# Add agent-shell-quote-reply and DWIM context prefill for queue-compose#12
Draft
timvisher-dd wants to merge 4 commits into
Draft
# Add agent-shell-quote-reply and DWIM context prefill for queue-compose#12timvisher-dd wants to merge 4 commits into
timvisher-dd wants to merge 4 commits into
Conversation
`agent-shell-quote-reply' pops the queue-compose buffer prefilled with the agent's last prose reply wrapped as a GFM block quote, and `agent-shell-queue-request' now gathers any DWIM context available in the source buffer (active region, error at point, dired marked files) and prefills the compose buffer with it. Regions from buffers with no file association are wrapped as GFM block quotes; file-visiting and dired context keep their existing purpose-built decoration. Cannot use `shell-maker--command-and-response-at-point' to find the last reply because `shell-maker--extract-history' returns the entire rendered turn span, including tool-call drawers and thought blocks. `agent-shell--last-agent-message-text' walks back to the most recent `agent_message_chunk' fragment and pulls the raw body from `agent-shell-ui--content-store' instead. `text-property-search-backward' returns nil when its predicate rejects the nearest match rather than continuing the walk, so the implementation loops over property changes and applies the suffix check manually — otherwise a tool-call fragment between point-max and the agent message silently swallows the result. Bound to `q' in the `agent-shell-help-menu' "Insert" group. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
agent-shell--insert-to-shell-buffer parked point at the start of the inserted text — the DWIM "land on your context" behavior — for every insertion, including the :submit t path used by agent-shell--process-pending-request and interactive C-c C-c on a queued request. After submit, that text becomes transcript above the new prompt, so point was left at the end of the last agent message. shell-maker only auto-scrolls streamed output while point is at point-max, so the cursor stayed stranded above the prompt for the whole turn. Leave point at point-max on the submit path so it tracks the prompt; keep the start-of-context placement only for non-submit insertions. Add a regression test asserting point is at end-of-buffer when shell-maker-submit fires (fails without the fix). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The interactive submit path (typing at the shell prompt and pressing RET) stranded point above the streamed reply for the whole turn. shell-maker-submit skips its own goto-char when shell-maker-point-at-last-prompt-p is true, so after comint-send-input point sits at the response boundary just below the echoed input — not at point-max. The new prompt and the streamed reply land below point, and shell-maker only auto-scrolls while point is at end-of-buffer, so the cursor stayed above the reply until the turn finished. 3ac141c fixed only the programmatic :submit t path (goto point-max before shell-maker-submit); the interactive RET path was never covered. The bug reproduces with or without agent-shell--with-preserved-point, so fix it at the submission entry point: agent-shell--handle (the :execute-command callback) now snaps point to point-max whenever a command is submitted, so the cursor follows streaming output down to the new prompt. A user who scrolls up mid-stream is still honored. Verified live against a real Claude turn: without the fix point sat 4 lines above point-max after the turn; with it point lands at point-max. Add two regression tests asserting agent-shell--handle moves point to point-max for a real command and leaves it alone for a nil (bootstrap) command. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…lected The queue-compose C-c C-c path submits with :no-focus t, so the shell window is never selected. The submit-time point fixes (3ac141c, 531fcc4) moved only buffer point via goto-char, which does not update the window-point of a non-selected window. The shell window therefore kept the position the user had parked point at before invoking agent-shell-queue-request, and the visible cursor snapped back there for the whole turn. with-preserved-point cannot cover this — it manages buffer point only. agent-shell--handle now also pushes window-point to point-max for every window showing the shell, mirroring the set-window-point pattern in agent-shell-jump-to-latest-permission-button-row. Add a regression test that displays the shell in an unselected window, parks its window-point above the prompt, and asserts submit moves window-point to point-max (fails without the fix). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Two related compose-buffer enhancements:
agent-shell-quote-reply(new) pops the queue-compose bufferprefilled with the agent's last prose reply, wrapped as a GFM
block quote. Bound to
qin theagent-shell-help-menu"Insert" group.
agent-shell-queue-request, when called interactively withouta prompt, now gathers any DWIM context available in the source
buffer (active region, error at point, dired marked files) and
prefills the compose buffer with it. Regions from buffers with
no file association are wrapped as GFM block quotes; file-
visiting and dired context keep their existing purpose-built
decoration.
Why not just use
shell-maker--command-and-response-at-point?It returns the entire rendered turn span —
shell-maker--extract- history' takes abuffer-substring' between`' and the next prompt, so tool-call
drawers and thought blocks are included. For a quote-reply that's
not what you want.
agent-shell--last-agent-message-text' walks back frompoint-max' to the most recent fragment whoseagent-shell-ui-state' qualified-id ends in-agent_message_chunk' and pulls the raw body from`agent-shell-ui--content-store'. Tool-call and thought fragments
are skipped by the suffix check.
`text-property-search-backward' gotcha
When its predicate rejects the nearest match, it returns nil
rather than continuing the walk. The first implementation used
the built-in predicate path and silently returned nil whenever a
tool-call fragment sat between point-max and the agent message.
The fix loops over property changes manually and applies the
suffix check each step — the regression test
(`agent-shell--last-agent-message-text-skips-tool-calls') fails
without the loop.
Test plan
tests/agent-shell-tests.el:nested code fence, already-quoted input)
skips tool calls, picks last of many, nil when empty)
no reply, pops compose with quoted reply)
agent-shell-queue-compose-pop':initial-content' (insertinto fresh buffer, append to existing draft with blank-line
separator, no-op when omitted)
agent-shell--compose-initial-content' (quotes non-file region, delegates to--context' otherwise)agent-shell-queue-request-without-prompt-pops- compose' for the new&key' signature.bin/test: 259/260 passing, 1 pre-existing skip,all CI checks green.
when prepping the PR). Worth a manual run that produces a
reply with tool calls in the same turn, then exercising
`agent-shell-quote-reply' to confirm the quoted output
contains only the prose.
Refs
agent-shell-wafa