|
| 1 | +--- |
| 2 | +title: A first foray into agentic coding |
| 3 | +description: "I\u2019ve been largely steering clear of the AI bandwagon up to now, |
| 4 | + mainly because the last thing I needed while working on Relocatable OCaml was to |
| 5 | + be trying to learn a new tool at the same time as finishing something which felt |
| 6 | + desperately late and which, at times, I myself didn\u2019t fully understand! However, |
| 7 | + as the buzz of the announcement dies down, I had a first go at pointing Claude Code |
| 8 | + to the OCaml compiler codebase, and thought I\u2019d add to the general noise of |
| 9 | + AI memoirs\u2026" |
| 10 | +url: https://www.dra27.uk/blog/platform/2025/09/17/late-to-the-party.html |
| 11 | +date: 2025-09-17T00:00:00-00:00 |
| 12 | +preview_image: |
| 13 | +authors: |
| 14 | +- "" |
| 15 | +source: |
| 16 | +ignore: |
| 17 | +--- |
| 18 | + |
| 19 | +<p>I’ve been largely steering clear of the AI bandwagon up to now, mainly because |
| 20 | +the last thing I needed while working on <a href="https://www.dra27.uk/(/blog/platform/2025/09/15/relocatable-ocaml.html)">Relocatable OCaml</a> |
| 21 | +was to be trying to learn a new tool at the same time as finishing something |
| 22 | +which felt desperately late and which, at times, I myself didn’t fully |
| 23 | +understand! However, as the buzz of <a href="https://discuss.ocaml.org/t/relocatable-ocaml/17253">the announcement</a> |
| 24 | +dies down, I had a first go at pointing <a href="https://claude.ai/">Claude Code</a> to the |
| 25 | +OCaml compiler codebase, and thought I’d add to the general noise of AI |
| 26 | +memoirs…</p> |
| 27 | + |
| 28 | +<p>These are just my observations at this “day 1” stage, so there’ll be |
| 29 | +inadequacies in the way I’m doing things, and so forth. The result can be seen <a href="https://github.com/dra27/ocaml/compare/535463082ea48f71fac945bdc242ab0774a6f18b...dra27:ocaml:f364eade826314b1b066597e6e126f12e9f8b29a?expand=1">on my OCaml fork</a>.</p> |
| 30 | + |
| 31 | +<p>The task I had a play with is related to an interest I have in ultimately being |
| 32 | +able to use OCaml as a scripting language within OCaml programs themselves. Many |
| 33 | +of the things I do <em>to</em> OCaml often end up being curiously related or having |
| 34 | +seemingly unrelated tasks, and one of the things implied by this goal is |
| 35 | +unifying the bytecode and native runtime libraries (<code class="language-plaintext highlighter-rouge">libcamlrun</code> and |
| 36 | +<code class="language-plaintext highlighter-rouge">libasmrun</code>). At the moment, these are separately compiled, with some files |
| 37 | +which are just for the bytecode runtime or just for the native runtime, most |
| 38 | +files which are shared between the two, and some files which are peppered with |
| 39 | +<code class="language-plaintext highlighter-rouge">#ifdef NATIVE_CODE</code>-style sections for the bits which are slightly different |
| 40 | +between the two implementations.</p> |
| 41 | + |
| 42 | +<p>I’m pleased to say that in my few hours of playing with this, I managed not to |
| 43 | +edit any files, so every change in each commit is physically Claude’s.</p> |
| 44 | + |
| 45 | +<p>As a warm-up, I got Claude to produce a very simple plugin demo (the fourth |
| 46 | +commit). I was particularly impressed at Claude’s understanding how to use a |
| 47 | +compiler build in-tree, without having to install it. Although it slipped up |
| 48 | +occasionally on where directories were relatively, it was inferring quite |
| 49 | +readily the need for <code class="language-plaintext highlighter-rouge">-I ../stdlib</code> (because the compiler isn’t installed) and |
| 50 | +so forth and although it didn’t grasp the need to pass bytecode programs |
| 51 | +directly to the interpreter (you have, for example, to run |
| 52 | +<code class="language-plaintext highlighter-rouge">../runtime/ocamlrun ../ocamlc</code> rather than <code class="language-plaintext highlighter-rouge">../ocamlc</code>), there was an |
| 53 | +impressive lack of going-in-circles when I suggested alternate commands |
| 54 | +(incidentally, the <code class="language-plaintext highlighter-rouge">Makefile</code> is hideous because I requested it to be a |
| 55 | +complete expansion without variables and extraneous rubbish, not because it was |
| 56 | +offered by default!).</p> |
| 57 | + |
| 58 | +<p>So far, so interesting, but what about actually working on the compiler itself? |
| 59 | +Interesting side-line with this - there’s a reasonable amount of scepticism |
| 60 | +about LLM-based contributions in OCaml core development at the moment. This |
| 61 | +branch isn’t ever heading upstream, but if it were to be opened as a PR, I need |
| 62 | +to sign off on literally every edit that’s been made as being owned by |
| 63 | +me - which means I wasn’t expecting ever to leave Claude in “auto mode”. More on |
| 64 | +that later.</p> |
| 65 | + |
| 66 | +<p>We started exploring a few of the source files which differ between bytecode and |
| 67 | +native code. <em>I</em> suggested the first two commits while examining |
| 68 | +<code class="language-plaintext highlighter-rouge">backtrace_byt.c</code>, <code class="language-plaintext highlighter-rouge">backtrace_nat.c</code> and <code class="language-plaintext highlighter-rouge">backtrace.c</code>, but Claude readily did |
| 69 | +the work. I changed my mind about what I wanted to look at at that point and |
| 70 | +decided to explore something different, so looked at a file which is shared, but |
| 71 | +which has different implementations of the bytecode and native code versions of |
| 72 | +the same functions - <code class="language-plaintext highlighter-rouge">callback.c</code>. The functions here are what are used from C |
| 73 | +code to be able to call back into the OCaml runtime. That’s an very interesting |
| 74 | +thing to look at when unifying the code, because it’s an entry-point from user |
| 75 | +code into OCaml - i.e. the functions themselves will need to know whether |
| 76 | +they’re operating on bytecode or native code.</p> |
| 77 | + |
| 78 | +<p>We did it in two stages: the first part was to tweak the representation of |
| 79 | +closures in the runtime to allow the closures to describe if they are bytecode |
| 80 | +or native code. Claude unquestionably shone at this stage. I’m aware at a |
| 81 | +high-level of <a href="https://github.com/ocaml/ocaml/pull/9619">ocaml/ocaml#9619</a> (I |
| 82 | +remember very well the meeting where it was discussed, because it’s also the |
| 83 | +meeting where I first aired the idea for Relocatable OCaml!), but I certainly |
| 84 | +didn’t have it at the front of my mind. Claude was able in a very short amount |
| 85 | +of time both to remind me how the closure representation works (with reference |
| 86 | +to the runtime code) and furthermore produced an explanation which - for me, |
| 87 | +knowing how it works - was convincing that a bit could be stolen from the |
| 88 | +environment offset in the descriptor to be able to record whether a closure is |
| 89 | +actually bytecode (pass the first field to <code class="language-plaintext highlighter-rouge">caml_bytecode_interpreter</code>) or is |
| 90 | +native code (execute it directly, via one of the assembly callback stubs).</p> |
| 91 | + |
| 92 | +<p>Getting the runtime updated and building was quite painless. There were two |
| 93 | +interesting bits after that: the build failed quite a way in with segfaults |
| 94 | +coming from the compiler. While I was impressed that Claude had inferred how |
| 95 | +our bytecode bootstrap procedure works (perhaps it was trained on one of our |
| 96 | +compiler courses!), that was a “yeah, but no” moment - bytecode closures are |
| 97 | +always constructed at runtime, so how could <code class="language-plaintext highlighter-rouge">boot/ocamlc</code> be involved (“You’re |
| 98 | +so right!”). The next suggestion was to attempt to start gdb. At that point, we |
| 99 | +used the human brain instead of the LLM - I ran the build myself and could see |
| 100 | +the command that what was actually segfaulting was not <code class="language-plaintext highlighter-rouge">ocamlc</code> but |
| 101 | +<strong><code class="language-plaintext highlighter-rouge">ocamlc.opt</code></strong>, and for me there was an instant crashing-sound of a penny on |
| 102 | +the floor. Claude had dutifully updated the runtime for this new closure |
| 103 | +representation, and bytecode was therefore fine - but the native code compiler |
| 104 | +<em>does</em> emit closures at compile-time and needed updating.</p> |
| 105 | + |
| 106 | +<p>This part I must say was terribly impressive. I merely had to prompt that the |
| 107 | +native compiler hadn’t been updated and Claude had leapt to |
| 108 | +<code class="language-plaintext highlighter-rouge">asmcomp/cmm_helpers.ml</code> before I’d had a chance to remember the filename. The |
| 109 | +build then worked and we moved on to running the testsuite.</p> |
| 110 | + |
| 111 | +<p>Only one failure - again, Claude was very ready to start going into gdb and to |
| 112 | +use all sorts of other sledgehammers, but it was clear to the human brain that |
| 113 | +this was just a reference file which needed updating.</p> |
| 114 | + |
| 115 | +<p>Quite impressive - we committed the work so far. Less impressive was Claude’s |
| 116 | +ability to craft two commits from the working directory. I confess I got bored |
| 117 | +trying to persuade it to unstage a single hunk.</p> |
| 118 | + |
| 119 | +<p>Having updated the representation, we then tackled the much bigger job of |
| 120 | +merging the functions. I was unimpressed by Claude’s suggestions about how it |
| 121 | +might structure the changes, but it was very impressive at inferring the |
| 122 | +connection between the stated target and my suggestions about what to do |
| 123 | +instead.</p> |
| 124 | + |
| 125 | +<p>The refactor here has a small subtlety: beforehand, the files are being compiled |
| 126 | +with one version where <code class="language-plaintext highlighter-rouge">NATIVE_CODE</code> is <code class="language-plaintext highlighter-rouge">#define</code>‘d, and another where it’s not. |
| 127 | +Although the ultimate aim is to have just one function, it doesn’t actually |
| 128 | +reduce it to just one version of the code. We still have two: one of them is the |
| 129 | +version of OCaml where both bytecode and native are available, but there is |
| 130 | +another where there is no native code version at all. Claude readily inferred |
| 131 | +why I wanted <code class="language-plaintext highlighter-rouge">BYTECODE_ONLY</code> to be introduced for this second case (and, if you |
| 132 | +look at the <code class="language-plaintext highlighter-rouge">bytecode_callbackN_exn</code> and similar declarations which are |
| 133 | +<code class="language-plaintext highlighter-rouge">#ifdef</code>‘d with this, you can see the idea), but when further problems happened |
| 134 | +in the build, it too readily wanted to reach back for <code class="language-plaintext highlighter-rouge">#ifdef NATIVE_CODE</code>. This |
| 135 | +kind of refactoring I describe as “infectious” - there’s a relatively small |
| 136 | +first change in one file, but then further non-trivial changes have to be made |
| 137 | +to propagate <em>the semantics</em> of that update. There were four changes needed to |
| 138 | +make this one work:</p> |
| 139 | + |
| 140 | +<ol> |
| 141 | + <li>Some bytecode-only shim versions of some native code functions are needed, |
| 142 | +in order to be able to link the native code support object (<code class="language-plaintext highlighter-rouge">amd64.S</code> et al) |
| 143 | +in the bytecode runtime.</li> |
| 144 | + <li>The fiber stack C declarations need to be unified between bytecode and |
| 145 | +native code, as they use different pointer types.</li> |
| 146 | + <li>Some native-only functions, while not needed in the bytecode runtime, need to |
| 147 | +be available just so that other code links. This is to support the code path |
| 148 | +of invoking a native code closure on the bytecode runtime. It’s not going to |
| 149 | +happen (yet), but the code still needs to compile.</li> |
| 150 | + <li>Some native-code shim versions of some bytecode functions are needed for the |
| 151 | +dual reason of 1! Noteworthy because…</li> |
| 152 | +</ol> |
| 153 | + |
| 154 | +<p>Claude inferred none of these steps, always preferring to go back to |
| 155 | +<code class="language-plaintext highlighter-rouge">NATIVE_ONLY</code>. <em>But</em> it was very good at executing the changes when I suggested |
| 156 | +them, and inferring how they were alternate solutions to the problem.</p> |
| 157 | + |
| 158 | +<p>We got through those changes fairly painlessly and, at this point, the |
| 159 | +distribution built. Claude seems desperately keen to ignore the testsuite and |
| 160 | +write test programs of its own, but I worried that its fingers would wear out |
| 161 | +and dogmatically kept saying, “no, run the testsuite”. A lot of failures. Again, |
| 162 | +not the greatest insights from it straight-away. At this stage, the human brain |
| 163 | +was doing some staring at the code (maybe in the future I’d more readily let it |
| 164 | +sit there crunching tokens). It turns out both the human brain and the LLM had |
| 165 | +made a silly mistake with the header change and forgotten about pointer |
| 166 | +arithmetic. <em>Nil point</em> to Claude for the lack of inference originally - but I |
| 167 | +was certainly impressed that the prompt “The failing tests are all in bytecode. |
| 168 | +I’m wondering if you made a mistake with the change to the header” which |
| 169 | +<em>immediately</em> caused the LLM to identify that the switch from <code class="language-plaintext highlighter-rouge">value *</code> to |
| 170 | +<code class="language-plaintext highlighter-rouge">void *</code> had totally omitted the need for more pointer casts (this is a large |
| 171 | +part of what I don’t like about the change at the moment, but it’s a WIP… I |
| 172 | +often find when hacking these ideas that it’s necessary to go through some very |
| 173 | +ugly intermediate C states!).</p> |
| 174 | + |
| 175 | +<p>After a fair bit of “think very hard”, all casts were updated and the entire |
| 176 | +testsuite passed. As soon as that happened, Claude wanted to re-test the tree |
| 177 | +with a bytecode-only build (to verify that the new <code class="language-plaintext highlighter-rouge">BYTECODE_ONLY</code> version was |
| 178 | +working correctly too). And at that point we committed this little foray.</p> |
| 179 | + |
| 180 | +<p>Impressions, perhaps to revisit in a few months:</p> |
| 181 | +<ul> |
| 182 | + <li>Claude was excellent at getting the first version up and putting the outline |
| 183 | +plan for the change together. It was certainly slower (than me) at editing the |
| 184 | +files having done it. That, in fairness, fits with my experience of pair |
| 185 | +programming, and perhaps reflects more on me than Claude.</li> |
| 186 | + <li>The first ideas Claude wanted to make for how to proceed were almost always |
| 187 | +dreadful, and given the complexity of the code, I’m not sure I’d ever want to |
| 188 | +leave it doing large amounts of work on its own. That probably says a lot |
| 189 | +about my “prompt engineering”.</li> |
| 190 | + <li>In these few hours of “vibing”, I poked at builds, viewed source files on my |
| 191 | +own in order to agree, but didn’t edit anything directly myself. Will |
| 192 | +certainly be interesting to see what happens if Claude’s doing less of the |
| 193 | +editing, but watching and doing more of the analysis.</li> |
| 194 | + <li>Given the need to sign off personally on every change made, while the task |
| 195 | +felt like it took longer, it should be tempered with the fact I would feel |
| 196 | +less need to review the change afterwards (although the final commit is still |
| 197 | +not at a state which would be submitted in a PR).</li> |
| 198 | + <li>It’s possible that left to its own devices, Claude would have arrived at a |
| 199 | +working version of this small step without intervention. However, based on |
| 200 | +its attempt to describe a high-level plan of an earlier bigger idea, I dread |
| 201 | +to think what the code would have looked like!</li> |
| 202 | + <li>Its response to error messages and the speed with which it gets to a |
| 203 | +resolution definitely makes the process “feel” faster, even though it was |
| 204 | +actually slower than doing it myself (a social media, echo-chambery feeling)</li> |
| 205 | +</ul> |
| 206 | + |
| 207 | +<p><em>The commits referenced on my GitHub fork of OCaml in this post are included for |
| 208 | +information and illustration only, and are not intended to be upstreamed to |
| 209 | +OCaml.</em></p> |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4&size=40){kind=avatar}
0 commit comments