The reasoning trace was rendered twice: once as its own "thinking" row on inference.completed, and again as a ✼ block on the following tool row, whose Reasoning was copied from Model.lastReasoning. The fallback branch was worse — lastReasoning is current model state, so every tool row with no Reasoning of its own (all non-diff rows, all snapshot-restored rows) got the newest trace stamped on it retroactively. Drop the block, the copy, and the now-dead RouterEntry.Reasoning / Model.lastReasoning; the standalone row already covers tool turns. actionToolText clipped every summary to 48 columns, which cut tool output and — worse — harness coach text: read-before-write rejections, write blocks, gate feedback, exactly the messages that say whether the agent was steered or silently blocked. The action renderer already wraps to panel width, so the clip was the only single-line ceiling; raise it to 4000 chars. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
@@ -218,7 +218,8 @@ func PreviewFrame(kind string, w, h int) string {
|
||||
m.routerMessages["04a546aa"] = []RouterEntry{
|
||||
{Role: "user", Content: "fix the healthcheck script and search for all usages"},
|
||||
{Role: "router", Content: "I'll read the current file, grep for references, then write the fix."},
|
||||
{Role: "tool", Content: "--- a/healthcheck.sh\n+++ b/healthcheck.sh\n@@ -0,0 +1,4 @@\n+#!/usr/bin/env bash\n+curl -sf http://localhost:8080/health\n+echo ok\n+exit 0\n", Reasoning: "The user wants a script that checks a health endpoint and reports the status."},
|
||||
{Role: "thinking", Content: "The user wants a script that checks a health endpoint and reports the status."},
|
||||
{Role: "tool", Content: "--- a/healthcheck.sh\n+++ b/healthcheck.sh\n@@ -0,0 +1,4 @@\n+#!/usr/bin/env bash\n+curl -sf http://localhost:8080/health\n+echo ok\n+exit 0\n"},
|
||||
{Role: "action", Icon: "✎", Content: "wrote healthcheck.sh (+12 −0)"},
|
||||
{Role: "action", Icon: "✓", Content: "ReadFile (path=/etc/hosts, offset=0, limit=200) · 28 lines"},
|
||||
{Role: "action", Icon: "✓", Content: "ListDir (path=/home/kami, pattern=*.sh, recursive=false) · 3 entries"},
|
||||
|
||||
@@ -75,7 +75,6 @@ type RouterEntry struct {
|
||||
Content string
|
||||
Icon string // action role only: the gutter glyph (✓ ✎ ✗ ⌘ ✕ ⊞ ⊟)
|
||||
Metrics *TurnMetrics
|
||||
Reasoning string // model reasoning/CoT that preceded this entry (tool/artifact turns)
|
||||
}
|
||||
|
||||
// TurnMetrics carries optional latency + token cost for a ROUTER chat turn.
|
||||
@@ -442,10 +441,6 @@ type Model struct {
|
||||
// (0 = tail-follow the newest output). PgUp/PgDn + ctrl+u/d move it; esc snaps back.
|
||||
outputScroll int
|
||||
|
||||
// lastReasoning holds the model's reasoning/CoT trace from the most recent
|
||||
// inference.completed, surfaced on the following tool-call and artifact rows.
|
||||
lastReasoning string
|
||||
|
||||
// event-inspector filter (OverlayEventInspector): narrows the event list by a substring
|
||||
// of type/detail. eventFilterTyping is true while the operator is editing the query after /.
|
||||
eventFilter string
|
||||
|
||||
@@ -0,0 +1,42 @@
|
||||
package app
|
||||
|
||||
import (
|
||||
"strings"
|
||||
"testing"
|
||||
)
|
||||
|
||||
// #414: tool output and harness coach text (read-before-write, write blocks) must survive
|
||||
// into the transcript in full and wrap across rows — not get clipped to a single line.
|
||||
func TestActionToolTextKeepsFullSummary(t *testing.T) {
|
||||
coach := "write rejected: read apps/server/src/main/kotlin/Dtos.kt before editing it — " +
|
||||
"the file is not in this stage's read manifest, so the edit would be blind."
|
||||
got := actionToolText("file_write blocked", coach)
|
||||
if !strings.Contains(got, "read manifest") {
|
||||
t.Fatalf("coach text was clipped: %q", got)
|
||||
}
|
||||
|
||||
m := inSessionModel(120, 30)
|
||||
m.routerMessages[m.selectedID] = []RouterEntry{{Role: "action", Icon: "✕", Content: got}}
|
||||
w, _ := m.outputViewport()
|
||||
rows, _ := m.buildTranscriptRows(w)
|
||||
if len(rows) < 2 {
|
||||
t.Fatalf("long action content rendered in %d row(s), want it wrapped over several", len(rows))
|
||||
}
|
||||
}
|
||||
|
||||
// #415: the reasoning trace renders once per turn (its own "thinking" row) — the following
|
||||
// tool row must not repeat it.
|
||||
func TestReasoningRendersOncePerTurn(t *testing.T) {
|
||||
const cot = "unique-cot-marker: check the health endpoint first"
|
||||
m := inSessionModel(120, 30)
|
||||
m.thinkingShown = true
|
||||
m.routerMessages[m.selectedID] = []RouterEntry{
|
||||
{Role: "thinking", Content: cot},
|
||||
{Role: "tool", Content: "--- a/x.sh\n+++ b/x.sh\n@@ -0,0 +1 @@\n+echo ok\n"},
|
||||
}
|
||||
w, _ := m.outputViewport()
|
||||
rows, _ := m.buildTranscriptRows(w)
|
||||
if n := strings.Count(stripANSI(strings.Join(rows, "\n")), "unique-cot-marker"); n != 1 {
|
||||
t.Fatalf("reasoning rendered %d times, want 1", n)
|
||||
}
|
||||
}
|
||||
@@ -194,7 +194,6 @@ func (m *Model) applyServer(msg protocol.ServerMessage) {
|
||||
}
|
||||
}
|
||||
case protocol.TypeInferenceStarted:
|
||||
m.lastReasoning = ""
|
||||
if s := m.session(msg.SessionID); s != nil {
|
||||
s.Active = true
|
||||
s.addEvent(nowMillis(), "InferenceStarted", msg.StageID)
|
||||
@@ -214,7 +213,6 @@ func (m *Model) applyServer(msg protocol.ServerMessage) {
|
||||
// row (revealed via the palette). Skipped when the model emits no separate channel.
|
||||
if strings.TrimSpace(msg.Reasoning) != "" {
|
||||
m.appendRouter(msg.SessionID, RouterEntry{Role: "thinking", Content: msg.Reasoning})
|
||||
m.lastReasoning = msg.Reasoning
|
||||
}
|
||||
}
|
||||
case protocol.TypeInferenceTimeout:
|
||||
@@ -255,7 +253,7 @@ func (m *Model) applyServer(msg protocol.ServerMessage) {
|
||||
// existing collapsed diff row (^x opens the full diff).
|
||||
path, add, del := diffSummary(*msg.Diff)
|
||||
m.appendAction(msg.SessionID, "✎", "wrote "+path+countSuffix(add, del))
|
||||
m.appendRouter(msg.SessionID, RouterEntry{Role: "tool", Content: *msg.Diff, Reasoning: m.lastReasoning})
|
||||
m.appendRouter(msg.SessionID, RouterEntry{Role: "tool", Content: *msg.Diff})
|
||||
} else {
|
||||
label := msg.ToolName
|
||||
// Prefer the actual call args (path=…, command="…") over the affected-entities
|
||||
@@ -598,16 +596,21 @@ func paramSuffix(params []string) string {
|
||||
return " (" + clip(strings.Join(params, " · "), 200) + ")"
|
||||
}
|
||||
|
||||
// actionToolText joins a tool label with a short, clipped result summary.
|
||||
// actionToolText joins a tool label with its result summary. The action row wraps its
|
||||
// content to the panel width, so the summary is kept whole rather than clipped to one
|
||||
// line — tool output and harness coach text (read-before-write, write blocks, gate
|
||||
// feedback) are the payload, not decoration.
|
||||
func actionToolText(label, summary string) string {
|
||||
// Collapse to a single line first: tool summaries (dir listings, file heads)
|
||||
// often carry newlines, and a multi-line action row paints a background stripe
|
||||
// per line with the raw content leaking underneath.
|
||||
// Newlines are normalised away because the renderer re-wraps to panel width anyway;
|
||||
// leaving them in would paint a background stripe per raw line.
|
||||
s := strings.Join(strings.Fields(summary), " ")
|
||||
if s == "" {
|
||||
return label
|
||||
}
|
||||
return label + " · " + clip(s, 48)
|
||||
// ponytail: flat 4000-char ceiling so one recursive list_dir can't flood the
|
||||
// transcript. Swap for expand-on-select against the diff/preview surface if that
|
||||
// ceiling starts cutting real output.
|
||||
return label + " · " + clip(s, 4000)
|
||||
}
|
||||
|
||||
func approvalIcon(outcome string) string {
|
||||
|
||||
@@ -823,23 +823,9 @@ func (m Model) buildTranscriptRows(w int) ([]string, []int) {
|
||||
rows = append(rows, t.span(s, t.P.Faint))
|
||||
}
|
||||
case "tool":
|
||||
if e.Reasoning != "" {
|
||||
if m.thinkingShown {
|
||||
for _, ln := range strings.Split(strings.TrimRight(e.Reasoning, "\n"), "\n") {
|
||||
rows = append(rows, t.span(" ✼ "+ln, t.P.Faint))
|
||||
}
|
||||
} else {
|
||||
rows = append(rows, t.span(" ✼ reasoning — palette: thinking", t.P.Faint))
|
||||
}
|
||||
} else if m.lastReasoning != "" {
|
||||
if m.thinkingShown {
|
||||
for _, ln := range strings.Split(strings.TrimRight(m.lastReasoning, "\n"), "\n") {
|
||||
rows = append(rows, t.span(" ✼ "+ln, t.P.Faint))
|
||||
}
|
||||
} else {
|
||||
rows = append(rows, t.span(" ✼ reasoning — palette: thinking", t.P.Faint))
|
||||
}
|
||||
}
|
||||
// No reasoning block here: the trace already renders as its own "thinking" row
|
||||
// (appended on inference.completed), and repeating it on the following tool row
|
||||
// showed the same CoT twice per turn.
|
||||
gutter := t.span(" ┈", t.P.Faint)
|
||||
summary := toolRowSummary(e.Content)
|
||||
avail := w - 4
|
||||
|
||||
Reference in New Issue
Block a user