_knitr-hooks},
  langid = {en}
For attribution, please cite this work as:
Navarro, Danielle. 2023. “Writing Knitr Hooks.” December 30, 2023. https://blog.djnavarro.net/posts/2023-12-30_knitr-hooks.