fix(turns): keep the run task in every stage's input

Later visits' input.md held only the previous stage's output, so the
operator's task reached only the first stage; when a Prepare stage
summarised (paths, HEAD), Code never saw the task or its acceptance
checklist and either blocked (run #202) or did part of it (#195, #197,
#198). Later inputs now carry the run prompt under '# Run task' (headings
demoted) before '# Previous turn output'.
This commit is contained in:
2026-10-31 02:52:16 +00:00
parent ea0f22dfc7
commit 13c4035175
8 changed files with 35 additions and 23 deletions
+3 -2
View File
@@ -662,8 +662,9 @@ files as that step's artifacts:
<artifacts>/<run>/public/
```
`input.md` is the run prompt on the first visit and the previous visit's output
body (without its outcome section) afterwards. `memory.md` is the accumulated markdown of completed turns plus the
`input.md` is the run prompt on the first visit; later visits carry the run
prompt under `# Run task` followed by the previous visit's output body (without
its outcome section) under `# Previous turn output`. `memory.md` is the accumulated markdown of completed turns plus the
files each turn added to the shared `public/` directory. A text-only stage (an
explicit empty `tools` block) receives both files' contents in its prompt
instead, because it has no file-reading tool. The model returns the complete
+4 -2
View File
@@ -16,8 +16,10 @@ Every model visit gets its own directory under the run's artifact root:
<artifacts>/<run>/public/
```
- `input.md` is the run prompt on the first visit and the previous visit's body
after that, under a `# Previous turn output` heading.
- `input.md` is the run prompt on the first visit. Later visits keep the run
prompt under a `# Run task` heading (its own headings demoted two levels) and
add the previous visit's body under a `# Previous turn output` heading, so
every stage sees what the run is for even when an earlier stage summarised.
- `memory.md` is one `## <stage> · visit <n>` section per completed turn: the
turn body followed by a system-written, sorted list of the files that turn
added to `public/`.
+5 -3
View File
@@ -15,9 +15,11 @@ the run's artifact root:
zero-padded to two digits). `Paths.Ensure` creates the turn, private and public
directories idempotently; it never deletes or rewrites existing files.
`input.md` is the run prompt on the first visit and the previous visit's output
body (without the outcome section) under a `# Previous turn output` heading
afterwards. `memory.md` is plain
`input.md` is the run prompt on the first visit. Afterwards it is the run
prompt under a `# Run task` heading (headings demoted two levels) followed by
the previous visit's output body (without the outcome section) under a
`# Previous turn output` heading. Carrying the task forward means a stage that
only summarised its own work is never the sole carrier of what the run is for. `memory.md` is plain
markdown: one `## <stage> · visit <n>` section per completed turn, with that
turn's body — its own ATX headings demoted two levels, so the turn heading stays
the structural one — followed by a system-written list of files it added to
+1 -1
View File
@@ -57,7 +57,7 @@ explicitly asks to manage a schedule; scheduling alone is not workflow authoring
## Turns and artifacts
Each visit has `input.md` (the run prompt or previous body), `memory.md`
Each visit has `input.md` (the run prompt; later visits add the previous body), `memory.md`
(completed turns), `output.md`, and `private/`. Every run also has one shared
`public/` directory, addressed by its absolute path; files there are visible to
later stages and are the stage's real artifacts. `input.md`, `memory.md`,
+2 -2
View File
@@ -89,8 +89,8 @@ func TestTurnsAccumulateMemoryAndListFilesAsArtifacts(t *testing.T) {
if detail.Status != "succeeded" || len(detail.Steps) != 2 {
t.Fatalf("run = %q steps=%d: %+v", detail.Status, len(detail.Steps), detail.RunSummary)
}
if secondInput != "# Previous turn output\n\nfirst body\n" {
t.Fatalf("second input.md = %q, want the previous turn's output", secondInput)
if secondInput != "# Run task\n\noriginal task\n\n# Previous turn output\n\nfirst body\n" {
t.Fatalf("second input.md = %q, want the run task and the previous turn's output", secondInput)
}
for _, want := range []string{"## first · visit 1", "first body", "Files added to public/:\n- public/shared.txt"} {
if !strings.Contains(secondMemory, want) {
+4 -4
View File
@@ -89,16 +89,16 @@ func turnInputWithRetry(base string, retryInput stage.Values) string {
return base + "\n# Retry input\n\nThis retry was given this complete input:\n\n```json\n" + string(encoded) + "\n```\n"
}
// turnInput is the run prompt on the first visit and the previous visit's
// output afterwards.
// turnInput is the run prompt on the first visit; later visits carry the run
// prompt and the previous visit's output.
func (s *runState) turnInput() string {
if len(s.turns) == 0 && strings.TrimSpace(s.lastOutput) == "" {
return s.prompt
}
if strings.TrimSpace(s.lastOutput) == "" {
return "# Previous turn output\n\n_(no output)_\n"
return turn.RenderInput(s.prompt, "_(no output)_\n")
}
return turn.RenderInput("", s.lastOutput)
return turn.RenderInput(s.prompt, s.lastOutput)
}
// turnOutput renders a completed step for later memory. Model stages store their
+10 -4
View File
@@ -6,14 +6,20 @@ import (
"strings"
)
// RenderInput renders the markdown a turn receives. The first turn receives the
// run prompt as-is. Later turns receive the previous turn's output verbatim
// under a heading; prompt is not repeated.
// RenderInput is a visit's input.md: the run prompt on the first visit, and on
// later visits the run prompt under "# Run task" followed by the previous
// visit's body under "# Previous turn output". Every stage keeps the operator's
// task: a stage that summarised its own work must not be the only carrier of
// what the run is for.
func RenderInput(prompt string, previousOutput string) string {
if strings.TrimSpace(previousOutput) == "" {
return prompt
}
return "# Previous turn output\n\n" + previousOutput
if strings.TrimSpace(prompt) == "" {
return "# Previous turn output\n\n" + previousOutput
}
task := strings.TrimRight(prompt, "\n")
return "# Run task\n\n" + demoteHeadings(task) + "\n\n# Previous turn output\n\n" + previousOutput
}
// MemoryEntry is one completed turn in a later turn's memory file. Output is
+6 -5
View File
@@ -15,15 +15,16 @@ func TestRenderInputFirstTurnIsTheRunPrompt(t *testing.T) {
}
}
func TestRenderInputLaterTurnShowsOnlyPreviousOutput(t *testing.T) {
func TestRenderInputLaterTurnKeepsTheTaskAndThePreviousOutput(t *testing.T) {
previous := "## Result\n\n- changed `internal/turn`\n\n```diff\n+new line\n```\n"
got := RenderInput("the original prompt", previous)
want := "# Previous turn output\n\n" + previous
got := RenderInput("# Goal\n\nThe original prompt.\n", previous)
want := "# Run task\n\n### Goal\n\nThe original prompt.\n\n# Previous turn output\n\n" + previous
if got != want {
t.Fatalf("later turn input = %q, want %q", got, want)
}
if strings.Contains(got, "the original prompt") {
t.Fatal("later turn input repeated the run prompt")
// Without a recorded prompt the previous output stands alone.
if got := RenderInput("", previous); got != "# Previous turn output\n\n"+previous {
t.Fatalf("promptless later input = %q", got)
}
}