Re: Assertion failure with git cat-file --batch-command

On Mon Jul 27, 2026 at 11:57 AM CEST, Jeff King wrote:
> On Mon, Jul 27, 2026 at 10:30:43AM +0100, Alan Stokes wrote:
>
>> I first observed this in 2.43.0, but it still seems to be present in
>> 2.54.0.
>
> Yeah, I think this has been there since --batch-command was added.
>
>> Note that if I ask git cat-file --batch-command to include the
>> objecttype in the output it is fine (which gives me a workaround). Or
>> if I use git cat-file --batch.
>>
>> IIUC git only fetches the metadata that it needs for each object, and
>> that is determined from the format. For --batch I guess the type is
>> always requested, since it is needed to print the object contents. But
>> for --batch-command that doesn't seem to happen.
>
> Yes, exactly. In the normal --batch code path we have this code:
>
>         /*
>          * If we are printing out the object, then always fill in the type,
>          * since we will want to decide whether or not to stream.
>          */
>         if (opt->batch_mode == BATCH_MODE_CONTENTS)
>                 data.info.typep = &data.type;
>
> But for command mode, we don't do the same. This makes your case work:
>
> diff --git a/builtin/cat-file.c b/builtin/cat-file.c
> index 1458dd76d6..78eab9723d 100644
> --- a/builtin/cat-file.c
> +++ b/builtin/cat-file.c
> @@ -690,6 +690,7 @@ static void parse_cmd_contents(struct batch_options *opt,
>  			     struct expand_data *data)
>  {
>  	opt->batch_mode = BATCH_MODE_CONTENTS;
> +	data->info.typep = &data->type;
>  	batch_one_object(line, output, opt, data);
>  }
>
>
> but there's a slight catch. That expand_data is used for every request,
> not just the current one. In normal --batch mode, every request wants
> the same data (the user-specified format plus the object contents). But
> in command mode, some may be "contents" requests and some may just be
> "info". The code above turns on type-checking for every request, making
> the "info" ones pay to look up the type.

Yes, for example, both 'info' and the 'remote-object-info' series
(marked to 'master' in the last "What's cooking") [1] act on
data->info.typep.

This would make 'info' do a type lookup, and 'remote-object-info'
request "type" even if it wasn't present on the format.

>
> A type lookup isn't all that expensive, but it might matter for some
> formats (e.g., just "%(objectname)" does an existence check and nothing
> else, so we never even access the object data).

Yes, and only the atoms in the format get expanded, a populated type
without its atom in the format won't be shown.
the wasted lookup or a bigger request are the only effect.

>
> I guess saving and restore data->info.typep would work.

Yes I think that too, I tried this and it worked fine:

static void parse_cmd_contents(struct batch_options *opt,
			     const char *line,
			     struct strbuf *output,
			     struct expand_data *data)
{
	enum object_type *saved = data->info.typep;

	opt->batch_mode = BATCH_MODE_CONTENTS;
	data->info.typep = &data->type;
	batch_one_object(line, output, opt, data);
	data->info.typep = saved;
}

nit: On the current code the parameters aren't indented correctly.

>
>> I'm not sure what the correct fix is - always request the type in
>> --batch-command, or perhaps only if a "contents" command is issued?
>
> Yeah, in general if you are asking about "contents" I'd expect you to
> get the full name/type/size triple. But it's not wrong to ask for less,
> and certainly we should never hit a BUG(). So I think we'd want a fix
> along the lines above.
>
> Do you want to try your hand at a patch? It would need to do the
> save/restore, and most importantly add a new test to t1006.
>
> -Peff

[1]: https://lore.kernel.org/git/[email protected]/

Hope this helps,
Pablo