diff --git a/Pantograph/Goal.lean b/Pantograph/Goal.lean
index e190d5d..52562e7 100644
--- a/Pantograph/Goal.lean
+++ b/Pantograph/Goal.lean
@@ -177,16 +177,51 @@ protected def GoalState.getMVarEAssignment (goalState: GoalState) (mvarId: MVarI
 
 --- Tactic execution functions ---
 
-protected def GoalState.step (state: GoalState) (goal: MVarId) (tacticM: Elab.Tactic.TacticM Unit)
+-- Mimics `Elab.Term.logUnassignedUsingErrorInfos`
+private def collectAllErroredMVars (src : MVarId) : Elab.TermElabM (List MVarId) := do
+  -- These descendants serve as "seed" mvars. If a MVarError's mvar is related
+  -- to one of these seed mvars, it means an error has occurred when a tactic
+  -- was executing on `src`. `evalTactic`, will not capture these mvars, so we
+  -- need to manually find them and save them into the goal list.
+  let descendants ←  Meta.getMVars $ ← instantiateMVars (.mvar src)
+  --let _ ← Elab.Term.logUnassignedUsingErrorInfos descendants
+  let mut alreadyVisited : MVarIdSet := {}
+  let mut result : MVarIdSet := {}
+  for { mvarId, .. } in (← get).mvarErrorInfos do
+    unless alreadyVisited.contains mvarId do
+      alreadyVisited := alreadyVisited.insert mvarId
+      /- The metavariable `mvarErrorInfo.mvarId` may have been assigned or
+         delayed assigned to another metavariable that is unassigned. -/
+      let mvarDeps ← Meta.getMVars (.mvar mvarId)
+      if mvarDeps.any descendants.contains then do
+        result := mvarDeps.foldl (·.insert ·) result
+  return result.toList
+
+private def mergeMVarLists (li1 li2 : List MVarId) : List MVarId :=
+  let li2' := li2.filter (¬ li1.contains ·)
+  li1 ++ li2'
+
+/--
+Set `guardMVarErrors` to true to capture mvar errors. Lean will not
+automatically collect mvars from text tactics (vide
+`test_tactic_failure_synthesize_placeholder`)
+-/
+protected def GoalState.step (state: GoalState) (goal: MVarId) (tacticM: Elab.Tactic.TacticM Unit) (guardMVarErrors : Bool := false)
   : Elab.TermElabM GoalState := do
   unless (← getMCtx).decls.contains goal do
     throwError s!"Goal is not in context: {goal.name}"
   goal.checkNotAssigned `GoalState.step
-  let (_, newGoals) ← tacticM { elaborator := .anonymous } |>.run { goals := [goal] }
+  let (_, { goals }) ← tacticM { elaborator := .anonymous } |>.run { goals := [goal] }
   let nextElabState ← MonadBacktrack.saveState
+  Elab.Term.synthesizeSyntheticMVarsNoPostponing
+
+  let goals ← if guardMVarErrors then
+      pure $ mergeMVarLists goals (← collectAllErroredMVars goal)
+    else
+      pure goals
   return {
     state with
-    savedState := { term := nextElabState, tactic := newGoals },
+    savedState := { term := nextElabState, tactic := { goals }, },
     parentMVar? := .some goal,
     calcPrevRhs? := .none,
   }
@@ -203,13 +238,13 @@ inductive TacticResult where
   | invalidAction (message: String)
 
 /-- Executes a `TacticM` monad on this `GoalState`, collecting the errors as necessary -/
-protected def GoalState.tryTacticM (state: GoalState) (goal: MVarId) (tacticM: Elab.Tactic.TacticM Unit):
+protected def GoalState.tryTacticM (state: GoalState) (goal: MVarId) (tacticM: Elab.Tactic.TacticM Unit) (guardMVarErrors : Bool := false):
       Elab.TermElabM TacticResult := do
   try
-    let nextState ← state.step goal tacticM
+    let nextState ← state.step goal tacticM guardMVarErrors
 
     -- Check if error messages have been generated in the core.
-    let newMessages ← (← Core.getMessageLog).toList.drop state.coreState.messages.toList.length
+    let newMessages ← (← Core.getMessageLog).toList --.drop state.coreState.messages.toList.length
       |>.filterMapM λ m => do
         if m.severity == .error then
           return .some $ ← m.toString
@@ -233,7 +268,7 @@ protected def GoalState.tryTactic (state: GoalState) (goal: MVarId) (tactic: Str
     (fileName := ← getFileName) with
     | .ok stx => pure $ stx
     | .error error => return .parseError error
-  state.tryTacticM goal $ Elab.Tactic.evalTactic tactic
+  state.tryTacticM goal (Elab.Tactic.evalTactic tactic) true
 
 protected def GoalState.tryAssign (state: GoalState) (goal: MVarId) (expr: String):
       Elab.TermElabM TacticResult := do
diff --git a/README.md b/README.md
index e070fff..02de68c 100644
--- a/README.md
+++ b/README.md
@@ -7,7 +7,7 @@ A Machine-to-Machine interaction system for Lean 4.
 Pantograph provides interfaces to execute proofs, construct expressions, and
 examine the symbol list of a Lean project for machine learning.
 
-See [documentations](doc/) for design rationale and references.
+See [documentations](doc/rationale.md) for design rationale and references.
 
 ## Installation
 
diff --git a/Test/Proofs.lean b/Test/Proofs.lean
index 65f099f..a6b5487 100644
--- a/Test/Proofs.lean
+++ b/Test/Proofs.lean
@@ -701,7 +701,7 @@ def test_nat_zero_add_alt: TestM Unit := do
       }
     ])
 
-def test_composite_tactic_failure: TestM Unit := do
+def test_tactic_failure_unresolved_goals : TestM Unit := do
   let state? ← startProof (.expr "∀ (p : Nat → Prop), ∃ (x : Nat), p (0 + x + 0)")
   let state0 ← match state? with
     | .some state => pure state
@@ -720,6 +720,37 @@ def test_composite_tactic_failure: TestM Unit := do
   let .failure messages ← state1.tacticOn 0 tactic | addTest $ assertUnreachable s!"{tactic} should fail"
   checkEq s!"{tactic} fails" messages #[s!"{← getFileName}:0:12: error: unsolved goals\np : Nat → Prop\n⊢ p 0\n"]
 
+
+def test_tactic_failure_synthesize_placeholder : TestM Unit := do
+  let state? ← startProof (.expr "∀ (p q r : Prop) (h : p → q), q ∧ r")
+  let state0 ← match state? with
+    | .some state => pure state
+    | .none => do
+      addTest $ assertUnreachable "Goal could not parse"
+      return ()
+
+  let tactic := "intro p q r h"
+  let state1 ← match ← state0.tacticOn 0 tactic with
+    | .success state => pure state
+    | other => do
+      addTest $ assertUnreachable $ other.toString
+      return ()
+
+  let tactic := "simpa [h] using And.imp_left h _"
+  let state2 ← match ← state1.tacticOn 0 tactic with
+    | .success state => pure state
+    | other => do
+      addTest $ assertUnreachable $ other.toString
+      return ()
+
+  checkEq tactic ((← state2.serializeGoals).map (·.devolatilize))  #[
+    buildGoal [("p", "Prop"), ("q", "Prop"), ("r", "Prop"), ("h", "p → q")] "p ∧ r"
+  ]
+
+  --let .failure messages ← state1.tacticOn 0 tactic | addTest $ assertUnreachable s!"{tactic} should fail"
+  --let message := s!"<Pantograph>:0:31: error: don't know how to synthesize placeholder\ncontext:\np q r : Prop\nh : p → q\n⊢ p ∧ r\n"
+  --checkEq s!"{tactic} fails" messages #[message]
+
 def suite (env: Environment): List (String × IO LSpec.TestSeq) :=
   let tests := [
     ("identity", test_identity),
@@ -732,7 +763,8 @@ def suite (env: Environment): List (String × IO LSpec.TestSeq) :=
     ("calc", test_calc),
     ("Nat.zero_add", test_nat_zero_add),
     ("Nat.zero_add alt", test_nat_zero_add_alt),
-    ("composite tactic failure", test_composite_tactic_failure),
+    ("tactic failure with unresolved goals", test_tactic_failure_unresolved_goals),
+    ("tactic failure with synthesize placeholder", test_tactic_failure_synthesize_placeholder),
   ]
   tests.map (fun (name, test) => (name, proofRunner env test))
 
diff --git a/doc/rationale.md b/doc/rationale.md
index 87c1606..4209474 100644
--- a/doc/rationale.md
+++ b/doc/rationale.md
@@ -24,6 +24,22 @@ The name Pantograph is a pun. It means two things
   a locomotive. In comparison the (relatively) simple Pantograph software powers
   theorem proving projects.
 
+## Caveats
+
+Pantograph does not exactly mimic Lean LSP's behaviour. That would not grant the
+flexibility it offers.  To support tree search means Pantograph has to act
+differently from Lean in some times, but never at the sacrifice of soundness.
+
+- When Lean LSP says "don't know how to synthesize placeholder", this indicates
+  the human operator needs to manually move the cursor to the placeholder and
+  type in the correct expression. This error therefore should not halt the proof
+  process, and the placeholder should be turned into a goal.
+- When Lean LSP says "unresolved goals", that means a proof cannot finish where
+  it is supposed to finish at the end of a `by` block. Pantograph will raise the
+  error in this case, since it indicates the termination of a proof search branch.
+- `pick_goal` or `swap` will not work since they run contrary to tree search
+  paradigms.
+
 ## References
 
 * [Pantograph Paper](https://arxiv.org/abs/2410.16429)