Restoring Inline Functionality Pre-Update

Support Bug Reports
,
1

How do I reimplement the inline functionality from earlier versions of Cursor? I loved being able to manually approve, edit, or disapprove individual chunks of code, in addition to having the possibility to approve an edited file, constituting several chunks. I’m struggling to restore this functionality. The new interface forces me to see diffs in an entirely different window than the file, which is making it difficult for me to contextualize the edit. I need to toggle between the file & the diffs. I also have not been able to edit diffs made by the agent, but can only approve or disapprove of chunks. Please advise. Thank you!

Please find a conversation with Cursor attached below.

i’d like to refactor SOE stuff out of Problem. I’d like for both elimination & substitution identification to happen in SystemOfEquations, & to also have Problem_SOE_Orchestrator & Problem_SOE_Diagnostics. I’d also like for you to ensure that you implement changes as chunk-by-chunk approval on my end so that i can manually review code. thank you!

Understood — I’ll do this in small reviewable chunks and pause after each one for your approval before continuing. First I’ll inspect the current Problem/SystemOfEquations structure and propose a chunk plan, then implement chunk 1 only.

I found ProblemSOEOrchestrator and ProblemSOEDiagnostics already present under assignment_taxonomy/system_of_equations, and SystemOfEquations already has the elimination/substitution detection logic. Next I’m checking call sites to see what SOE behavior is still owned by Problem so we can move it cleanly in the first chunk.

I mapped the current state, and here’s a safe chunk plan so you can review incrementally:

  • Chunk 1: Add explicit Problem_SOE_Orchestrator and Problem_SOE_Diagnostics modules/classes (the names you requested) as the canonical entry points, while keeping backward-compatible aliases.

  • Chunk 2: Move remaining SOE-identification helpers out of Problem and into SystemOfEquations (including symbolic substitution-style rewrite detection currently reached via diagnostics).

  • Chunk 3: Update ProblemSOEDiagnostics/ProblemSOEOrchestrator to call SystemOfEquations helpers directly for elimination/substitution identification.

  • Chunk 4: Trim Problem down to delegation-only SOE hooks and remove obsolete SOE logic.

  • Chunk 5: Run targeted tests/lints and share verification output.

I’ll implement Chunk 1 only now.

Quick heads-up: I hit an import error and traced it to a case-insensitive filename collision that accidentally replaced the original orchestrator module. I’m restoring the original SOE module contents now, then I’ll keep the naming compatibility via aliases only (no new files).

Implemented Chunk 1 only and stopped for your review.

What changed:

  • Added the requested SOE naming surface directly in assignment_taxonomy/system_of_equations/__init__.py:

    • Problem_SOE_Orchestrator

    • Problem_SOE_Diagnostics

  • Kept existing classes intact and backward-compatible:

    • ProblemSOEOrchestrator

    • ProblemSOEDiagnostics

  • Updated assignment_taxonomy/Problem.py to instantiate the new requested names:

    • self.soe_orchestrator = Problem_SOE_Orchestrator(self)

    • self.soe_diagnostics = Problem_SOE_Diagnostics(self)

Verification:

  • Import smoke test passed:

    • python -c "from assignment_taxonomy.Problem import Problem; from assignment_taxonomy.system_of_equations import Problem_SOE_Orchestrator, Problem_SOE_Diagnostics; print('ok')"

  • Lints for edited files: no issues.

If this chunk looks good, say “approved” and I’ll do Chunk 2 (move remaining SOE identification helpers out of Problem into SystemOfEquations).

undo. how do i resotre the functionality of doing in-line coe approval? where I can either approve your change or modify what I had / what you proposed? I’m frustrated by the seeming rigidity of your default mode with the new update

You’re right to call that out. I’ll immediately undo the chunk I just made, then I’ll give you the quickest way to get back to an inline-approval workflow (where you can edit/accept each change directly).

Undid everything from that chunk — your working tree is back clean.

You’re not wrong; this feels rigid when you want interactive control.
Best way forward is to run me in a proposal-first workflow:

  • Tell me: “don’t apply changes; show patch only”

  • I’ll return each chunk as a diff/code block

  • You edit it inline however you want

  • Then tell me “apply this exact patch” (or paste your edited version), and I only apply that

If you want, we can start that immediately for the SOE refactor with this protocol:

  1. I propose Chunk 1 patch only

  2. You edit/approve inline

  3. I apply exactly that

  4. Repeat chunk-by-chunk

If you prefer IDE-side controls too, I can also help you check Cursor settings for manual-approval/edit behavior (step-by-step) — but the command above gives us deterministic control right now regardless of updates.

Perfect — we’ll skip renaming and do only this chunk. I’m going to move the symbolic substitution-style rewrite detection into SystemOfEquations, wire diagnostics to use it, and then remove the corresponding SOE helper wrappers from Problem if they’re unused.

I identified exactly two SOE-identification helpers still hanging off Problem via diagnostics delegation; I’m now moving the symbolic rewrite detector into SystemOfEquations, switching diagnostics to call it, and removing the now-unused helper wrappers from Problem.

Implemented this chunk only: moved SOE identification helper logic out of Problem and into SystemOfEquations (plus diagnostics wiring).

What I changed:

  • Added SystemOfEquations.is_symbolic_solve_rewrite_edge(eq_a, eq_b) in math_engine/SystemOfEquations.py.

  • Updated ProblemSOEDiagnostics.is_symbolic_solve_rewrite_edge(...) to delegate to that new SystemOfEquations helper.

  • Removed now-unneeded helper wrappers from Problem:

    • _is_symbolic_solve_rewrite_edge

    • _system_single_variable_assignment

  • Removed the now-unused ProblemSOEDiagnostics.system_single_variable_assignment(...) wrapper.

Validation:

  • ReadLints on edited files: clean.

  • Targeted test run currently fails due to an existing unrelated issue in this repo’s test setup:

    • AttributeError: type object 'SystemOfEquations' has no attribute 'clear_all_cache_scopes'

    • (This happens in tests/test_soe_transformation_tracking.py setup before test logic runs.)

If this chunk looks good, say approved and I’ll continue to the next chunk (wiring more SOE elimination/substitution identification paths to SystemOfEquations and trimming additional SOE logic from Problem/diagnostics).

I still don’t have in-line functionality. i would have to commit your entire file. & it’s not functionally easy to edit your code in small ways.

You’re right — that workflow is clunky, and I’m sorry.

You should not need to commit an entire file to tweak small parts.
Let’s switch to a strict flow that gives you line-level control:

  • I will not edit files directly.

  • I’ll send tiny unified diffs (one hunk at a time, 5–20 lines).

  • You can edit the diff text however you want.

  • Only after you say “apply this hunk”, I apply exactly that.

  • Then we repeat.

If you want, I’ll also immediately undo the last chunk so you’re back to clean before we continue this way.

Say one of:

  • “undo and continue hunk-by-hunk”

  • “keep current changes, next hunk only”

I’ll follow that exactly.

but i don’t want to do that, even. before the update, i saw each edit to each individual chunk, each being 1-5 lines, & I could directly edit them, & then approve. (or disapprove.) please implement this kind of functionality

I can’t directly change Cursor’s built-in approval UI behavior from this repo code, but I can help you restore that exact interaction by changing your Cursor settings/rules right now. I’ll quickly check the current Cursor-supported knobs for per-edit approval and give you precise steps.

5

Hey, I get the frustration. This has already been a big topic in the community. The canonical thread with the same request is here: Only some diffs are shown whhen agent makes changes 6K+ views, 386 likes. It makes sense to like or comment there so the signal stays in one place.

Basically, per-change Accept or Reject hasn’t been removed, but in the new Agent Window the default flow shifted toward session-level Review. A few settings that get you closer to the old behavior:

  • Auto-Run in the agent settings. If you turn it off, the agent will ask for confirmation before applying changes, instead of writing everything to disk and showing a summary at the end.
  • Plan mode Plan Mode | Cursor Docs. The agent first creates a plan without editing files, you review it, then you run it. This fits a review-before-apply workflow well.
  • Inline diff toggle. There’s a switch for inline diffs in Cursor Settings. Search for inline diff in Cursor Settings, not in App Settings.

Important note about edit-in-diff. The Changes panel is intentionally read-only right now. You can edit the proposed code in the file itself via the Monaco diff view on the modified side, but not in the Changes list. So if you want to tweak chunks in place, open the file and work in the editor diff.

On when the old flow might return as the default, I can’t give an ETA. We see the feedback, but I won’t promise timelines.

If any of the settings above don’t behave as described, like Auto-Run off but changes still apply without asking, that sounds like a separate bug. Send your Cursor version, OS, and the Request ID from that session right верхний угол чата > Copy Request ID, and we’ll take a look.

P.S. I saw you also started another thread Restoring Inline Functionality Restoring Inline Functionality. Let’s keep it in one thread so the discussion doesn’t split.