Deep diving into a major mode - Part 2 (IDE Features)

In <a href='/post/major-mode-part-1/'>Part 1</a> I addressed: syntax tables, indentation, font locking, and context-sensitive syntax.

We now build out a "Hyde" with: shell/async process integration, Eldoc support, and Autocompletion. Also "shift-K" documentation lookup for the symbol-at-point.


Building Inferior Hy

The basics

The easiest way to add a REPL to your major-mode is through setting the variable inferior-lisp-program, possibly adding setup code through inferior-lisp-load-command, then running the inferior-lisp command.

;; within the define-derived-mode hy-mode setup...
(setq-local inferior-lisp-program "hy")
(setq-local inferior-lisp-load-command "(print \"Hy there!\")")

This isn't sufficient for advanced shells. The custom is to create a function named run-hy (eg. there is run-python, run-haskell…) which starts up the shell, sets inferior-hy-mode , and switches to the shell.


It is a "Major mode for interacting with an inferior interpreter". Most-all inferior modes will derive from comint-mode. It provides many utilities for interacting with shell-like processes.

Hy mode's prompt is a right arrow.

(define-derived-mode inferior-hy-mode comint-mode "Inferior Hy"
  "Major mode for Hy inferior process."
  (setq-local indent-tabs-mode nil)

  ;; How to dispaly the process status in the mode-line
  (setq mode-line-process '(":%s"))
  ;; This disables editing and traversing the "=>" prompts
  (setq-local comint-prompt-read-only t)
  ;; Lets comint mode recognize the prompt
  (setq-local comint-prompt-regexp (rx bol "=>" space))

  ;; ... other specialized config introduced later ...

There are several comint components we will make use of:

  1. comint-last-prompt - a cons cell of begin/end markers of last prompt.

  2. comint-send-string - perform process-send-string with comint bookkeeping.

  3. comint-redirect-send-command-to-process and comint-redirect-completed - for sending strings asynchronously.

  4. comint-(pre)output-filter-functions - entry points into capturing and cleaning process output.

  5. make-comint-in-buffer - entry point into comint mode.

Managing buffers and processes

We must manage: the standard hy shell process, the internal hy process used for autocompletion and eldoc, and temporal buffers for more advanced buffer transformations of the standard hy shell process.

The configuration required of inferior-hy-mode :

;;; Configuration

(defconst hy-shell-interpreter "hy"
  "Default Hy interpreter name.")

(defvar hy-shell-interpreter-args "--spy"
  "Default arguments for Hy interpreter.")

;;; Internal

(defconst hy-shell-buffer-name "Hy"
  "Default buffer name for Hy interpreter.")

(defconst hy-shell-internal-buffer-name "Hy Internal"
  "Default buffer name for the internal Hy process.")

(defvar hy-shell-buffer nil
  "The current shell buffer for Hy.")

(defvar hy--shell-output-filter-in-progress nil
  "Whether we are waiting for output in `hy-shell-send-string-no-output'.")

(defvar hy--shell-font-lock-enable t
  "Whether the shell should font-lock the current line.")

Now the building blocks of the shell can be defined.

The implementations are rather straightforward. To keep space down, only name and docstring are provided:

(defun hy--shell-format-process-name (proc-name)
  "Format a PROC-NAME with closing astericks.")

(defun hy-shell-get-process (&optional internal)
  "Get process corr. to `hy-shell-buffer-name'/`hy-shell-internal-buffer-name'.")

(defun hy--shell-current-buffer-process ()
  "Get process associated with current buffer.")

(defun hy--shell-current-buffer-a-process? ()
  "Is `current-buffer' a live process?")

(defun hy--shell-get-or-create-buffer ()
  "Get or create `hy-shell-buffer' buffer for current hy shell process.")

(defun hy--shell-buffer? ()
  "Is `hy-shell-buffer' set and does it exist?")

(defun hy--shell-kill-buffer ()
  "Kill `hy-shell-buffer'.")

(defun hy--shell-calculate-command (&optional internal)
  "Calculate the string used to execute the inferior Hy process.")
  ;; Straightforward string formatting - see: `shell-quote-argument'

Starting up the shell

The commands above are enough to build out some basic shell support. Lets look at run-hy :

(defun run-hy (&optional cmd)
  "Run an inferior Hy process.

CMD defaults to the result of `hy--shell-calculate-command'."
  (unless (executable-find "hy")
    (message "Hy not found, activate a virtual environment with Hy."))

  (-> (or cmd (hy--shell-calculate-command))
     (hy--shell-make-comint hy-shell-buffer-name 'show)

Most of the work is delegated to hy--shell-make-comint as we also must have the internal variant:

(defun run-hy-internal ()
  "Start an inferior hy process in the background for autocompletion."
  (unless (executable-find "hy")
    (message "Hy not found, activate a virtual environment containing Hy to use
Eldoc, Anaconda, and other hy-mode features."))

  (when (and (not (hy-shell-get-process 'internal))
             (executable-find "hy"))
    (-let [hy--shell-font-lock-enable
          (-> (hy--shell-calculate-command 'internal)
             (hy--shell-make-comint hy-shell-internal-buffer-name nil 'internal)
        (message "Hy internal process successfully started")))))

This is a simple variation of run-hy that passes the internal argument through the hy shell building blocks and also sends setup code for eldoc-mode and company-mode.

Now we are ready to startup our inferior-hy-mode :

(defun hy--shell-make-comint (cmd proc-name &optional show internal)
  "Create and return comint process PROC-NAME with CMD, opt. INTERNAL and SHOW."
  (-when-let* ((proc-buffer-name
                (hy--shell-format-process-name proc-name))
                (not (comint-check-proc proc-buffer-name)))
                (split-string-and-unquote cmd))
                (apply 'make-comint-in-buffer proc-name proc-buffer-name
                       (car cmdlist) nil (cdr cmdlist)))
                (get-buffer-process buffer)))
    (with-current-buffer buffer
    (when show
      (display-buffer buffer))
    (if internal
        (set-process-query-on-exit-flag process nil)
      (setq hy-shell-buffer buffer))

All the work is once again delegated to our shell building blocks. There are several things to notice:

  1. If the process is meant to be autostarted/quited, make sure to use set-process-query-on-exit-flag to nil.

  2. The cmdlist car is "hy" and cdr is the hy interpreter arguments.

  3. Further accessing of the shell is done with the hy-shell-buffer variable.

Working with the shell

The shell is now functional, but we still don't have methods to send strings to the shell (for instance, sending the current-form or the buffer for evaluation). Nor do we have any support for asynchronously sending and extracting information from our internal process.

Like always, lets define some utilities:

(defun hy--shell-end-of-output? (string)
  "Return non-nil if STRING ends with the prompt."
  (s-matches? comint-prompt-regexp string))

(defun hy--shell-output-filter (string)
  "If STRING ends with input prompt then set filter in progress done."
  (when (hy--shell-end-of-output? string)
    (setq hy--shell-output-filter-in-progress nil))
  "\n=> ")

hy--shell-output-filter-in-progress is the critical component. Lets see how it is used:

(defun hy--shell-send-string (string &optional process internal)
  "Internal implementation of shell send string functionality."
  (let ((process (or process (hy-shell-get-process internal)))
        (hy--shell-output-filter-in-progress t))
    (comint-send-string process string)
    (while hy--shell-output-filter-in-progress
      (accept-process-output process))))

The shell process is obtained, we set it to be in progress, and send it off to comint. But how and when is the filter reset?

We come back to the comint-(pre)output-filter-functions. When we send the string via comint-send-string, part of its bookkeeping is to apply these filter functions to the output. However, the output can come in chunks, so simply accepting the process output is not sufficient. We must recognize when the last of the expected process output is retrieved and signal to stop accepting output.

The difference between the pre and standard filters is when they are applied. The pre variation is executed before the process output is inserted into the buffer.

Looking back to hy--shell-output-filter, what are the outcomes of using it as a pre or standard filter? Lets look at the exposed send strings:

(defun hy-shell-send-string-no-output (string &optional process internal)
  "Send STRING to hy PROCESS and inhibit printing output."
  (-let [comint-preoutput-filter-functions
    (hy--shell-send-string string process internal)))

(defun hy-shell-send-string (string &optional process)
  "Send STRING to hy PROCESS."
  (-let [comint-output-filter-functions
    (hy--shell-send-string string process)))

Lastly, these functions won't work for asynchronous ops like Eldoc and Autocompletion. You will see a Blocking call inhibiting process output error messaged in the minibuffer.

The asynchronous version is different. We redirect the process output to a temporary buffer and capture its output.

The key is the 100ms timeout argument passed to accept-process-output.

(defun hy--shell-send-async (string)
  "Send STRING to internal hy process asynchronously."
  (let ((output-buffer " *Comint Hy Redirect Work Buffer*")
        (proc (hy-shell-get-process 'internal)))
    (with-current-buffer (get-buffer-create output-buffer)
      (comint-redirect-send-command-to-process string output-buffer proc nil t)

      (set-buffer (process-buffer proc))
      (while (and (null comint-redirect-completed)
                  (accept-process-output proc nil 100 t)))
      (set-buffer output-buffer)

Our shell is now ready for autocompletion, eldoc, and other awesome IDE features.

I originally planned to go into font-locking the prompt input (highly non-trivial), but given the length of this post I will provide and link to it as a separate future post.


All the work for autocompletion was in setting up the asynchronous process support. Lets see how easy autocompletion becomes:

(defconst hy-company-setup-code
  "(import [hy.completer [Completer]])
(setv --HYCOMPANY (Completer))"
  "Autocompletion setup code to send to the internal process.")

(defconst hy--company-regexp
  (rx "'"
      (group (1+ (not (any ",]"))))
      (any "," "]"))
  "Regex to extra candidates from --HYCOMPANY.")

(defun hy--company-format-str (string)
  "Format STRING to send to hy for completion candidates."
  (when string
    (format "(.%s --HYCOMPANY \"%s\")"
            (cond ((s-starts-with? "#" string)  ; Tag matches broken in Hy atm
                  ((s-contains? "." string)

(defun hy--company-candidates (string)
  "Get candidates for completion of STRING."
  (-when-let* ((command (hy--company-format-str string))
               (candidates (hy--shell-send-async command))
               (matches (s-match-strings-all hy--company-regexp candidates)))
    (-select-column 1 matches)))  ; Get match-data-1 for each match

(defun company-hy (command &optional arg &rest ignored)
  (interactive (list 'interactive))
  (cl-case command
    (prefix (company-grab-symbol))
    (candidates (hy--company-candidates arg))
    (meta (-> arg hy--eldoc-get-docs hy--str-or-empty))))

Completer is a hy builtin that completes a given string and does all the work here. We simply call it's appropriate method, extract the items in the retrieved list, and hand it off to company.

company-grab-symbol gets the current symbol which is handed off as arg in the subsequent call.

The meta argument shows the eldoc output for the current selected company candidate in the minibuffer, as seen in this post's initial image.

company-hy can then be enabled either through adding to company-backends or for Spacemacs users adding:

;; Technically this should be within a hy layer, but this still works uncaptured
  :backends company-hy
  :modes hy-mode inferior-hy-mode)

Developing a major-mode that accommodates Spacemacs users will be touched on in future posts.


For those unfamiliar, in the initial image eldoc-mode provides the formatted docstring and arguments in the minibuffer for the symbol-at-point (or completion candidate).

Perhaps surprisingly, Eldoc is a lot more challenging than autocompletion.

For starters, your language won't provide Eldoc like strings (formatted argument list + first line of docstring) by default. How difficult inspecting language constructs is entirely dependent on the language. Hy in particular is difficult due to how macros are implemented and namespaced. I won't provide the hy-eldoc-setup-code here, it can be found within the source.

Next, your implementation must mirror any relevant DSLs. For lisps, Eldoc inspects the form opener. It is hydiomatic to:

(setv x "hi")
(.format "{} there" x)
(setv a-list [])
(.append a-list "friend")

We need to send str.format and a-list.append - the form opener alone is insufficient.


Eldoc is setup via the eldoc-documentation-function :

(defun hy-eldoc-documentation-function ()
  "Drives `eldoc-mode', retrieves eldoc msg string for inner-most symbol."
  (-> (hy--eldoc-get-inner-symbol)

(defun hy--mode-setup-eldoc ()
  (make-local-variable 'eldoc-documentation-function)
  (setq-local eldoc-documentation-function 'hy-eldoc-documentation-function)
  (eldoc-mode +1))

There are three core components:

  1. hy--eldoc-send for sending a formatted string and cleaning its output.

  2. hy--eldoc-get-inner-symbol getting opening form and completing the dot DSL.

  3. hy--eldoc-fontify-text for highlighting the final text string like in the image.

Lets look at fontifying first. We can't blindly apply Hy's font-locks as the docstring isn't captured in quotes. Since the text is static, we just add the faces to the string ourselves.

(defun hy--fontify-text (text regexp &rest faces)
  "Fontify portions of TEXT matching REGEXP with FACES."
  (when text
        (s-matched-positions-all regexp text)
      (-lambda ((beg . end))
        (--each faces
          (add-face-text-property beg end it nil text))))))

(defun hy--eldoc-fontify-text (text)
  "Fontify eldoc TEXT."
  (let ((kwd-rx
         (rx string-start (1+ (not (any space ":"))) ":"))
         (rx symbol-start "&" (1+ word)))
         (rx "`" (1+ (not space)) "`")))
     text kwd-rx 'font-lock-keyword-face)
     text kwargs-rx 'font-lock-type-face)
     text quoted-args-rx 'font-lock-constant-face 'bold-italic))

Next lets see the sending and formatting of the shell's raw eldoc output.

(defun hy--eldoc-send (string)
  "Send STRING for eldoc to internal process returning output."
  (-> string

The string/output formatting are implementation details specific to Hy and so won't be detailed. If we are dealing with an empty string, we return nil rather than the empty string to pass-by parent when clauses.

The meat of Eldoc is in extracting the innermost symbol of the current point

(defun hy--eldoc-get-inner-symbol ()
  "Traverse and inspect innermost sexp and return formatted string for eldoc."
    (-when-let* ((_ (hy-shell-get-process 'internal))
                 (state (syntax-ppss))
                 (start-pos (hy--sexp-inermost-char state))
                 (_ (progn (goto-char start-pos)
                           (not (hy--not-function-form-p))))
                 (function (progn (forward-char)
                                  (thing-at-point 'symbol))))

      ;; Attribute method call (eg. ".format str") needs following sexp
      (if (s-starts-with? "." function)
          (when (ignore-errors (forward-sexp) (forward-char) t)
            (pcase (char-after)
              ;; Can't send just .method to eldoc
              (?\) (setq function nil))
              (?\s (setq function nil))
              (?\C-j (setq function nil))  ; newline

              ;; Dot dsl doesn't work on literals
              (?\[ (concat "list" function))
              (?\{ (concat "dict" function))
              (?\  (concat "str" function))  ; the " is deleted in blog as breaks rainbow.js

              ;; Otherwise complete the dot dsl
              (_ (progn
                   (concat (thing-at-point 'symbol) function)))))

So Eldoc's path is to call hy--eldoc-get-inner-symbol if an internal process is active, syntax-ppss indicates we are within a form, and that the innermost form is a symbol. The completed string is sent off to the internal process we've built up, the output is chomped of quote characters and the prompt and syntax errors (eg. completing "str." while we are still typing) are ignored. The result is fontified and returned by the documentation function.

Spacemacs shift-k documentation lookup

A feature of Spacemacs is typing "K" to perform spacemacs/evil-smart-doc-lookup to get the full documentation of the symbol-at-point in a separate buffer.

Using Eldoc's documentation functions, with slightly different formatting, we already have most of shift-K implemented.

We moved most of hy-eldoc-documentation-function into hy--eldoc-get-docs which distinctly accepts an optional argument for buffer-style rather than eldoc-style formatting.

We then create a mirror of the documentation function as hy--docs-for-thing-at-point. We format the text to account for newlines (newlines from process output are escaped so we must trim one backslash from each newline).

(defun hy--docs-for-thing-at-point ()
  "Mirrors `hy-eldoc-documentation-function' formatted for a buffer, not a msg."
  (-> (thing-at-point 'symbol)
     (hy--eldoc-get-docs t)

(defun hy--format-docs-for-buffer (text)
  "Format raw hydoc TEXT for inserting into hyconda buffer."
  (when text
    (-let [kwarg-newline-regexp
           (rx ","
               (1+ (not (any "," ")")))
               (group-n 1 "\\\n")
               (1+ (not (any "," ")"))))]
      (--> text
         (s-replace "\\n" "\n" it)
         (replace-regexp-in-string kwarg-newline-regexp "newline" it nil t 1)))))

It is interesting how "K" is actually called, I'm not sure if any other function operates quite the same. "K" calls the function bound to "SPC m h h".

(spacemacs/set-leader-keys-for-major-mode 'hy-mode
  "hh" 'hy-describe-thing-at-point)

Lastly we need to create, switch-to, and insert the retrieved docs as hy-describe-thing-at-point.

(defun hy-describe-thing-at-point ()
  "Implement shift-k docs lookup for `spacemacs/evil-smart-doc-lookup'."
  (-when-let* ((text (hy--docs-for-thing-at-point))
               (doc-buffer "*Hyconda*"))
    (with-current-buffer (get-buffer-create doc-buffer)
      (switch-to-buffer-other-window doc-buffer)

      (insert text)
      (goto-char (point-min))

      (insert "------\n")
      (fill-region (point) (point-max))

      ;; Eventually make hyconda-view-minor-mode, atm this is sufficient
      (local-set-key "q" 'quit-window)
      (when (fboundp 'evil-local-set-key)
        (evil-local-set-key 'normal "q" 'quit-window)))))


There are several other features worth discussion like font-locking the shell prompt input and the send-(form/region/buffer)-to-shell that are sizable enough to warrant their own posts later. With this post, the series is caught up to the current featureset. Linting would be the next big problem to attack. I'd also like to integrate ert.

Going without Company and Eldoc has helped me appreciate the value in IDEs. Hy is quickly becoming a pleasant development experience.

My only guidance was source code. I hope this series make the problem more tractable for prospective major mode authors.

comments powered by Disqus