🤖 feat: add Flow Prompting workspace prompts#2844
Conversation
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: acc20746d3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5a80485cd4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
@codex review |
1 similar comment
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 152a70b23e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 514781851b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 62152607ed
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d1e8bf49f3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
Codex Review: Didn't find any major issues. 👍 ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 919e8572e4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6b64caa242
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
Codex Review: Didn't find any major issues. Keep it up! ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 34a56f2347
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5c3d25cdd4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 88dd1f9866
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f9e70a24d8
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Replace the ad-hoc Flow Prompt Agent Scope textarea with Next-heading extraction from the prompt file so each prompt update carries the current Next section, while the system prompt only tells the agent to re-read the full file when the diffs are insufficient. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
Mock the full App entry in AppLoader's auth-path unit tests so the suite doesn't drag react-dnd and ProjectSidebar into a branch that only needs pre-App loading/auth behavior. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
Keep AppLoader's auth-path unit tests from importing the real react-dnd stack so bun's cross-file DOM teardown cannot trip MutationObserver initialization before ProjectSidebar's test file loads. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
Switch the AppLoader auth and ProjectSidebar suites to load their modules after the DOM and drag-and-drop mocks are in place so bun's cross-file cleanup cannot resolve react-dnd before happy-dom exists. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
Re-register the AppLoader auth and ProjectSidebar drag-and-drop module mocks in beforeEach so `mock.restore()` cannot leave later tests resolving the real react-dnd stack before happy-dom exists. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
Keep the mock-before-load fixes in AppLoader auth and ProjectSidebar tests while switching the deferred loads to lint-annotated `require()` calls, which Bun test already uses elsewhere for post-mock module loading. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 57c7cb657e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
ammar-agent
force-pushed
the
flow-prompting-2aqb
branch
from
57c7cb6 to
0f18706
Compare
|
@codex review Rebased onto current |
Keep a failed-fingerprint guard for auto-sent flow prompt revisions so an idle send failure leaves the preview ready for an explicit retry instead of re-dispatching the same synthetic message every poll interval. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0f18706aee
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
@codex review Addressed the failed-flow-prompt retry comment by keeping a failed fingerprint guard for idle auto-send failures. The same revision now stays visible for explicit retry instead of being re-enqueued every refresh tick. |
Keep workspace-scoped agent state unset until the user actually picks an override, so the provider preserves an existing backend agent instead of pushing the default picker value back into metadata on first load. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
|
@codex review Addressed the AgentContext regression by leaving workspace-scoped agent state unset until the user explicitly picks an override. Existing backend agent metadata now survives first load, which keeps Flow Prompting follow-up sends on the intended agent. |
Raise the long end-to-end send-mode integration test timeout from 60s to 90s so slow runners can finish the full context-menu, stream, and background-bash sequence without tripping the global test budget mid-step. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
|
@codex review One more test-only tweak: the long send-mode integration test now gets 90s instead of 60s so slow CI runners can finish the full context-menu + streaming/background-bash path without timing out mid-step. |
Double the local timeout for the relative-path runtime file-editing integration case, which exercises three provider-backed turns and was timing out on busy Linux runners even though the path handling itself eventually succeeds. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
|
@codex review Also raised the local-only timeout for the relative-path runtime file-editing integration case. It performs three provider-backed turns (create/edit/read), and the Linux runner was hitting the 25s local budget even when the path logic eventually succeeded. |
Keep the Flow Prompting unit tests aligned with the new failed-fingerprint guard by passing the extra shouldEmitUpdate argument in the legacy auto-send suppression cases. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
|
@codex review Small follow-up after the failed-fingerprint guard: I updated the older |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1e14a41388
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
src/browser/components/FlowPromptComposerCard/FlowPromptComposerCard.tsx
Outdated
Show resolved
Hide resolved
Strip the injected `Current Next heading` block before parsing Flow Prompt preview bodies so the composer still renders diff/content previews instead of falling back to raw text when a prompt file defines a Next section. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
|
@codex review Fixed the Flow Prompt preview parser so it strips the injected |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 62c72e7f91
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
src/browser/components/FlowPromptComposerCard/FlowPromptComposerCard.tsx
Outdated
Show resolved
Hide resolved
Wrap the injected `Current Next heading` block in a longer outer fence and teach the preview parser to strip any leading fenced Next section generically, so structured Flow Prompt previews still render when the Next heading contains fenced code. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
|
@codex review Follow-up fix: |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8025bc3a29
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Choose an outer fence longer than any backtick run inside the rendered Flow Prompt markdown sections so `Next` previews and full prompt snapshots remain structurally intact even when they contain nested fenced examples. --- _Generated with `mux` • Model: `openai:gpt-5.4` • Thinking: `xhigh` • Cost: `$447.91`_ <!-- mux-attribution: model=openai:gpt-5.4 thinking=xhigh costs=447.91 -->
|
@codex review Switched the Flow Prompt markdown wrappers to dynamic fence lengths based on the content, so |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cb0049bcfd
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const prefix = `${getFlowPromptPathMarkerLine(attachment.flowPromptPath)}\n\nCurrent flow prompt contents:\n\`\`\`md\n`; | ||
| const suffix = "\n```"; |
There was a problem hiding this comment.
The flow-prompt attachment renderer always wraps flowPromptContent with fixed triple-backtick fences, so any prompt that already contains a fenced block like `````mdor even plain ``` will terminate the outer wrapper early. This makes the post-compaction` payload malformed and can distort the prompt snapshot the model sees after compaction, especially for prompt files that include embedded examples. Please derive the outer fence length from the content (like the update-message path already does) before building this block.
Useful? React with 👍 / 👎.
1 participant
Summary
Flow Prompting adds a repo-local
.mux/prompts/<workspace-name>.mdworkflow that Mux can open, monitor, and turn into durable prompt updates while still leaving inline chat available for quick follow-ups. This iteration restores end-of-turn auto-send, replaces the composer-side Agent Scope textarea withNext-heading steering parsed from the markdown prompt file so every update carries the currentNextsection, moves the full-file reread guidance into the conditional system prompt, preserves backend-selected workspace agents when no local override exists so follow-up sends keep the intended agent, keeps failed auto-sent revisions guarded until the user retries or edits the file, fixes the first-save preview race, keeps tiny post-compaction prompt/plan attachments inside budget, prevents stale restored prompt retries from replaying after a newer save is accepted, and keeps the composer accessory compact enough that it does not sprawl vertically at wider chat widths.Background
The goal is to reduce the “wait vs. context-switch” tax of agentic work by moving durable task guidance into a workspace artifact that can stay open in the user’s editor. The implementation also needs to work across local and remote runtimes, avoid duplicate synthetic sends while prompt updates are in flight, preserve the exact prompt-file path in model context, and keep the accessory compact enough that it does not crowd the composer or diverge from the other accessory banners.
Implementation
workspaceFlowPromptServicefor prompt-file lifecycle management, diff/full snapshot generation, session-side baseline tracking, rename/fork handling, backend-owned auto-send preference storage, in-flight fingerprint tracking, failed-fingerprint retry guards, pending-state clearing, queued follow-up refresh bookkeeping so a save that lands during an in-flight read still updates the preview immediately, andNext-heading extraction from the markdown prompt filefilePartsstripping from reused active send optionsNextpanel sourced from the flow prompt file, left-align the collapsed banner copy to match the other accessories, and let the helper text span the full accessory width beneath the header/actions so it wraps cleanly across the containerNextsection alongside prompt diffs/full snapshots, move the “read the full prompt file when diffs are insufficient” rule into the conditional system prompt, and clamp flow-prompt/plan attachment truncation to the available budget so tight compaction windows still preserve a minimal marker plus partial payload instead of dropping the block entirelyreact-dndbefore the focused auth and sidebar suites runNext-heading extraction, failed-revision retry suppression, clear-update visibility, first-save refresh timing, stale retry dispatch ordering, compact accessory controls, attachment-budget edge cases, backend-agent preservation, and the AppLoader/ProjectSidebar unit isolationValidation
bun test src/browser/components/FlowPromptComposerCard/FlowPromptComposerCard.test.tsxbun test src/node/services/workspaceFlowPromptService.test.ts src/node/services/workspaceService.test.ts -t "Flow Prompt"bun test src/common/utils/ui/flowPrompting.test.ts src/node/services/streamContextBuilder.test.ts -t "flow prompt"bun test src/browser/contexts/AgentContext.test.tsxbun test --coverage --coverage-reporter=lcov src/browser/components/AppLoader/AppLoader.auth.test.tsx src/browser/components/ProjectSidebar/ProjectSidebar.test.tsxTEST_INTEGRATION=1 bun x jest tests/ui/chat/sendModeDropdown.test.ts -t "click sends tool-end by default while context menu + keybind dispatch modes remain" --runInBand --silent(3x)TEST_INTEGRATION=1 bun x jest tests/ui/git/status.test.ts -t "git status reflects ahead count" --runInBand --silent(3x)make static-checkRisks
Most regression risk sits in the Flow Prompting preview/update path and the composer accessory layout. The backend changes stay scoped to flow-prompt bookkeeping,
Next-heading extraction, failure guards, agent preservation, and compaction attachments, while the latest UI changes stay inside the accessory’s own layout classes, tooltip-backed action buttons, collapsed-strip alignment, helper-text row, and a narrowly-scoped AppLoader/ProjectSidebar test harness adjustment.Generated with
mux• Model:openai:gpt-5.4• Thinking:xhigh• Cost:$447.91