fix(core): preserve provider prompt prefixes

commit: 1a6392e94847c3e95db12d29c8799047b354c23c
author: brice · 2mo ago · 2026-08-18 13:17 UTC
parent: aa7307c
signature: GPG VerifiedCD88F9E564E0102646FB6077A1A07C7C7C3E349D · arch-gpg Verified for the committer email

2 files · +49 -81

M packages/core/context_budget.go +13 -31


@@ -9,48 +9,30 @@ import (
9	9		)
10	10
11	11		const (
12		-	maxToolResultTextBytes = 32 * 1024
13		-	maxToolResultTotalTextBytes = 128 * 1024
14		-	toolResultOmissionMarker = "[tool result content omitted]"
	12	+	maxToolResultTextBytes = 32 * 1024
	13	+	toolResultOmissionMarker = "[tool result content omitted]"
15	14		)
16	15
17	16		// projectToolResultMessages returns a copied provider-input view of msgs.
18		-	// Text inside each tool result is bounded independently and then against the
19		-	// total budget, with newer results receiving the remaining budget first.
20		-	// Tool-result blocks and all non-text content remain in the projection so
21		-	// tool-call/result pairing stays valid.
	17	+	// Each tool result is bounded independently. Its projection must depend only
	18	+	// on that result: retroactively shrinking older results when a new one arrives
	19	+	// mutates the prompt prefix and defeats provider prompt caching. Compaction
	20	+	// owns the aggregate transcript budget.
22	21		func projectToolResultMessages(msgs []provider.Message) []provider.Message {
23	22		projected := copyToolResultMessages(msgs)
24		-	remaining := maxToolResultTotalTextBytes
25	23
26		-	for i := len(projected) - 1; i >= 0; i-- {
27		-	for j := len(projected[i].Content) - 1; j >= 0; j-- {
28		-	result, ok := projected[i].Content[j].(provider.ToolResultBlock)
	24	+	for i := range projected {
	25	+	for j, content := range projected[i].Content {
	26	+	result, ok := content.(provider.ToolResultBlock)
29	27		if !ok {
30	28		continue
31	29		}
32	30
33		-	timingText := ""
34		-	if result.Timing != nil {
35		-	candidate := formatToolTiming(result.Timing)
36		-	if len(candidate) <= remaining {
37		-	timingText = candidate
38		-	remaining -= len(candidate)
39		-	}
40		-	}
41		-
42		-	textBytes := toolResultTextBytes(result.Content)
43		-	if textBytes > 0 {
44		-	limit := maxToolResultTextBytes
45		-	if remaining < limit {
46		-	limit = remaining
47		-	}
48		-	var retained int
49		-	result.Content, retained = projectToolResultContent(result.Content, limit)
50		-	remaining -= retained
	31	+	if toolResultTextBytes(result.Content) > 0 {
	32	+	result.Content, _ = projectToolResultContent(result.Content, maxToolResultTextBytes)
51	33		}
52		-	if timingText != "" {
53		-	result.Content = append(result.Content, provider.TextBlock{Text: timingText})
	34	+	if result.Timing != nil {
	35	+	result.Content = append(result.Content, provider.TextBlock{Text: formatToolTiming(result.Timing)})
54	36		}
55	37		projected[i].Content[j] = result
56	38		}

 )
 const (
-	maxToolResultTextBytes      = 32 * 1024
+	maxToolResultTextBytes   = 32 * 1024
-	maxToolResultTotalTextBytes = 128 * 1024
+	toolResultOmissionMarker = "[tool result content omitted]"
-	toolResultOmissionMarker    = "[tool result content omitted]"
 )
 // projectToolResultMessages returns a copied provider-input view of msgs.
-// Text inside each tool result is bounded independently and then against the
+// Each tool result is bounded independently. Its projection must depend only
-// total budget, with newer results receiving the remaining budget first.
+// on that result: retroactively shrinking older results when a new one arrives
-// Tool-result blocks and all non-text content remain in the projection so
+// mutates the prompt prefix and defeats provider prompt caching. Compaction
-// tool-call/result pairing stays valid.
+// owns the aggregate transcript budget.
 func projectToolResultMessages(msgs []provider.Message) []provider.Message {
 	projected := copyToolResultMessages(msgs)
-	remaining := maxToolResultTotalTextBytes
-	for i := len(projected) - 1; i >= 0; i-- {
+	for i := range projected {
-		for j := len(projected[i].Content) - 1; j >= 0; j-- {
+		for j, content := range projected[i].Content {
-			result, ok := projected[i].Content[j].(provider.ToolResultBlock)
+			result, ok := content.(provider.ToolResultBlock)
 			if !ok {
 				continue
 			}
-			timingText := ""
+			if toolResultTextBytes(result.Content) > 0 {
-			if result.Timing != nil {
+				result.Content, _ = projectToolResultContent(result.Content, maxToolResultTextBytes)
-				candidate := formatToolTiming(result.Timing)
-				if len(candidate) <= remaining {
-					timingText = candidate
-					remaining -= len(candidate)
-				}
-			}
-			textBytes := toolResultTextBytes(result.Content)
-			if textBytes > 0 {
-				limit := maxToolResultTextBytes
-				if remaining < limit {
-					limit = remaining
-				}
-				var retained int
-				result.Content, retained = projectToolResultContent(result.Content, limit)
-				remaining -= retained
 			}
-			if timingText != "" {
+			if result.Timing != nil {
-				result.Content = append(result.Content, provider.TextBlock{Text: timingText})
+				result.Content = append(result.Content, provider.TextBlock{Text: formatToolTiming(result.Timing)})
 			}
 			projected[i].Content[j] = result
 		}

M packages/core/context_budget_test.go +36 -50


@@ -2,6 +2,7 @@ package core
2	2
3	3		import (
4	4		"context"
	5	+	"fmt"
5	6		"reflect"
6	7		"strings"
7	8		"testing"

@@ -93,37 +94,28 @@ func TestProjectToolTimingIsModelOnly(t *testing.T) {
93	94		}
94	95		}
95	96
96		-	func TestProjectToolTimingCountsTowardAggregateBudget(t *testing.T) {
97		-	started := time.Date(2026, 8, 8, 12, 0, 0, 0, time.UTC)
98		-	timing := &provider.ToolTiming{StartedAt: started, CompletedAt: started.Add(time.Second), Duration: time.Second}
99		-	msgs := make([]provider.Message, 0, maxToolResultTotalTextBytes/maxToolResultTextBytes+1)
100		-	for i := 0; i < maxToolResultTotalTextBytes/maxToolResultTextBytes; i++ {
101		-	msgs = append(msgs, provider.Message{Role: provider.RoleTool, Content: []provider.Content{provider.ToolResultBlock{
102		-	CallID: string(rune('a' + i)),
103		-	Content: []provider.Content{provider.TextBlock{Text: strings.Repeat("x", maxToolResultTextBytes)}},
104		-	}}})
105		-	}
106		-	msgs = append(msgs, provider.Message{Role: provider.RoleTool, Content: []provider.Content{provider.ToolResultBlock{
107		-	CallID: "timed-empty", Timing: timing,
108		-	}}})
109		-
110		-	projected := projectToolResultMessages(msgs)
111		-	if got := toolResultTextForCall(projected, "timed-empty"); got != formatToolTiming(timing) {
112		-	t.Fatalf("empty timed result = %q, want timing annotation", got)
113		-	}
114		-	var total int
115		-	for _, message := range projected {
116		-	for _, content := range message.Content {
117		-	if result, ok := content.(provider.ToolResultBlock); ok {
118		-	total += retainedToolResultTextBytes(result.Content)
119		-	}
120		-	}
121		-	}
122		-	if total > maxToolResultTotalTextBytes {
123		-	t.Fatalf("projected tool-result text = %d bytes, want <= %d", total, maxToolResultTotalTextBytes)
	97	+	func TestProjectToolResultMessagesKeepsHistoricalPrefixStable(t *testing.T) {
	98	+	messages := make([]provider.Message, 0, 4)
	99	+	for i := 0; i < 4; i++ {
	100	+	messages = append(messages, provider.Message{Role: provider.RoleTool, Content: []provider.Content{
	101	+	provider.ToolResultBlock{
	102	+	CallID: fmt.Sprintf("call-%d", i),
	103	+	Content: []provider.Content{provider.TextBlock{Text: strings.Repeat(string(rune('a'+i)), maxToolResultTextBytes)}},
	104	+	},
	105	+	}})
124	106		}
125		-	if got := toolResultTextForCall(projected, "a"); !strings.Contains(got, toolResultOmissionMarker) {
126		-	t.Fatalf("oldest result = %q, want truncation after timing reservation", got)
	107	+
	108	+	before := projectToolResultMessages(messages)
	109	+	messages = append(messages, provider.Message{Role: provider.RoleTool, Content: []provider.Content{
	110	+	provider.ToolResultBlock{
	111	+	CallID: "new-call",
	112	+	Content: []provider.Content{provider.TextBlock{Text: strings.Repeat("new", maxToolResultTextBytes)}},
	113	+	},
	114	+	}})
	115	+	after := projectToolResultMessages(messages)
	116	+
	117	+	if !reflect.DeepEqual(after[:len(before)], before) {
	118	+	t.Fatal("appending a tool result changed the provider-visible historical prefix")
127	119		}
128	120		}
129	121

@@ -147,7 +139,6 @@ func TestProjectToolResultMessagesBoundsAndUTF8(t *testing.T) {
147	139		}
148	140
149	141		projected := projectToolResultMessages(msgs)
150		-	total := 0
151	142		for _, message := range projected {
152	143		for _, content := range message.Content {
153	144		result, ok := content.(provider.ToolResultBlock)

@@ -166,15 +157,11 @@ func TestProjectToolResultMessagesBoundsAndUTF8(t *testing.T) {
166	157		t.Fatalf("result %q contains invalid UTF-8", result.CallID)
167	158		}
168	159		}
169		-	total += retainedBytes
170	160		}
171	161		}
172		-	if total > maxToolResultTotalTextBytes {
173		-	t.Fatalf("projected tool-result text = %d bytes, want <= %d", total, maxToolResultTotalTextBytes)
174		-	}
175	162		}
176	163
177		-	func TestProjectToolResultMessagesRetainsNewestResultsFirst(t *testing.T) {
	164	+	func TestProjectToolResultMessagesBoundsEveryResultIndependently(t *testing.T) {
178	165		msgs := make([]provider.Message, 0, 10)
179	166		for i := 1; i <= 5; i++ {
180	167		id := "call-" + string(rune('0'+i))

@@ -192,18 +179,17 @@ func TestProjectToolResultMessagesRetainsNewestResultsFirst(t *testing.T) {
192	179		}
193	180
194	181		projected := projectToolResultMessages(msgs)
195		-	oldest := toolResultTextForCall(projected, "call-1")
196		-	newest := toolResultTextForCall(projected, "call-5")
197		-	if oldest != toolResultOmissionMarker {
198		-	t.Fatalf("oldest result = %q, want omission marker after newer results filled budget", oldest)
199		-	}
200		-	if !strings.HasPrefix(newest, "result-call-5 ") \|\| !strings.Contains(newest, toolResultOmissionMarker) {
201		-	t.Fatalf("newest result was not retained and truncated first: %q", newest[:min(len(newest), 80)])
	182	+	for i := 1; i <= 5; i++ {
	183	+	id := "call-" + string(rune('0'+i))
	184	+	got := toolResultTextForCall(projected, id)
	185	+	if !strings.HasPrefix(got, "result-"+id+" ") \|\| !strings.Contains(got, toolResultOmissionMarker) {
	186	+	t.Fatalf("result %s was not independently retained and truncated: %q", id, got[:min(len(got), 80)])
	187	+	}
202	188		}
203	189		}
204	190
205		-	func TestProjectToolResultMessagesMarksResultsAfterAggregateBudgetIsExhausted(t *testing.T) {
206		-	const resultCount = maxToolResultTotalTextBytes/maxToolResultTextBytes + 1
	191	+	func TestProjectToolResultMessagesDoesNotExhaustAnAggregateBudget(t *testing.T) {
	192	+	const resultCount = 5
207	193		msgs := make([]provider.Message, 0, resultCount)
208	194		for i := 0; i < resultCount; i++ {
209	195		id := "call-" + string(rune('1'+i))

@@ -230,11 +216,11 @@ func TestProjectToolResultMessagesMarksResultsAfterAggregateBudgetIsExhausted(t
230	216		retainedTotal += retainedToolResultTextBytes(result.Content)
231	217		}
232	218		}
233		-	if retainedTotal != maxToolResultTotalTextBytes {
234		-	t.Fatalf("retained tool-result text = %d bytes, want exactly %d", retainedTotal, maxToolResultTotalTextBytes)
	219	+	if retainedTotal != resultCount*maxToolResultTextBytes {
	220	+	t.Fatalf("retained tool-result text = %d bytes, want %d independent bytes", retainedTotal, resultCount*maxToolResultTextBytes)
235	221		}
236		-	if got := toolResultTextForCall(projected, "call-1"); got != toolResultOmissionMarker {
237		-	t.Fatalf("aggregate-exhausted result = %q, want omission marker", got)
	222	+	if got := toolResultTextForCall(projected, "call-1"); got == toolResultOmissionMarker {
	223	+	t.Fatalf("old result was discarded after later results: %q", got)
238	224		}
239	225		}
240	226

@@ -291,7 +277,7 @@ func TestAgentNormalRequestProjectsToolResultsAfterPairRepair(t *testing.T) {
291	277		t.Fatalf("Continue: %v", err)
292	278		}
293	279		projectedText := toolResultTextForCall(client.lastReq.Messages, "call-1")
294		-	if retainedToolResultTextBytesForCall(client.lastReq.Messages, "call-1") > maxToolResultTotalTextBytes \|\| !strings.Contains(projectedText, toolResultOmissionMarker) {
	280	+	if !strings.Contains(projectedText, toolResultOmissionMarker) {
295	281		t.Fatalf("normal request result was not bounded: %d bytes, %q", len(projectedText), projectedText[:min(len(projectedText), 80)])
296	282		}
297	283		if !strings.Contains(projectedText, formatToolTiming(timing)) {


@@ -2,6 +2,7 @@ package core
2			2
3		import (	3		import (
4		"context"	4		"context"
			5	+	"fmt"
5		"reflect"	6		"reflect"
6		"strings"	7		"strings"
7		"testing"	8		"testing"

@@ -93,37 +94,28 @@ func TestProjectToolTimingIsModelOnly(t *testing.T) {
93		}	94		}
94		}	95		}
95			96
96	-	func TestProjectToolTimingCountsTowardAggregateBudget(t *testing.T) {	97	+	func TestProjectToolResultMessagesKeepsHistoricalPrefixStable(t *testing.T) {
97	-	started := time.Date(2026, 8, 8, 12, 0, 0, 0, time.UTC)	98	+	messages := make([]provider.Message, 0, 4)
98	-	timing := &provider.ToolTiming{StartedAt: started, CompletedAt: started.Add(time.Second), Duration: time.Second}	99	+	for i := 0; i < 4; i++ {
99	-	msgs := make([]provider.Message, 0, maxToolResultTotalTextBytes/maxToolResultTextBytes+1)	100	+	messages = append(messages, provider.Message{Role: provider.RoleTool, Content: []provider.Content{
100	-	for i := 0; i < maxToolResultTotalTextBytes/maxToolResultTextBytes; i++ {	101	+	provider.ToolResultBlock{
101	-	msgs = append(msgs, provider.Message{Role: provider.RoleTool, Content: []provider.Content{provider.ToolResultBlock{	102	+	CallID: fmt.Sprintf("call-%d", i),
102	-	CallID: string(rune('a' + i)),	103	+	Content: []provider.Content{provider.TextBlock{Text: strings.Repeat(string(rune('a'+i)), maxToolResultTextBytes)}},
103	-	Content: []provider.Content{provider.TextBlock{Text: strings.Repeat("x", maxToolResultTextBytes)}},	104	+	},
104	-	}}})	105	+	}})
105	-	}
106	-	msgs = append(msgs, provider.Message{Role: provider.RoleTool, Content: []provider.Content{provider.ToolResultBlock{
107	-	CallID: "timed-empty", Timing: timing,
108	-	}}})
109	-
110	-	projected := projectToolResultMessages(msgs)
111	-	if got := toolResultTextForCall(projected, "timed-empty"); got != formatToolTiming(timing) {
112	-	t.Fatalf("empty timed result = %q, want timing annotation", got)
113	-	}
114	-	var total int
115	-	for _, message := range projected {
116	-	for _, content := range message.Content {
117	-	if result, ok := content.(provider.ToolResultBlock); ok {
118	-	total += retainedToolResultTextBytes(result.Content)
119	-	}
120	-	}
121	-	}
122	-	if total > maxToolResultTotalTextBytes {
123	-	t.Fatalf("projected tool-result text = %d bytes, want <= %d", total, maxToolResultTotalTextBytes)
124		}	106		}
125	-	if got := toolResultTextForCall(projected, "a"); !strings.Contains(got, toolResultOmissionMarker) {	107	+
126	-	t.Fatalf("oldest result = %q, want truncation after timing reservation", got)	108	+	before := projectToolResultMessages(messages)
			109	+	messages = append(messages, provider.Message{Role: provider.RoleTool, Content: []provider.Content{
			110	+	provider.ToolResultBlock{
			111	+	CallID: "new-call",
			112	+	Content: []provider.Content{provider.TextBlock{Text: strings.Repeat("new", maxToolResultTextBytes)}},
			113	+	},
			114	+	}})
			115	+	after := projectToolResultMessages(messages)
			116	+
			117	+	if !reflect.DeepEqual(after[:len(before)], before) {
			118	+	t.Fatal("appending a tool result changed the provider-visible historical prefix")
127		}	119		}
128		}	120		}
129			121

@@ -147,7 +139,6 @@ func TestProjectToolResultMessagesBoundsAndUTF8(t *testing.T) {
147		}	139		}
148			140
149		projected := projectToolResultMessages(msgs)	141		projected := projectToolResultMessages(msgs)
150	-	total := 0
151		for _, message := range projected {	142		for _, message := range projected {
152		for _, content := range message.Content {	143		for _, content := range message.Content {
153		result, ok := content.(provider.ToolResultBlock)	144		result, ok := content.(provider.ToolResultBlock)

@@ -166,15 +157,11 @@ func TestProjectToolResultMessagesBoundsAndUTF8(t *testing.T) {
166		t.Fatalf("result %q contains invalid UTF-8", result.CallID)	157		t.Fatalf("result %q contains invalid UTF-8", result.CallID)
167		}	158		}
168		}	159		}
169	-	total += retainedBytes
170		}	160		}
171		}	161		}
172	-	if total > maxToolResultTotalTextBytes {
173	-	t.Fatalf("projected tool-result text = %d bytes, want <= %d", total, maxToolResultTotalTextBytes)
174	-	}
175		}	162		}
176			163
177	-	func TestProjectToolResultMessagesRetainsNewestResultsFirst(t *testing.T) {	164	+	func TestProjectToolResultMessagesBoundsEveryResultIndependently(t *testing.T) {
178		msgs := make([]provider.Message, 0, 10)	165		msgs := make([]provider.Message, 0, 10)
179		for i := 1; i <= 5; i++ {	166		for i := 1; i <= 5; i++ {
180		id := "call-" + string(rune('0'+i))	167		id := "call-" + string(rune('0'+i))

@@ -192,18 +179,17 @@ func TestProjectToolResultMessagesRetainsNewestResultsFirst(t *testing.T) {
192		}	179		}
193			180
194		projected := projectToolResultMessages(msgs)	181		projected := projectToolResultMessages(msgs)
195	-	oldest := toolResultTextForCall(projected, "call-1")	182	+	for i := 1; i <= 5; i++ {
196	-	newest := toolResultTextForCall(projected, "call-5")	183	+	id := "call-" + string(rune('0'+i))
197	-	if oldest != toolResultOmissionMarker {	184	+	got := toolResultTextForCall(projected, id)
198	-	t.Fatalf("oldest result = %q, want omission marker after newer results filled budget", oldest)	185	+	if !strings.HasPrefix(got, "result-"+id+" ") \|\| !strings.Contains(got, toolResultOmissionMarker) {
199	-	}	186	+	t.Fatalf("result %s was not independently retained and truncated: %q", id, got[:min(len(got), 80)])
200	-	if !strings.HasPrefix(newest, "result-call-5 ") \|\| !strings.Contains(newest, toolResultOmissionMarker) {	187	+	}
201	-	t.Fatalf("newest result was not retained and truncated first: %q", newest[:min(len(newest), 80)])
202		}	188		}
203		}	189		}
204			190
205	-	func TestProjectToolResultMessagesMarksResultsAfterAggregateBudgetIsExhausted(t *testing.T) {	191	+	func TestProjectToolResultMessagesDoesNotExhaustAnAggregateBudget(t *testing.T) {
206	-	const resultCount = maxToolResultTotalTextBytes/maxToolResultTextBytes + 1	192	+	const resultCount = 5
207		msgs := make([]provider.Message, 0, resultCount)	193		msgs := make([]provider.Message, 0, resultCount)
208		for i := 0; i < resultCount; i++ {	194		for i := 0; i < resultCount; i++ {
209		id := "call-" + string(rune('1'+i))	195		id := "call-" + string(rune('1'+i))

@@ -230,11 +216,11 @@ func TestProjectToolResultMessagesMarksResultsAfterAggregateBudgetIsExhausted(t
230		retainedTotal += retainedToolResultTextBytes(result.Content)	216		retainedTotal += retainedToolResultTextBytes(result.Content)
231		}	217		}
232		}	218		}
233	-	if retainedTotal != maxToolResultTotalTextBytes {	219	+	if retainedTotal != resultCount*maxToolResultTextBytes {
234	-	t.Fatalf("retained tool-result text = %d bytes, want exactly %d", retainedTotal, maxToolResultTotalTextBytes)	220	+	t.Fatalf("retained tool-result text = %d bytes, want %d independent bytes", retainedTotal, resultCount*maxToolResultTextBytes)
235		}	221		}
236	-	if got := toolResultTextForCall(projected, "call-1"); got != toolResultOmissionMarker {	222	+	if got := toolResultTextForCall(projected, "call-1"); got == toolResultOmissionMarker {
237	-	t.Fatalf("aggregate-exhausted result = %q, want omission marker", got)	223	+	t.Fatalf("old result was discarded after later results: %q", got)
238		}	224		}
239		}	225		}
240			226

@@ -291,7 +277,7 @@ func TestAgentNormalRequestProjectsToolResultsAfterPairRepair(t *testing.T) {
291		t.Fatalf("Continue: %v", err)	277		t.Fatalf("Continue: %v", err)
292		}	278		}
293		projectedText := toolResultTextForCall(client.lastReq.Messages, "call-1")	279		projectedText := toolResultTextForCall(client.lastReq.Messages, "call-1")
294	-	if retainedToolResultTextBytesForCall(client.lastReq.Messages, "call-1") > maxToolResultTotalTextBytes \|\| !strings.Contains(projectedText, toolResultOmissionMarker) {	280	+	if !strings.Contains(projectedText, toolResultOmissionMarker) {
295		t.Fatalf("normal request result was not bounded: %d bytes, %q", len(projectedText), projectedText[:min(len(projectedText), 80)])	281		t.Fatalf("normal request result was not bounded: %d bytes, %q", len(projectedText), projectedText[:min(len(projectedText), 80)])
296		}	282		}
297		if !strings.Contains(projectedText, formatToolTiming(timing)) {	283		if !strings.Contains(projectedText, formatToolTiming(timing)) {