Skip to content
Closed
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
126 changes: 126 additions & 0 deletions lycaon/internal/toolexecution/program_preparation_review_test.go
Original file line number Diff line number Diff line change
@@ -0,0 +1,126 @@
package toolexecution

import (
"bytes"
"context"
"io"
"os"
"path/filepath"
"testing"

"github.com/lycaon/lycaon/internal/commandprogram"
"github.com/lycaon/lycaon/internal/hitl"
"github.com/lycaon/lycaon/internal/platform"
"github.com/lycaon/lycaon/internal/progress"
"github.com/lycaon/lycaon/internal/sandbox"
"github.com/lycaon/lycaon/internal/settings"
"github.com/lycaon/lycaon/internal/testutil"
"github.com/lycaon/lycaon/internal/toolcontract"
"github.com/lycaon/lycaon/internal/toolprofiles"
"github.com/lycaon/lycaon/internal/tools"
"github.com/lycaon/lycaon/internal/tools/projectpaths"
)

// Preparation can read source bytes before the final producer is invoked.
// A human reply cannot preserve an operation removed during that wait.
func TestNativePreparationRechecksProfileAfterApproval(t *testing.T) {
for _, withdraw := range []bool{false, true} {
name := "unchanged"
if withdraw {
name = "withdrawn"
}
t.Run(name, func(t *testing.T) {
parent, tc := frameFixture(t)
ctx := sandbox.WithSessionID(t.Context(), parent.SessionID)
testutil.FailErr(t, "seed preparation source", os.WriteFile(filepath.Join(tc.ActiveRootPath(), "input"), []byte("source bytes"), 0600))
classes, err := toolcontract.NewCommandEffects(toolcontract.CommandReadOnly)
testutil.FailErr(t, "read effect class", err)
profile := sandbox.ToolProfile{ID: "fixture", CommandEffects: classes}
boundary := sandbox.NewBoundary(sandbox.Config{ProjectRootRequired: true, RejectSymlinkEscape: true}, []sandbox.ToolProfile{profile})
boundary.SetProfileSource(func(context.Context, string) []sandbox.ToolProfile { return []sandbox.ToolProfile{profile} })
store, err := settings.NewApprovalStoreAt(filepath.Join(t.TempDir(), "approvals.yaml"))
testutil.FailErr(t, "create policy store", err)
testutil.FailErr(t, "require read review", store.PutProject(tc.ActiveRootPath(), settings.ApprovalConfig{Rules: []settings.ApprovalRule{{Category: settings.ApprovalCategoryTool, Pattern: "read", Effect: settings.ApprovalEffectAsk}}}))
gate := settings.NewRuleApprovalGate(store, settings.NoSources())
executor := NewExecutor(NewApprovalPolicyEngine(toolprofiles.NewProfilePolicyEngine(boundary), gate), tools.NewDefaultRegistry(), "fixture")
definition := nativeDefinition(t, "read")
surface, err := executor.Metadata.FreezeProgramSurface(ctx, platform.ToolFilter{ProfileID: "fixture"}, []toolcontract.OperationDefinition{definition})
testutil.FailErr(t, "freeze initial profile", err)
frameMemory := stageMemory(t, 64<<20)
frame, err := FreezeProgramFrame(ctx, parent, tc, ProgramModelInvocation{ToolName: "command", Arguments: map[string]any{"command": "cat input"}}, frameMemory)
testutil.FailErr(t, "capture program", err)
t.Cleanup(frame.Close)
child := parent
child.Identity.ExecutionID, child.Identity.ParentExecutionID = "child", parent.Identity.ExecutionID
bound, err := toolcontract.BindOperation(definition, child, "cat", definition.Facts().Commands[0].Effects[0].Mode, nil)
testutil.FailErr(t, "bind read", err)
var preparedBytes, output bytes.Buffer
asks, records := 0, 0
manager := &nativeRecorderReviewManager{invariantCheckpointMgr: invariantCheckpointMgr{status: hitl.DecisionStatusApproved, result: &hitl.DecisionResult{Approved: true}}}
manager.onRequest = func(req hitl.CheckpointRequest) {
asks++
if asks != 1 || preparedBytes.Len() != 0 || records != 0 {
t.Fatal("review repeated or followed preparation")
}
matched := false
for _, rule := range req.ApprovalMatches {
matched = matched || rule.Category == "tool" && rule.Pattern == "read"
}
if !matched {
t.Fatal("read requirement absent from review")
}
if withdraw {
profile.DenyTools = []string{"read"}
}
}
executor.Approvals.SetCheckpointManager(ctx, manager, gate)
t.Cleanup(func() {
testutil.FailErr(t, "release egress resolver", executor.Approvals.ReleaseEgressResolver(context.Background()))
})
effect := NativeEffect{Operation: bound, Stage: toolcontract.ExecutionStage{ExecutionID: "child", OperationID: "read", Argv: []string{"cat", "input"}}, Arguments: map[string]any{"path": "input"}, Streams: commandprogram.Streams{Output: &output}}
effect.PrepareReviewed = func(ctx context.Context, args map[string]any, actual tools.ToolContext) (PreparedNativeEffect, error) {
paths, err := projectpaths.NewProgramPaths(ctx, boundary, actual, ".")
if err != nil {
return PreparedNativeEffect{}, err
}
path, err := paths.Bind(ctx, args["path"].(string), false)
if err != nil {
return PreparedNativeEffect{}, err
}
file, err := path.Open()
if err != nil {
return PreparedNativeEffect{}, err
}
defer file.Close()
if _, err := io.Copy(&preparedBytes, file); err != nil {
return PreparedNativeEffect{}, err
}
return PreparedNativeEffect{Operation: bound, InvokeOwner: func(_ context.Context, _ tools.ToolContext, streams commandprogram.Streams) (toolcontract.ProducerResult, error) {
_, err := io.Copy(streams.Output, bytes.NewReader(preparedBytes.Bytes()))
return toolcontract.ProducerResult{Terminal: toolcontract.ProducerCompleted, Complete: true, PublicationEligible: true, EffectCertainty: toolcontract.NoEffect}, err
}}, nil
}
owner := progress.NewMemoryStore()
testutil.FailErr(t, "seed progress", owner.Set(parent.SessionID, "- [ ] Inspect"))
memory := stageMemory(t, 64<<20)
result, err := executor.InvokeNative(ctx, frame, surface, memory, nativeProgressFixture{owner}, nativeStartFunc(func(context.Context, toolcontract.BoundOperation, string, toolcontract.ReviewedOperationInputs) error {
records++
return nil
}), effect, tc)
frame.Close()
if asks != 1 || memory.InUse().State != 0 || frameMemory.InUse().State != 0 {
t.Fatal("review or retention contract failed")
}
if withdraw {
if err == nil || preparedBytes.Len() != 0 || output.Len() != 0 || records != 0 || result.OwnerInvoked {
t.Fatalf("withdrawn operation read source: err=%v prepared=%q output=%q records=%d owner=%t", err, preparedBytes.String(), output.String(), records, result.OwnerInvoked)
}
} else {
testutil.FailErr(t, "resume approved preparation", err)
if preparedBytes.String() != "source bytes" || output.String() != "source bytes" || records != 1 || !result.OwnerInvoked || !result.OutputPublishable {
t.Fatalf("authorized preparation failed: prepared=%q output=%q records=%d result=%+v", preparedBytes.String(), output.String(), records, result)
}
}
})
}
}