Skip to content

Introduce a "code view" of transcripts (git blame style) #23

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Draft
btucker wants to merge 92 commits into
base: main
Choose a base branch
from
Draft

Introduce a "code view" of transcripts (git blame style) #23

btucker wants to merge 92 commits into from

Conversation

@btucker
Copy link

@btucker btucker commented Dec 29, 2025 •
edited
Loading

image

When reviewing work aided by coding agents I've often wanted to explore the chronological transcript side-by-side with the code. I find having the provanence easily explorable is valuable to keeping my mental model aligned with the codebase. I've been messing with different UI concepts around this & was inspired by the release of claude-code-transcripts to bring these ideas here.

Here's an example of what this looks like, the complete sessions of building this feature

Or a more reasonably sized session, care of @simonw: https://gistpreview.github.io/?df987f5892de7a02f003944d88211533/code.html

Along the way building this, "we" (Claude Code & I) discovered an infinite loop bug in python-markdown. The discovery & fix are memorialized in this session. Humorously, this lead to the discovery of another bug in this PR if the transcript contains unterminated HTML comments.

In a nutshell, this PR adds a new flag, --code-view that works for local, json, & web generation. Under the hood, this is how it works:

  1. Extract Operations
    - Parses session loglines for Write, Edit, and Bash tool calls
    - Creates FileOperation objects with file path, content/diff, timestamp, and message ID
    - Detects rm commands in Bash calls → OP_DELETE operations

  2. Build Temp Git Repo
    - Creates a temporary git repository
    - Replays operations chronologically as commits:

    • Write → writes file content
    • Edit → applies string replacement (with resync from actual repo if needed)
    • Delete → removes file from repo
      - Each commit stores metadata (tool_id, msg_id, timestamp) in the commit message

  3. Generate Blame Data
    - Runs git blame on each file in the final repo state
    - Groups consecutive lines by commit → BlameRange objects
    - Maps ranges back to the original prompt/message that made the change

  4. Render HTML
    - File tree: Built from files that exist in final repo state (deleted files excluded)
    - Code panel: Uses CodeMirror 6 with custom decorations for blame highlighting
    - Transcript panel: Full conversation with click-to-navigate links
    - Blame ranges link code lines ↔ transcript messages bidirectionally

The UI is far from perfect. I tried to stay true to not introducing any frontend frameworks with the exception of CodeMirror (which is loaded from esm.sh). We did some work on performance optimizations for huge sessions. For smaller sessions, it's much more nimble. For example:

$ uv run claude-code-transcripts json https://gist.githubusercontent.com/simonw/bfe117b6007b9d7dfc5a81e4b2fd3d9a/raw/31e9df7c09c8a10c6fbd257aefa47dfa3f7863e5/3f5f590c-2795-4de2-875a-aa3686d523a1.jsonl --code-view --gist

Fetching session from URL...
Warning: Could not auto-detect GitHub repo. Commit links will be disabled.
Generated page-001.html
Generated /private/var/folders/sl/rhfr008x7s56dc6bsbnwh3qh0000gn/T/claude-session-tmp9n674xpt/index.html (2 prompts, 1 pages)
Generated code.html (7 files)
Output: /private/var/folders/sl/rhfr008x7s56dc6bsbnwh3qh0000gn/T/claude-session-tmp9n674xpt
Creating GitHub gist...
Gist: https://gist.github.com/btucker/c0fdb0e0a763a983a04a5475bb63954e
Preview: https://gistpreview.github.io/?c0fdb0e0a763a983a04a5475bb63954e/index.html

-> See resulting code view

I ran into an issue that gistpreview relies listing all files in the gist. This fails if you exceed the size limit as then every file after you hit the limit is blank. To work around this I adopted an approach of putting data files in a separate gist. These can then be loaded via the raw API & doesn't have the same limitations. This was a pre-existing issue, but exacerbated by the code view. Search is also updated to make use of these data files instead of the HTML if available. This ended up being a large part of the PR.

Also added are e2e tests with python-playwright. This was very helpful as the complexity of the code view UI increased.

This is obviously a huge, wide-ranging PR and no hard feelings if you do not want to accept it into the main codebase. It could also be reasonably split up into 3 changes:

  1. fixing content escaping/sanitization bugs with nh3
  2. adding support for json to be a url (already added in main)
  3. supporting large sessions with gistpreview
  4. adding code view UI

btucker and others added 30 commits December 27, 2025 10:23
@btucker @claude
Add code viewer with git blame-style annotations
58d542b 
Add a --code-view flag that generates a Code tab showing files modified
during the session:

- Tree view of modified files on the left panel
- CodeMirror editor on the right with syntax highlighting
- Git blame-style gutter showing which transcript message wrote each line
- Click blame annotation to jump to the transcript message
- Tab navigation between Transcript and Code views

Supports both local git repos and public GitHub URLs:
- Local repos: uses git show HEAD:filepath
- GitHub URLs: fetches via GitHub API

Key additions:
- FileOperation and FileState dataclasses for tracking file changes
- extract_file_operations() to parse Write/Edit tool calls
- get_initial_file_content() for git/GitHub integration
- reconstruct_file_with_blame() for blame attribution algorithm
- CodeMirror 6 integration with custom blame gutter extension
- Tab bar in index.html and page.html templates

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix JSON escaping in code viewer to prevent script injection
1959dec 
Escape `</` and `<!--` sequences in embedded JSON to prevent the browser
from prematurely terminating the <script> block when file contents
contain these HTML-significant patterns.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Simplify code viewer to diff-only mode
d076aaf 
Remove git file fetching functionality that was causing performance issues.
The code viewer now shows:
- For new files (first op is Write): full content with blame attribution
- For existing files (first op is Edit): diff-only view showing changes

Removed:
- get_initial_file_content, _get_file_from_local_git, _get_file_from_github
- session_cwd extraction (no longer needed)
- mode parameter from generate_code_view_html
- Debug print statements
- Related tests for git fetching

Full file reconstruction from git can be added as a future enhancement.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add progress indicator for large page rendering
a27d922 
Shows "Page X/Y: rendering message N/M..." when a page has more than 50
messages. Uses carriage return to update in place and clears when done.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add three-pane code viewer with git-based blame and full transcript
53cfbf7 
Major redesign of the code view:
- Three resizable panes: file tree, code editor, and transcript
- Uses temporary git repo to track file changes for accurate blame
- Click code lines to scroll transcript to the relevant message
- Full transcript in right pane (same as Transcript tab)
- Draggable separators between all panels

UI improvements:
- Tab styling connects to header border (active tab "emerges" from content)
- File status indicators (green dot for added, orange for modified)
- Header stays full-width on both tabs for stable navigation
- Transcript content centered at 800px max-width

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Simplify file detection with per-file git repo lookup
1f1c9ae 
Replace the global source_repo_path parameter with automatic per-file
git repository detection. For Edit operations on files not yet in the
temp repo, we now walk up from each file's directory to find its git
repo and fetch initial content from HEAD. Falls back to reading from
disk if no git repo is found.

This handles cases where files span multiple git repositories and
removes the need for the --repo CLI flag when files already exist
in git repos on disk.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add chunked rendering for transcript pane performance
4e78390 
For large transcripts, having all messages in the DOM at once can cause
slow renders and high memory usage. This adds chunked rendering:

- Store messages as a JS array instead of embedding all HTML in DOM
- Render first 50 messages initially
- Use IntersectionObserver to load more as user scrolls down
- scrollToMessage() ensures target message is rendered before scrolling
- Build ID-to-index map for fast message lookup

This keeps the initial DOM small while still supporting full transcript
navigation when clicking on blame sections in the code viewer.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Skip blame highlighting for pre-existing content
aecc019 
Lines that existed before the session (no msg_id) now have a plain
white background instead of being colored. Only lines that were
actually written or edited during the session get blame colors.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add overflow hidden to index item containers
d2404e8 
Defensive CSS to prevent markdown content from breaking out of
containers if it contains malformed HTML.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix truncation for dynamically rendered messages
c2cfa01 
The chunked rendering broke message truncation because the truncation
initialization only ran on page load. Now we initialize truncation for
newly rendered messages after each chunk is added to the DOM.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix blame highlighting to only color actually changed lines
c571051 
When fetching initial content for Edit operations, we now commit it
first (with empty metadata) before applying the edit. This allows git
blame to correctly attribute unchanged lines to the initial commit
(which has no msg_id) and only the actually changed lines to the edit
commit.

Previously, all lines in an edited file were attributed to the first
edit operation, even lines that weren't changed.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add tests for git-based blame attribution
183c9fa 
Tests verify that:
- Write operations attribute all lines to that operation
- Edit operations only attribute changed lines, not context
- Multiple edits to the same file are tracked separately

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add scrollbar minimap showing blame range locations
1592307 
A thin vertical strip next to the editor shows colored markers for each
blame range, making it easy to see where changes are in long files.
Clicking a marker scrolls to that location and highlights it.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker
currently non-functional codemirror state
ff535c6
@btucker @claude
Extract code_view.py module from __init__.py
617ae5e 
Refactors the code viewer functionality into a separate module:
- Moves ~1000 lines of code viewer logic to code_view.py
- Reduces __init__.py from ~3100 to ~2433 lines
- Includes FileOperation, FileState, CodeViewData, BlameRange dataclasses
- Includes all git-based blame attribution functions
- Removes obsolete escape_json_for_script test (fixed in python-markdown)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Apply DRY refactorings: CSS template, helpers for grouping and colors
13d5516 
1. Extract CSS to templates/styles.css (~260 lines)
   - Improves maintainability with proper CSS formatting
   - Allows future CSS tooling integration

2. Add group_operations_by_file() helper in code_view.py
   - Consolidates duplicate file grouping logic
   - Used by build_file_states() and generate_code_view_html()

3. Add buildRangeColorMap() helper in code_view.js
   - Pre-computes color assignments for blame ranges
   - Ensures consistent colors between editor decorations and minimap
   - Eliminates duplicate color tracking logic

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix JSON escaping for script tag embedding
e44272e 
Escape </ sequences in JSON data embedded in <script> tags to prevent
premature script tag closing when file content contains </script>.

Adds test to verify the escaping works correctly.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker
resolved issue when reconstructing file state from edit operations, i…
a01b324 
…f any edit fails or is malformed, our reconstruction diverges from reality. Subsequent edits also fail because their old_string doesn't match our corrupted state.
@btucker
include the prompt number in the code view
97e5a00
@btucker
handling continuation messages properly
5e2aa58
@btucker
update readme
11882a0
@btucker @claude
Fix blame highlighting and strip common prefix from file tree
e860505 
- Strip common top-level directories from file tree paths for cleaner display
- Fix blame highlighting not appearing by using StateField instead of ViewPlugin
- Remove final sync step that was destroying blame attribution
- Add tests for common prefix stripping behavior

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Handle empty git repo in get_file_content_from_repo
e0a280c 
Add ValueError to the exception handler to gracefully handle cases
where the temporary git repo has no commits yet.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Extract originalFile from tool results for remote session support
ca2bc04 
When viewing sessions from remote URLs (like gists), the local filesystem
isn't available to read initial file content before edits. This change
extracts the originalFile field from toolUseResult in JSONL sessions and
uses it as a fallback when reconstructing file state.

- Add original_content field to FileOperation dataclass
- Extract originalFile from toolUseResult entries in extract_file_operations()
- Use original_content as fallback in build_file_history_repo() when file
  doesn't exist locally
- Add test coverage for originalFile extraction

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix JSONL parsing to preserve toolUseResult for remote sessions
a86705f 
The JSONL parser was stripping out the toolUseResult field, which contains
the originalFile content needed for code reconstruction in remote sessions.
This caused code.html to have empty fileData when processing JSONL URLs.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add test for JSONL parsing preserving toolUseResult
7ebad63 
Ensures the toolUseResult field (containing originalFile content) is
preserved when parsing JSONL files, which is needed for remote session
code reconstruction.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Treat skill expansions (isMeta) as continuations for prompt numbering
75b01da 
Skill expansion messages have isMeta=True in JSONL sessions. These are
system-injected, not user-typed prompts. They should be treated as
continuations of the previous conversation for prompt numbering purposes.

This fixes the code viewer showing all blame as "prompt simonw#2" when a skill
was invoked - now all operations are correctly attributed to prompt #1
(the original user request to invoke the skill).

- Preserve isMeta field in JSONL parsing
- Treat isMeta entries like isCompactSummary (continuations)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add test for isMeta field preservation in JSONL parsing
bf60b03 
🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Assign blame colors by prompt number, not by operation
75c7035 
Previously each Write/Edit operation got a unique color based on its
msg_id. Now all operations from the same prompt get the same color,
making it easier to visually identify which prompt caused which changes.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Show assistant context in code viewer tooltips
f0bf40d 
Replace the stats/summary in tooltips with the assistant text that
immediately preceded the tool call. This provides more useful context
about what Claude was thinking when making each change.

- Capture assistant text blocks before tool_use
- Display as "Assistant context" in tooltip
- Truncate long assistant text to 500 chars
- Add CSS styling for tooltip-assistant section

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
btucker and others added 22 commits December 29, 2025 14:50
@btucker @claude
Add batched gist uploads and descriptive gist names
8c39fc0 
- Use gh gist edit --add to add files to same gist in batches instead
  of creating multiple separate gists
- Add descriptive gist names like "claude-code-transcripts local session-name"
- Simplify code by removing DATA_GIST_IDS handling since we only have
  one data gist now

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add index-data.json for large session index content
940f423 
- Generate index-data.json with index items HTML for large sessions
- Update index.html template to load from JSON on gistpreview
- Strip inline content from index.html when uploading to gist
- Increase batch size limits (10MB, 100 files) for adding files to gists
- Add tests for index-data.json generation

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Use templates to exclude content from HTML in large sessions
df4e7dc 
Instead of stripping content with regex after rendering, modify the
templates to conditionally exclude content when use_page_data_json or
use_index_data_json is true. This is cleaner and more reliable.

Also update tests to expect the new create + edit gist pattern.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Include content in HTML for local viewing, strip for gist upload
151361f 
Large sessions now always include content in the generated HTML files,
allowing local viewing without JavaScript. Content is stripped from
HTML during inject_gist_preview_js when uploading to gist with a
data gist ID, keeping gist files small.

- Always pass use_*_data_json=False to templates (include content)
- Add _strip_container_content helper for HTML content stripping
- Strip content and inject loader JS in inject_gist_preview_js only
- Move PAGE_DATA_LOADER_JS and INDEX_DATA_LOADER_JS to Python constants
- Templates stay clean - loader JS only added during gist upload
- Add test for large sessions including content in local HTML

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix gist upload: add files one at a time instead of batching
0126082 
The gh gist edit --add command with multiple --add flags was silently
failing to add some files. Change to add files one at a time to ensure
all files are uploaded reliably.

Also add verbose logging to show which files are being uploaded.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add retry logic and delays for gist file uploads
2622e65 
GitHub's API returns HTTP 409 conflicts when updating gists too rapidly.
Add retry logic with exponential backoff (2s, 4s, 6s) for 409 errors,
plus a 0.5s delay between file additions to avoid rate limiting.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Remove vestigial batching code from gist upload
628d818 
Since files are now added one at a time with delays, the batch grouping
logic was no longer needed and just added confusing output. Removed
_batch_files_for_gist and related constants.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker
Because of limitations in gistpreview we have to do some backflips to…
02ea004 
… get styles/js to work
@btucker
we don't need the fancy css/js loading if it's a small session on gis…
01a2879 
…tpreview
@btucker @claude
Fix code view performance: don't auto-scroll transcript
14f31e3 
For large transcripts (12K+ messages), auto-scrolling to messages
would render thousands of DOM nodes, causing the browser to hang
for several seconds on file switches and blame clicks.

Changes:
- Remove auto-scroll to transcript on file load (just highlight code)
- Remove auto-scroll on blame click (tooltip shows message on hover)
- Remove auto-scroll in navigateToBlame (user already sees transcript)
- Keep progressive rendering infrastructure for future use
- Update tests to match new behavior

The tooltip on hover still shows the full message content, and users
can manually scroll the transcript if needed. This keeps the UI
responsive even for very large transcripts.

Performance improvement:
- Before: 5+ second hang when switching files or clicking blame
- After: Instant response (<50ms)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Pre-compute prompt_num and color_index server-side for blame ranges
8034b65 
Move computation of user prompt numbers and color indices from the
client-side JavaScript to the server-side Python during git repo build.
This eliminates client-side walking of data for each blame block.

Changes:
- build_msg_to_user_html now returns 3 values including msg_to_prompt_num
- generate_code_view_html computes color_index per unique context_msg_id
- Each blame range now includes prompt_num and color_index fields
- Simplified buildRangeMaps JS to use pre-computed values directly

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix clicking blame blocks to scroll transcript to corresponding message
6f25b95 
Restored the call to scrollToMessage() in the blame click handler.
This was previously removed for performance but is needed for proper
navigation between code and transcript views.

Added e2e test to verify clicking a blame block scrolls to and highlights
the corresponding message in the transcript panel.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Implement windowed transcript rendering with teleportation
66f5803 
Instead of rendering all messages between the current view and the target
when navigating to a distant blame block, the transcript now "teleports"
to the target location by clearing the DOM and rebuilding from the user
prompt containing the target message.

Key changes:
- Track windowStart/windowEnd instead of just renderedCount
- Add findUserPromptIndex() to locate conversation boundaries
- Add teleportToMessage() to jump to distant messages instantly
- Add bi-directional lazy loading (scroll up loads earlier messages)
- Add top sentinel for upward IntersectionObserver
- Move bottom sentinel inside transcript-content div

This keeps DOM size small regardless of where in the transcript the
user navigates, while still allowing smooth scrolling in both directions.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add prompt numbers to user messages and improve pinned header behavior
d1e51f5 
Changes:
1. Auto-select first blame block on file load - now that teleportation
   makes transcript navigation instant, re-enable auto-scrolling

2. Add prompt # to user message headers - shows "User simonw#1", "User simonw#2" etc.
   for each user prompt (excluding tool result messages)

3. Add prompt # to pinned user message - shows "#N message text..."

4. Fix pinned user prompt overlap - hide the pinned header when the
   next user message would overlap with it, rather than showing both

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix pinned user message flashing when overlapping next prompt
8c5f75b 
The pinned message was flashing because hiding it changed offsetHeight
to 0, which shifted the threshold calculations and caused unstable
show/hide behavior.

Fix by caching the pinned message height and using the cached value
when the element is hidden. Also simplified the threshold logic to use
a single pinnedAreaBottom value for both determining what to pin and
detecting overlap.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Move prompt number to pinned message label
d08694e 
Changed pinned user message format from:
  Label: "User prompt:"
  Content: "#1 message text..."

To:
  Label: "User Prompt #1"
  Content: "message text..."

This puts the prompt number in the header where it's more prominent.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix teleport to always render up to target message
f7f93d9 
When teleporting to a distant message, the initial chunk was only
rendering CHUNK_SIZE messages from the user prompt. If the target
message (e.g., an assistant edit) was beyond that chunk, it wouldn't
be in the DOM and the scroll would fail or scroll to the wrong place.

Fixed by ensuring initialEnd is at least the target message index,
so the target is always rendered after teleporting.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix URL fragment navigation and use bright yellow highlight
e35fe6c 
- Fix URL fragment handling to scroll code and select blame block on load
- Call scrollToMessage when navigating from hash to update transcript
- Use setTimeout to wait for editor initialization when loading new file
- Change selected blame highlight color from blue to bright yellow

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Use consistent yellow highlight and skip initial blame highlight
da3972a 
- Use bright yellow highlight color for transcript message border
  (matches the blame block highlight color)
- Skip blame highlighting on initial file load, but still scroll
  the editor and transcript to the first blame
- Fix navigateToBlame timing when loading a different file

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Add progress output for code view generation
b7b39b1 
Shows progress for:
- Replaying file operations (when > 20 operations)
- Processing files for blame data (when > 5 files)

Uses carriage return for in-place updates, similar to page rendering.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Clear progress line when switching phases
ef27068 
Fixes leftover characters when switching from operations to files phase.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Change user message heading to "User Prompt #X"
ea2d6e6 
Adds "Prompt" to match the pinned message label format.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker btucker marked this pull request as draft December 30, 2025 17:12
btucker and others added 6 commits December 30, 2025 12:31
@btucker @claude
Fix prompt number mismatch between tooltip and pinned header
5aba4e2 
The pinned user prompt header was showing a different prompt number than
the tooltip when clicking blame blocks. This was caused by inconsistent
counting between server-side (tooltip) and client-side (pinned header):

- Server-side build_msg_to_user_html() was skipping continuation
  conversations, but the HTML renderer included them in the count
- Client-side getPromptNumber() was excluding messages with
  class="continuation" but this didn't match the server logic

Fix:
- Remove continuation skip in build_msg_to_user_html()
- Update _collect_conversation_messages() to only return current conv
- Update client-side getPromptNumber() to include continuation messages
- Change tooltip heading from "#X" to "User Prompt #X" for consistency

Also fixes:
- Race condition when loading file from URL hash (check hash first)
- CSS selector bug in updatePinnedUserMessage()
- Ensure user prompt is loaded when teleporting to distant messages

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Fix prompt number counting and pinned click after teleportation
e0978eb 
Two issues fixed:

1. Prompt number mismatch: Client-side getPromptNumber was counting ALL
   user messages including tool_results, but server only counts real
   prompts (messages with "User Prompt #N" label). Fixed by checking
   for ">User Prompt #" in message HTML instead of just class="message user".

2. Pinned click not working: After teleportation, the onclick handler
   held a stale reference to a DOM element that was removed. Fixed by
   storing the message ID and looking up the element on click, with
   fallback to scrollToMessage if element is not in DOM.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Use server-provided prompt_num instead of client-side counting
5114ad8 
Instead of computing prompt numbers client-side (which was error-prone
with off-by-one issues), include the canonical prompt_num in the
messagesData JSON from the server.

The server already computes prompt numbers correctly when rendering
HTML, so we just add that to the message data payload. Client-side
getPromptNumber() now reads directly from messagesData[i].prompt_num.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker
Compute all user prompt numbering at build
b73dcf3
@btucker @claude
Code quality improvements to code_view.py
607ce99 
Based on code quality review, addressed three issues:

1. Remove duplicate imports - removed redundant `import re` (appeared
   twice inside functions) and `import json` (inside function, already
   at module level)

2. Add proper is_recursive field to FileOperation instead of reusing
   replace_all field for delete operations. This was a code smell where
   a boolean meant for "replace all occurrences" was being used for
   "is recursive delete".

3. Consolidate read_blob_content and read_blob_bytes into single
   read_blob(decode=True/False) function, keeping backwards-compatible
   aliases for the original function names.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker @claude
Merge upstream/main and adopt URL handling
cb3fe47 
Merged upstream's JSON URL functionality, removing our duplicate
implementation of is_url and fetch_session_from_url in favor of
upstream's is_url and fetch_url_to_tempfile.

Kept our code-view options (--code-view, --exclude-deleted-files)
while adopting upstream's URL handling logic and variable naming.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@btucker
Copy link
Author

btucker commented Dec 31, 2025 •
edited
Loading

Quick update that I got inspired to spin off all the git pieces of this into it's own project: https://github.com/btucker/agentgit

Still very much a WIP, but I think there's potential. I'm going to keep this PR open as a draft for the moment, but come back and slim it down using agentgit as a library

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@btucker