Compare commits

..
Author SHA1 Message Date
Vasanthdev2004 c4fff2b1de test(sandbox): give the cross-platform anchor tests native paths
Moving these beside the other BuildWindowsACLPlan tests made them run on
Linux and macOS, where their Windows path literals stop meaning what they
say: pathWithinRoot is filepath.Rel underneath, a backslash is an ordinary
character off Windows, and C:\workspace\.git\hooks is then one component
that is not under C:\workspace, so every anchor came back empty. The
neighbouring tests get away with such literals because they only compare
strings they built the same way and never ask whether one contains another.

Paths are built with filepath.Join from temp roots now, so the containment
question is asked in the separator the running platform actually uses.
2026-09-12 09:47:03 +05:30
Vasanthdev2004 38c1e333b1 test(sandbox): check the plan anchoring on every platform, out-of-root case included
The plan builder is cross-platform and its anchor tests were in a Windows-only
file, so nothing checked the wiring on Linux or macOS. Moved beside the other
BuildWindowsACLPlan tests, with the case review raised: an out-of-root
ReadOnlySubpath stays unanchored, and an in-root one is still held to its
write root.
2026-09-12 09:47:03 +05:30
Vasanthdev2004 28360a2dff fix(sandbox): anchor only the carveouts that are under the root, and put the DACL back
Two from review, both right.

ReadOnlySubpaths is a profile field an operator can set to any path, and one
placed outside the write root is a configuration that works today; anchoring
it unconditionally turned that into a containment refusal. Only paths
lexically under the root are anchored now, which is where the derived
carveouts are anyway, and anything else keeps the final-component guard it
always had.

The successful-apply tests left their deny ACE in place, so t.TempDir could
fail to remove the tree. Restoring the snapshot exposed the sharper half:
the ACE denied the group the test runs as, which revoked its own WRITE_DAC
and left the rollback unable to reopen the target. The tests deny a group
this process is not a member of instead, which is what a capability SID is
in production, and roll back afterwards.
2026-09-12 09:46:12 +05:30
Vasanthdev2004 ad78725568 fix(sandbox): reject the reparse ancestor itself before materializing
The pre-create check asked where the deepest existing ancestor lives, and
a junction answers with its own path: the open does not follow a
final-component reparse point, so <root>/.git came back as <root>/.git and
matched. os.MkdirAll does follow it. What disqualifies that ancestor is
that it IS a reparse point, not where it reports living.

Found by reverting the check and watching the test still pass, which said
the guard was doing nothing rather than that the test was weak. Pinned
against the function now, because through the whole apply the create is
made and then removed on the failure path, so the filesystem afterwards
looks identical either way. The plan wiring gets its own test for the same
reason: every apply-level case here hands the group an anchor directly, so
none of them would notice the builder never setting one.
2026-09-12 09:46:12 +05:30
Vasanthdev2004 a9aa37e267 fix(sandbox): keep a derived ACL target inside the write root it came from
FILE_FLAG_OPEN_REPARSE_POINT refuses a reparse point at the final path
component and resolves every component above it, which is right for a path
the operator named and not enough for one this package derived. The
write-root carveouts are derived: <root>/.git/hooks and <root>/.git/config
are constructed from a root, and <root>/.git is a name an unprivileged
workspace writer can create before setup runs. mklink /J needs no
privilege, so a junction there had the apply open <junction-target>/hooks,
an ordinary directory with nothing wrong about its final component, and
zero sandbox setup wrote a deny ACE on it as Administrator, outside the
workspace.

A derived entry now carries the root it came from, and the apply requires
the object it finally holds to still live under it. The check is on the
handle rather than the name: GetFinalPathNameByHandle answers where the
open object actually is, and both sides go through it so the spelling
normalizes the same way. Materialization is checked before it creates,
since os.MkdirAll follows the same reparse points.

Deliberately not a refusal of every reparse point on the path. Above the
write root the path is the operator's, who may keep a workspace under a
junction or a mapped directory; only the tail the sandbox appended is held
strict.

Closes #1024
2026-09-12 09:46:12 +05:30
14 changed files with 711 additions and 2447 deletions

No files matched your search

+2 -3
View File
@@ -302,9 +302,8 @@ Conceptually, one requested tool goes through these phases:
the model can inspect it and decide what to do next.
6. **Render and persist** — the result is surfaced to the TUI, stream-JSON
writer, or ACP client. TUI and exec record rich tool/session events for
resume/history; ACP persists conversational messages and completed tool-call
starts/results for load replay. Live streaming details and permission
exchanges remain client notifications rather than durable history.
resume/history; ACP currently persists conversational messages and streams
tool/permission updates to the client.
Tool success and failure are both informative context. A successful `read_file`
result gives the model file contents; a failed `edit_file` result gives the model
+54 -546
View File
@@ -2,14 +2,10 @@ package acp
import (
"context"
"crypto/sha256"
"encoding/base64"
"encoding/json"
"errors"
"fmt"
"log"
"os"
"path/filepath"
"strings"
"sync"
@@ -44,12 +40,8 @@ type Deps struct {
// ResolveWorkspaceRoot validates + normalizes a client-supplied cwd (must be an
// existing directory; never the bare root). It is the file-tool confinement root.
ResolveWorkspaceRoot func(cwd string) (string, error)
// PersistEvents is an optional test seam for injecting event-store failures.
// It receives a COMPLETED TURN as one batch, never one event at a time: the
// batch is the unit the shared store keeps contiguous (see persistTurn).
PersistEvents func(sessionID string, inputs []sessions.AppendEventInput) error
Store *sessions.Store
AgentInfo Implementation
Store *sessions.Store
AgentInfo Implementation
}
// Workspace is the per-turn execution environment passed to the agent. ACP does
@@ -114,8 +106,6 @@ func NewAgent(conn *Conn, deps Deps) *Agent {
conn.Handle(MethodInitialize, a.handleInitialize)
conn.Handle(MethodSessionNew, a.handleSessionNew)
conn.Handle(MethodSessionLoad, a.handleSessionLoad)
conn.Handle(MethodSessionList, a.handleSessionList)
conn.Handle(MethodSessionResume, a.handleSessionResume)
conn.Handle(MethodSessionPrompt, a.handleSessionPrompt)
conn.Handle(MethodSessionSetMode, a.handleSetMode)
conn.Handle(MethodSessionSetConfigOption, a.handleSetConfigOption)
@@ -144,11 +134,11 @@ func (a *Agent) handleInitialize(_ context.Context, params json.RawMessage) (any
return InitializeResult{
ProtocolVersion: negotiated,
AgentCapabilities: AgentCapabilities{
// ACP v1 optional methods are advertised as empty capability objects;
// clients must gate session/list and session/resume on their presence.
LoadSession: true,
PromptCapabilities: PromptCapabilities{Image: true},
SessionCapabilities: &SessionCapabilities{List: &struct{}{}, Resume: &struct{}{}},
// Only advertise what ZERO actually implements: session/load (loadSession)
// and image prompts. session/resume + the session-capability sub-object
// are intentionally omitted since there is no resume handler yet.
LoadSession: true,
PromptCapabilities: PromptCapabilities{Image: true},
},
AgentInfo: &info,
// ZERO owns credentials (BYOK) and does not delegate auth to the editor.
@@ -158,41 +148,12 @@ func (a *Agent) handleInitialize(_ context.Context, params json.RawMessage) (any
// ---- session lifecycle ----
// requestedWorkspace is the single rule for a cwd that arrived from the client,
// applied before the value can reach ResolveWorkspaceRoot.
//
// THE REQUEST'S OWN CWD IS THE ONLY THING ALLOWED TO PICK A ROOT. Every params
// type spells cwd as a plain string, so an absent field and an empty one decode
// to exactly the same "", and the resolver treats "" as "use the directory this
// process was started in" and joins a relative cwd onto it. Those two defaults
// meet in the middle: `{"sessionId":"known"}` used to activate a persisted
// session, and `{}` used to create one, both rooted at wherever the editor
// happened to spawn `zero acp` — which then becomes the sandbox and file-tool
// confinement root. The client never named that directory, so it is a root the
// server invented, and ACP requires an absolute cwd on every method carrying
// one. Rejecting here is what keeps a missing field from being answered with a
// guess.
func requestedWorkspace(cwd string) (string, error) {
trimmed := strings.TrimSpace(cwd)
if trimmed == "" {
return "", RPCError(codeInvalidParams, "cwd is required and must be an absolute path")
}
if !filepath.IsAbs(trimmed) {
return "", RPCError(codeInvalidParams, "cwd must be an absolute path: "+trimmed)
}
return trimmed, nil
}
func (a *Agent) handleSessionNew(ctx context.Context, params json.RawMessage) (any, error) {
var p NewSessionParams
if err := json.Unmarshal(params, &p); err != nil {
return nil, RPCError(codeInvalidParams, "invalid session/new params")
}
cwd, err := requestedWorkspace(p.Cwd)
if err != nil {
return nil, err
}
root, err := a.deps.ResolveWorkspaceRoot(cwd)
root, err := a.deps.ResolveWorkspaceRoot(p.Cwd)
if err != nil {
return nil, RPCError(codeInvalidParams, err.Error())
}
@@ -204,7 +165,7 @@ func (a *Agent) handleSessionNew(ctx context.Context, params json.RawMessage) (a
if err != nil {
return nil, RPCError(codeInternalError, "create session: "+err.Error())
}
sess, _ := a.registerSession(meta.SessionID, root, nil, model, models, restrictModels)
sess := a.registerSession(meta.SessionID, root, nil, model, models, restrictModels)
return NewSessionResult{
SessionID: sess.id,
ConfigOptions: a.configOptions(sess),
@@ -217,90 +178,22 @@ func (a *Agent) handleSessionLoad(ctx context.Context, params json.RawMessage) (
if err := json.Unmarshal(params, &p); err != nil {
return nil, RPCError(codeInvalidParams, "invalid session/load params")
}
return a.activatePersistedSession(ctx, p, persistedSessionLoad)
}
func (a *Agent) handleSessionResume(ctx context.Context, params json.RawMessage) (any, error) {
var p ResumeSessionParams
if err := json.Unmarshal(params, &p); err != nil {
return nil, RPCError(codeInvalidParams, "invalid session/resume params")
}
return a.activatePersistedSession(ctx, p, persistedSessionResume)
}
type persistedSessionOperation uint8
const (
persistedSessionLoad persistedSessionOperation = iota
persistedSessionResume
)
// activatePersistedSession restores the agent's internal conversation context
// for both lifecycle methods. session/load additionally replays user-visible
// history as ordered session/update notifications; session/resume deliberately
// does not, which makes it safe for an already-rendered desktop reconnect.
func (a *Agent) activatePersistedSession(ctx context.Context, p LoadSessionParams, operation persistedSessionOperation) (any, error) {
// BOTH METHODS, BEFORE ANYTHING IS LOOKED UP. session/resume shares this
// params type with session/load, so the wire cannot tell an omitted cwd from
// an empty one; the blank case used to be answered with the persisted cwd,
// which made {"sessionId":"known"} a complete, successful activation request.
// Validating here rather than in one handler is what keeps the two entry
// points from drifting apart.
cwd, err := requestedWorkspace(p.Cwd)
if err != nil {
return nil, err
}
meta, err := a.deps.Store.Get(p.SessionID)
if err != nil || meta == nil {
return nil, RPCError(codeInvalidParams, "session not found: "+p.SessionID)
}
if operation == persistedSessionResume && !sessions.IsResumableKind(meta.SessionKind) {
return nil, RPCError(codeInvalidParams, "session is not resumable: "+p.SessionID)
cwdInput := p.Cwd
if strings.TrimSpace(cwdInput) == "" {
cwdInput = meta.Cwd
}
if strings.TrimSpace(meta.Cwd) == "" {
return nil, RPCError(codeInvalidParams, "session has no persisted workspace: "+p.SessionID)
}
// SAME RULE ON THE WAY IN. Omitting a relative entry from the listing is not
// enough: a stored "." still resolves against this process's directory, so a
// client that hands back that same directory as its cwd would be sold the
// invented root as a match. A workspace that cannot be identified is not one
// this session can be restored into.
if !filepath.IsAbs(meta.Cwd) {
return nil, RPCError(codeInvalidParams, "session workspace is not an absolute path, so it cannot be identified: "+p.SessionID)
}
persistedRoot, err := a.deps.ResolveWorkspaceRoot(meta.Cwd)
if err != nil {
return nil, RPCError(codeInvalidParams, "persisted session workspace is unavailable: "+err.Error())
}
root, err := a.deps.ResolveWorkspaceRoot(cwd)
root, err := a.deps.ResolveWorkspaceRoot(cwdInput)
if err != nil {
return nil, RPCError(codeInvalidParams, err.Error())
}
// ACP session cwd is immutable. Loading history under a different root
// would give a conversation from one workspace access to another
// workspace's configuration, files, and tools.
if !sameWorkspace(root, persistedRoot) {
return nil, RPCError(codeInvalidParams, "session cwd does not match its persisted workspace")
}
// Load history BEFORE publishing the session so no concurrent prompt observes
// a half-initialized session (registerSession sets history under the lock and
// reuses an already-live session rather than orphaning its in-flight turn).
requireHistoryLog := operation == persistedSessionResume && meta.EventCount > 0
history, messages, historyWarning, historyErr := a.loadHistory(meta.SessionID, requireHistoryLog)
// RESUME PROMISES RESTORED CONTEXT, SO A FAILED RESTORATION IS A FAILED
// RESUME. historyErr used to do nothing but suppress replay and raise a
// warning: the session was registered and reported ready regardless, so a
// corrupt record or an unreadable events file left the caller holding a live,
// promptable session ID whose next prompt ran as a brand-new conversation
// under the old identity. Nothing in the response said so.
//
// Publication is now gated on restoration for resume. LOAD keeps the
// best-effort policy deliberately rather than by inheriting this helper: it
// replays what it has and warns about the rest, which is the right trade for
// a client rebuilding a transcript it can still scroll. Reported by @jatmn.
if operation == persistedSessionResume && historyErr != nil {
return nil, RPCError(codeInternalError, "restore session history: "+historyErr.Error())
}
history, historyErr := a.loadHistory(meta.SessionID)
model, models, restrictModels, err := a.resolveModelChoices(ctx, root)
if err != nil {
return nil, RPCError(codeInternalError, "config: "+err.Error())
@@ -311,36 +204,12 @@ func (a *Agent) activatePersistedSession(ctx context.Context, p LoadSessionParam
models = append(models, SessionConfigOptionValue{Value: persistedModel, Name: persistedModel})
}
}
sess, existed := a.registerSession(meta.SessionID, root, history, model, models, restrictModels)
if existed {
_, messages, historyWarning, historyErr = a.refreshSessionHistory(sess, requireHistoryLog)
if operation == persistedSessionResume && historyErr != nil {
return nil, RPCError(codeInternalError, "restore session history: "+historyErr.Error())
}
}
note := &notifier{conn: a.conn, sessionID: sess.id}
if operation == persistedSessionLoad && historyErr == nil {
for _, message := range messages {
if message.tool != nil {
note.send(*message.tool)
continue
}
note.send(replayMessageChunk(message.role, replayMessageID(message.eventID), message.content))
}
}
sess := a.registerSession(meta.SessionID, root, history, model, models, restrictModels)
a.warnPersistence(
note,
"rehydrate session compaction",
"Could not apply session compaction. Raw session history was restored instead.",
historyWarning,
operation == persistedSessionLoad,
)
a.warnPersistence(
note,
&notifier{conn: a.conn, sessionID: sess.id},
"load session history",
"Could not load session history. The session is open, but earlier turns may be missing until storage recovers.",
historyErr,
operation == persistedSessionLoad,
)
return LoadSessionResult{
ConfigOptions: a.configOptions(sess),
@@ -348,80 +217,6 @@ func (a *Agent) activatePersistedSession(ctx context.Context, p LoadSessionParam
}, nil
}
func (a *Agent) handleSessionList(_ context.Context, params json.RawMessage) (any, error) {
var p ListSessionsParams
if len(params) > 0 {
if err := json.Unmarshal(params, &p); err != nil {
return nil, RPCError(codeInvalidParams, "invalid session/list params")
}
}
if p.Cursor != "" {
return nil, RPCError(codeInvalidParams, "invalid session/list cursor")
}
items, err := a.deps.Store.ListResumable()
if err != nil {
return nil, RPCError(codeInternalError, "list sessions: "+err.Error())
}
// The filter is the one cwd ACP leaves optional, so a blank one means "no
// filter" rather than an error — but a relative one is rebased onto this
// process's directory exactly like the others, and would then silently answer
// about a workspace the client never named.
var cwd string
if strings.TrimSpace(p.Cwd) != "" {
filter, err := requestedWorkspace(p.Cwd)
if err != nil {
return nil, err
}
cwd, err = a.deps.ResolveWorkspaceRoot(filter)
if err != nil {
return nil, RPCError(codeInvalidParams, err.Error())
}
}
result := ListSessionsResult{Sessions: make([]SessionInfo, 0, len(items))}
for _, item := range items {
// EVERY ENTRY IS RESOLVED, FILTER OR NO FILTER. Listing is a menu, and
// every item on it has to be orderable: activatePersistedSession
// resolves the persisted workspace and refuses what it cannot reach, so
// anything this loop cannot resolve is something the client would be
// offered and then denied.
//
// Resolving only when a cwd filter was supplied left two shapes through —
// a session whose workspace has since been deleted, and a legacy entry
// holding a relative path, which was then reported as cwd "." even though
// ACP requires SessionInfo.cwd to be absolute.
if strings.TrimSpace(item.Cwd) == "" {
continue
}
// A RELATIVE PERSISTED CWD HAS NO RECOVERABLE IDENTITY. Resolving one
// rebases it onto wherever this ACP server happens to be running and
// advertises that invented absolute path as the session's workspace — so a
// conversation created for one project could be resumed against another
// project's configuration, files and tools. The original base is not
// knowable from the metadata, so the honest answer is to omit the entry
// rather than to guess at it.
if !filepath.IsAbs(item.Cwd) {
continue
}
itemRoot, err := a.deps.ResolveWorkspaceRoot(item.Cwd)
if err != nil {
continue
}
if cwd != "" && !sameWorkspace(itemRoot, cwd) {
continue
}
result.Sessions = append(result.Sessions, SessionInfo{
SessionID: item.SessionID,
Title: item.Title,
// The RESOLVED root, not the stored string: absolute as ACP
// requires, and the same value the client will hand back on resume.
Cwd: itemRoot,
UpdatedAt: item.UpdatedAt,
Meta: &SessionInfoMeta{ModelID: item.ModelID, CreatedAt: item.CreatedAt},
})
}
return result, nil
}
// ---- prompt turn ----
func (a *Agent) handleSessionPrompt(ctx context.Context, params json.RawMessage) (any, error) {
@@ -498,15 +293,6 @@ func (a *Agent) runTurn(ctx context.Context, sess *acpSession, userText string,
note.text("\n\n[zero warning] " + text + "\n")
}
}
// Buffer the complete turn until RunAgent has a valid ACP outcome. A hard run
// failure is an aborted turn: neither live history nor durable history may
// retain its user/tool prefix under an RPC error. StopCancelled is a valid
// outcome and commits through the same boundary below.
pendingEvents := make([]sessions.AppendEventInput, 0, 4)
queue := func(input sessions.AppendEventInput) {
pendingEvents = append(pendingEvents, input)
}
queue(messageEvent("user", userText))
opts := agent.Options{
Cwd: sess.cwd,
@@ -521,12 +307,8 @@ func (a *Agent) runTurn(ctx context.Context, sess *acpSession, userText string,
Images: images,
OnText: note.text,
OnReasoning: note.thought,
OnToolCall: func(call agent.ToolCall) {
queue(toolCallEvent(call))
note.toolCall(call)
},
OnToolCall: note.toolCall,
OnToolResult: func(result agent.ToolResult) {
queue(toolResultEvent(result))
note.toolResult(result)
if result.Name == "update_plan" {
a.emitPlan(registry, note)
@@ -539,36 +321,19 @@ func (a *Agent) runTurn(ctx context.Context, sess *acpSession, userText string,
agentPrompt := buildPrompt(sess.snapshotHistory(), userText)
result, runErr := a.deps.RunAgent(ctx, agentPrompt, provider, opts)
if result.FinalAnswer != "" {
queue(messageEvent("assistant", result.FinalAnswer))
}
reason, stopErr := stopReasonFor(result, runErr)
if stopErr != nil {
return "", RPCError(codeInternalError, stopErr.Error())
}
// ONE BATCH, ONE LOCK. The outcome gate above decides WHETHER the turn is
// committed; the store batch decides WHICH RECORDS STAY TOGETHER relative to
// other writers, and these are different guarantees. Persisting the buffer
// one event per call took and released the store's session lock between
// events, so a second ACP instance holding the same session — its turnMu is
// its own, only the store is shared — could append its whole turn in the
// gap: user A, user B, answer B, answer A. Every write was individually
// locked and the log was valid, but a fresh load pairs prose by order and
// reconstructed three wrong turns from two right ones. Before this branch,
// persistTurn handed the user and assistant events to AppendEvents together;
// buffering until the outcome is known was correct, splitting the commit
// into singleton writes was not. The store writes a batch under one lock and
// one file write, which is the contiguity restored here — not a disk
// transaction, and not a promise about crashes mid-fsync. Reported by @jatmn.
persistenceErr := a.persistTurn(sess.id, pendingEvents)
sess.appendHistory(turnRecord{user: userText, assistant: result.FinalAnswer})
a.warnPersistence(
note,
"save session history",
"Could not save session history. This turn is available in memory, but future resume may miss it until storage recovers.",
persistenceErr,
true,
)
if err := a.persistTurn(sess, userText, result.FinalAnswer); err != nil {
a.warnPersistence(
note,
"save session history",
"Could not save session history. This turn is available in memory, but future resume may miss it until storage recovers.",
err,
)
}
return reason, nil
}
@@ -827,159 +592,39 @@ func (a *Agent) configOptions(s *acpSession) []SessionConfigOption {
// ---- persistence + continuity ----
// persistTurn commits a completed turn's buffered events as one store batch, so
// no other writer's records can land between them. A failure is the failure of
// the whole batch: nothing after a failed prerequisite is appended on its own.
func (a *Agent) persistTurn(sessionID string, inputs []sessions.AppendEventInput) error {
if len(inputs) == 0 {
return nil
}
if a.deps.PersistEvents != nil {
return a.deps.PersistEvents(sessionID, inputs)
}
func (a *Agent) persistTurn(sess *acpSession, user, assistant string) error {
defer sess.appendHistory(turnRecord{user: user, assistant: assistant})
if a.deps.Store == nil {
return nil
}
_, err := a.deps.Store.AppendEvents(sessionID, inputs)
events := []sessions.AppendEventInput{
{
Type: sessions.EventMessage,
Payload: map[string]any{"role": "user", "content": user},
},
}
if assistant != "" {
events = append(events, sessions.AppendEventInput{
Type: sessions.EventMessage,
Payload: map[string]any{"role": "assistant", "content": assistant},
})
}
_, err := a.deps.Store.AppendEvents(sess.id, events)
return err
}
func messageEvent(role, content string) sessions.AppendEventInput {
return sessions.AppendEventInput{
Type: sessions.EventMessage,
Payload: map[string]any{"role": role, "content": content},
}
}
func toolCallEvent(call agent.ToolCall) sessions.AppendEventInput {
return sessions.AppendEventInput{
Type: sessions.EventToolCall,
Payload: map[string]any{
"toolCallId": call.ID,
"name": call.Name,
"arguments": call.Arguments,
},
}
}
func toolResultEvent(result agent.ToolResult) sessions.AppendEventInput {
return sessions.AppendEventInput{
Type: sessions.EventToolResult,
Payload: map[string]any{
"toolCallId": result.ToolCallID,
"name": result.Name,
"status": result.Status,
"output": result.Output,
"changedFiles": append([]string(nil), result.ChangedFiles...),
},
}
}
// persistedMessage is one entry of the restored transcript in stored order. It
// is either a message (role/content) or a tool-call notification (tool), never
// both: keeping them in ONE ordered slice is what preserves the interleaving a
// client needs to redraw the conversation as it originally happened.
type persistedMessage struct {
eventID string
role string
content string
// tool is set for a replayed tool call or its result, and is already in wire
// shape — it comes from the same toolCallStart/toolCallResult mapping the
// live turn uses, so a replayed tool call renders identically to a fresh one
// and there is no second mapping to drift out of sync.
tool *ToolCallUpdate
}
func (a *Agent) loadHistory(sessionID string, requireHistoryLog bool) ([]turnRecord, []persistedMessage, error, error) {
func (a *Agent) loadHistory(sessionID string) ([]turnRecord, error) {
if a.deps.Store == nil {
return nil, nil, nil, nil
return nil, nil
}
// THE EFFECTIVE CONVERSATION, NOT THE RAW LOG. A compacted session keeps its
// original prefix on disk alongside an EventCompaction that names the events
// it replaced and carries their durable summary. ReadEvents hands back both,
// so restoring from it replayed superseded turns AND dropped the summary that
// replaced them — resume then seeded the next turn with context the
// conversation had already moved past, and load visibly resurrected
// transcript entries the user had seen compacted away.
//
// ReadRehydratedEvents is the projection the TUI and exec paths already use
// (internal/tui/session.go, internal/sessions/exec_session.go), so all three
// now agree on what the conversation IS. Switching readers alone is not
// enough: rehydration substitutes the compaction event in place of the events
// it replaced, so a loop that skips everything but EventMessage would drop the
// summary exactly as before. It is projected below. Reported by @jatmn.
events, eventLogPresent, err := a.deps.Store.ReadRehydratedEventsWithPresence(sessionID)
var rehydrateWarning error
events, err := a.deps.Store.ReadEvents(sessionID)
if err != nil {
rehydrateWarning = err
events, eventLogPresent, err = a.deps.Store.ReadEventsWithPresence(sessionID)
if err != nil {
return nil, nil, nil, err
}
}
if requireHistoryLog && !eventLogPresent {
return nil, nil, nil, fmt.Errorf("zero session %s event log is missing", sessionID)
return nil, err
}
var records []turnRecord
var messages []persistedMessage
activeToolCalls := make(map[string]string)
var pendingUser string
havePending := false
for _, e := range events {
if e.Type == sessions.EventCompaction {
// The summary stands in for the turns it replaced, so it enters the
// history as an assistant message: it is prior context the next turn
// must see, and the wire has no separate role for it.
raw, marshalErr := json.Marshal(e.Payload)
if marshalErr != nil {
continue
}
var payload struct {
Summary string `json:"summary"`
}
if json.Unmarshal(raw, &payload) != nil || strings.TrimSpace(payload.Summary) == "" {
continue
}
messages = append(messages, persistedMessage{
eventID: persistedMessageIdentity(sessionID, e),
role: "assistant",
content: payload.Summary,
})
records = append(records, turnRecord{user: pendingUser, assistant: payload.Summary})
pendingUser = ""
havePending = false
continue
}
// TOOL ACTIVITY IS PART OF THE TRANSCRIPT, NOT DECORATION. A restored
// conversation that shows only the prose reads as if the agent asserted
// its edits and command output from nowhere: the user sees "I updated
// scope.go" with no record that any tool ran. These are collected for
// replay only. They deliberately do NOT enter turnRecord, so the prompt
// context a resumed turn receives stays byte-identical to what it was —
// load and resume continue to consume the SAME effective history, and
// only the wire rendering differs. Reported by @jatmn.
if e.Type == sessions.EventToolCall || e.Type == sessions.EventToolResult {
if upd := replayToolUpdate(e); upd != nil {
rawID := upd.ToolCallID
if e.Type == sessions.EventToolResult {
replayID, ok := activeToolCalls[rawID]
if !ok {
continue
}
upd.ToolCallID = replayID
delete(activeToolCalls, rawID)
} else {
replayID := replayToolCallID(persistedMessageIdentity(sessionID, e))
activeToolCalls[rawID] = replayID
upd.ToolCallID = replayID
}
messages = append(messages, persistedMessage{
eventID: persistedMessageIdentity(sessionID, e),
tool: upd,
})
}
continue
}
if e.Type != sessions.EventMessage {
continue
}
@@ -996,14 +641,12 @@ func (a *Agent) loadHistory(sessionID string, requireHistoryLog bool) ([]turnRec
}
switch msg.Role {
case "user":
messages = append(messages, persistedMessage{eventID: persistedMessageIdentity(sessionID, e), role: msg.Role, content: msg.Content})
if havePending {
records = append(records, turnRecord{user: pendingUser})
}
pendingUser = msg.Content
havePending = true
case "assistant":
messages = append(messages, persistedMessage{eventID: persistedMessageIdentity(sessionID, e), role: msg.Role, content: msg.Content})
records = append(records, turnRecord{user: pendingUser, assistant: msg.Content})
pendingUser = ""
havePending = false
@@ -1012,89 +655,10 @@ func (a *Agent) loadHistory(sessionID string, requireHistoryLog bool) ([]turnRec
if havePending {
records = append(records, turnRecord{user: pendingUser})
}
return records, messages, rehydrateWarning, nil
return records, nil
}
// replayToolUpdate rebuilds the ACP notification for one stored tool event.
//
// The stored toolCallId is returned here as a correlation key. loadHistory maps
// each call occurrence to a stable session-wide replay identity and applies that
// identity to its result. Older records wrote the field as "id", so both
// spellings are accepted — the TUI projection already does the same
// (internal/tui/session.go).
//
// A record that carries no usable id is skipped rather than replayed under a
// synthesized one: an unpairable update is worse than a missing row.
func replayToolUpdate(event sessions.Event) *ToolCallUpdate {
raw, err := json.Marshal(event.Payload)
if err != nil {
return nil
}
var payload struct {
Name string `json:"name"`
ToolCallID string `json:"toolCallId"`
ID string `json:"id"`
Arguments string `json:"arguments"`
Status string `json:"status"`
Output string `json:"output"`
ChangedFiles []string `json:"changedFiles"`
}
if json.Unmarshal(raw, &payload) != nil {
return nil
}
id := payload.ToolCallID
if id == "" {
id = payload.ID
}
if id == "" {
return nil
}
if event.Type == sessions.EventToolCall {
upd := toolCallStart(agent.ToolCall{ID: id, Name: payload.Name, Arguments: payload.Arguments})
return &upd
}
status := tools.Status(payload.Status)
if status == "" {
status = tools.StatusOK
}
upd := toolCallResult(agent.ToolResult{
ToolCallID: id,
Name: payload.Name,
Status: status,
Output: payload.Output,
ChangedFiles: append([]string(nil), payload.ChangedFiles...),
})
return &upd
}
func persistedMessageIdentity(sessionID string, event sessions.Event) string {
if event.ID != "" {
return event.ID
}
return fmt.Sprintf("%s:%d", sessionID, event.Sequence)
}
// replayMessageID maps ZERO's stable event identity to a standards-shaped UUID
// without leaking or parsing the event id on the wire. The same stored message
// receives the same opaque id across loads, which also gives clients an exact
// chunk boundary when two adjacent persisted messages have the same role.
func replayMessageID(eventID string) string {
return replayOpaqueID("message", eventID)
}
func replayToolCallID(eventID string) string {
return replayOpaqueID("tool-call", eventID)
}
func replayOpaqueID(kind, eventID string) string {
sum := sha256.Sum256([]byte("zero-acp-" + kind + ":" + eventID))
b := sum[:16]
b[6] = (b[6] & 0x0f) | 0x50
b[8] = (b[8] & 0x3f) | 0x80
return fmt.Sprintf("%x-%x-%x-%x-%x", b[:4], b[4:6], b[6:8], b[8:10], b[10:16])
}
func (a *Agent) warnPersistence(note *notifier, action string, message string, err error, notify bool) {
func (a *Agent) warnPersistence(note *notifier, action string, message string, err error) {
if err == nil {
return
}
@@ -1103,7 +667,7 @@ func (a *Agent) warnPersistence(note *notifier, action string, message string, e
sessionID = note.sessionID
}
log.Printf("zero acp: failed to %s for session %s: %v", action, sessionID, err)
if note != nil && notify {
if note != nil {
note.text("\n\n[zero warning] " + message + "\n")
}
}
@@ -1150,34 +714,18 @@ func promptImages(blocks []ContentBlock) []zeroruntime.ImageBlock {
// registerSession publishes a session under the agent's lock. If one is already
// registered for id (e.g. a re-load of an in-flight session) the existing live
// session is returned rather than orphaning its turn or resetting its mode/model.
// The bool reports that case so persisted activation can re-read and apply disk
// history under turnMu. history is set BEFORE first publication so no concurrent
// prompt can read a half-initialized session.
func (a *Agent) registerSession(id, cwd string, history []turnRecord, model string, models []SessionConfigOptionValue, restrictModels bool) (*acpSession, bool) {
// session is returned unchanged rather than orphaning its turn or resetting its
// mode/model. history is set BEFORE publishing so no concurrent prompt can read a
// half-initialized session.
func (a *Agent) registerSession(id, cwd string, history []turnRecord, model string, models []SessionConfigOptionValue, restrictModels bool) *acpSession {
a.mu.Lock()
defer a.mu.Unlock()
if existing := a.sessions[id]; existing != nil {
return existing, true
return existing
}
sess := &acpSession{id: id, cwd: cwd, mode: agent.PermissionModeAuto, model: model, models: models, restrictModels: restrictModels, history: history}
a.sessions[id] = sess
return sess, false
}
// refreshSessionHistory is the activation apply boundary for an already-live
// session. The history is read while prompts are excluded, then applied to the
// same object the prompt path uses. Re-reading here avoids overwriting a turn
// that committed between activatePersistedSession's pre-publication read and
// discovering that another activation had already registered the session.
func (a *Agent) refreshSessionHistory(sess *acpSession, requireHistoryLog bool) ([]turnRecord, []persistedMessage, error, error) {
sess.turnMu.Lock()
defer sess.turnMu.Unlock()
history, messages, warning, err := a.loadHistory(sess.id, requireHistoryLog)
if err == nil {
sess.setHistory(history)
}
return history, messages, warning, err
return sess
}
func (a *Agent) session(id string) *acpSession {
@@ -1260,48 +808,8 @@ func (s *acpSession) appendHistory(rec turnRecord) {
s.mu.Unlock()
}
func (s *acpSession) setHistory(history []turnRecord) {
s.mu.Lock()
s.history = append([]turnRecord(nil), history...)
s.mu.Unlock()
}
func (s *acpSession) snapshotHistory() []turnRecord {
s.mu.Lock()
defer s.mu.Unlock()
return append([]turnRecord(nil), s.history...)
}
// sameWorkspace reports whether two resolved roots name the same directory.
//
// STRING EQUALITY IS NOT DIRECTORY IDENTITY. ResolveWorkspaceRoot is abs plus
// filepath.Clean plus a stat: it does not fold case and does not resolve
// junctions, so one directory reached by two spellings produces two different
// roots. A session persisted from the TUI was then unresumable from an editor
// holding a different spelling of the same project folder, and session/list
// filtered by the other spelling returned nothing — not a failed resume but an
// invisible one, on exactly the case this feature exists for.
//
// filepath.EvalSymlinks is NOT the fix on Windows: it normalises a drive letter
// but returns a junction path unchanged, so the alias survives it. Junctions need
// no privilege, so this is ordinary rather than exotic. os.SameFile compares the
// filesystem's own identity for the two directories, which is the question being
// asked.
//
// The string comparison stays as the fast path, and a stat failure falls back to
// it rather than widening the match — this gate refuses access to another
// workspace's files and configuration, so an unanswerable comparison denies.
func sameWorkspace(left, right string) bool {
if left == right {
return true
}
leftInfo, err := os.Stat(left)
if err != nil {
return false
}
rightInfo, err := os.Stat(right)
if err != nil {
return false
}
return os.SameFile(leftInfo, rightInfo)
}
+13 -1684
View File
File diff suppressed because it is too large. Load diff
+10 -43
View File
@@ -89,13 +89,6 @@ type Conn struct {
handlers map[string]HandlerFunc
notifiers map[string]NotifyFunc
// Notifications for one method are an ordered stream: session/update in
// particular is stateful, so dispatching every frame in an unrelated
// goroutine can make a later chunk overtake an earlier one. Different
// methods retain independent tails, keeping session/cancel responsive while
// an update handler is busy.
notifyMu sync.Mutex
notifyTails map[string]chan struct{}
mu sync.Mutex
nextID int64
@@ -113,12 +106,11 @@ type Conn struct {
// lifetime.
func NewConn(r io.Reader, w io.Writer) *Conn {
return &Conn{
rawReader: r,
w: w,
handlers: make(map[string]HandlerFunc),
notifiers: make(map[string]NotifyFunc),
notifyTails: make(map[string]chan struct{}),
pending: make(map[int64]chan rpcMessage),
rawReader: r,
w: w,
handlers: make(map[string]HandlerFunc),
notifiers: make(map[string]NotifyFunc),
pending: make(map[int64]chan rpcMessage),
}
}
@@ -311,7 +303,11 @@ func (c *Conn) handleLine(ctx context.Context, line []byte) {
}(msg)
case msg.isNotify():
if fn := c.notifiers[msg.Method]; fn != nil {
c.dispatchNotification(ctx, msg, fn)
c.wg.Add(1)
go func(m rpcMessage) {
defer c.wg.Done()
fn(ctx, m.Params)
}(msg)
}
default:
// Malformed frame; reply only if we can identify a request id.
@@ -321,35 +317,6 @@ func (c *Conn) handleLine(ctx context.Context, line []byte) {
}
}
// dispatchNotification preserves wire order within one method without making
// unrelated notification methods wait for each other. The read loop installs
// each tail before starting its goroutine, so goroutine scheduling cannot
// reorder the chain it observes.
func (c *Conn) dispatchNotification(ctx context.Context, msg rpcMessage, fn NotifyFunc) {
c.notifyMu.Lock()
previous := c.notifyTails[msg.Method]
done := make(chan struct{})
c.notifyTails[msg.Method] = done
c.notifyMu.Unlock()
c.wg.Add(1)
go func() {
defer c.wg.Done()
defer func() {
close(done)
c.notifyMu.Lock()
if c.notifyTails[msg.Method] == done {
delete(c.notifyTails, msg.Method)
}
c.notifyMu.Unlock()
}()
if previous != nil {
<-previous
}
fn(ctx, msg.Params)
}()
}
func (c *Conn) dispatchRequest(ctx context.Context, msg rpcMessage) {
fn := c.handlers[msg.Method]
if fn == nil {
-80
View File
@@ -360,86 +360,6 @@ func TestConnNotification(t *testing.T) {
}
}
func TestConnNotificationsForOneMethodStayInWireOrder(t *testing.T) {
a, b, stop := connPair(t)
defer stop()
firstStarted := make(chan struct{})
releaseFirst := make(chan struct{})
delivered := make(chan int, 2)
b.HandleNotify("update", func(_ context.Context, params json.RawMessage) {
var in struct{ Sequence int }
_ = json.Unmarshal(params, &in)
if in.Sequence == 1 {
close(firstStarted)
<-releaseFirst
}
delivered <- in.Sequence
})
if err := a.Notify("update", map[string]int{"Sequence": 1}); err != nil {
t.Fatalf("first notify: %v", err)
}
select {
case <-firstStarted:
case <-time.After(2 * time.Second):
t.Fatal("first notification did not start")
}
if err := a.Notify("update", map[string]int{"Sequence": 2}); err != nil {
t.Fatalf("second notify: %v", err)
}
select {
case sequence := <-delivered:
t.Fatalf("notification %d overtook the blocked first notification", sequence)
case <-time.After(25 * time.Millisecond):
}
close(releaseFirst)
for want := 1; want <= 2; want++ {
select {
case got := <-delivered:
if got != want {
t.Fatalf("notification order = ...,%d; want %d", got, want)
}
case <-time.After(2 * time.Second):
t.Fatalf("notification %d was not delivered", want)
}
}
}
func TestConnDifferentNotificationMethodsStayConcurrent(t *testing.T) {
a, b, stop := connPair(t)
defer stop()
updateStarted := make(chan struct{})
releaseUpdate := make(chan struct{})
cancelDelivered := make(chan struct{})
b.HandleNotify("update", func(context.Context, json.RawMessage) {
close(updateStarted)
<-releaseUpdate
})
b.HandleNotify("cancel", func(context.Context, json.RawMessage) {
close(cancelDelivered)
})
if err := a.Notify("update", nil); err != nil {
t.Fatalf("update notify: %v", err)
}
select {
case <-updateStarted:
case <-time.After(2 * time.Second):
t.Fatal("update notification did not start")
}
if err := a.Notify("cancel", nil); err != nil {
t.Fatalf("cancel notify: %v", err)
}
select {
case <-cancelDelivered:
case <-time.After(2 * time.Second):
t.Fatal("different notification method was blocked behind update")
}
close(releaseUpdate)
}
func TestConnMethodNotFound(t *testing.T) {
a, _, stop := connPair(t)
defer stop()
-8
View File
@@ -19,14 +19,6 @@ func agentMessageChunk(delta string) ContentChunk {
return ContentChunk{SessionUpdate: UpdateAgentMessageChunk, Content: TextBlock(delta)}
}
func replayMessageChunk(role, messageID, text string) ContentChunk {
update := UpdateAgentMessageChunk
if role == "user" {
update = UpdateUserMessageChunk
}
return ContentChunk{SessionUpdate: update, MessageID: messageID, Content: TextBlock(text)}
}
func agentThoughtChunk(delta string) ContentChunk {
return ContentChunk{SessionUpdate: UpdateAgentThoughtChunk, Content: TextBlock(delta)}
}
+5 -54
View File
@@ -12,8 +12,6 @@ const (
MethodAuthenticate = "authenticate"
MethodSessionNew = "session/new"
MethodSessionLoad = "session/load"
MethodSessionList = "session/list"
MethodSessionResume = "session/resume"
MethodSessionPrompt = "session/prompt"
MethodSessionCancel = "session/cancel" // notification
MethodSessionUpdate = "session/update" // notification (agent -> client)
@@ -64,16 +62,8 @@ type PromptCapabilities struct {
}
type AgentCapabilities struct {
LoadSession bool `json:"loadSession"`
PromptCapabilities PromptCapabilities `json:"promptCapabilities"`
SessionCapabilities *SessionCapabilities `json:"sessionCapabilities,omitempty"`
}
// Empty capability objects are presence flags in ACP v1. Pointers preserve
// the wire distinction between an advertised `{}` and an omitted capability.
type SessionCapabilities struct {
List *struct{} `json:"list,omitempty"`
Resume *struct{} `json:"resume,omitempty"`
LoadSession bool `json:"loadSession"`
PromptCapabilities PromptCapabilities `json:"promptCapabilities"`
}
type AuthMethod struct {
@@ -127,14 +117,9 @@ type McpServer struct {
}
type NewSessionParams struct {
Cwd string `json:"cwd"`
McpServers []McpServer `json:"mcpServers"`
// AdditionalDirectories is declared by the protocol and consumed nowhere in
// this repo yet. When it is wired up it must go through requestedWorkspace
// like Cwd does: these are client-supplied paths with no absoluteness rule of
// their own, which is the same shape as the resume-cwd hole that made
// {"sessionId":"known"} activate a session against this process's directory.
AdditionalDirectories []string `json:"additionalDirectories,omitempty"`
Cwd string `json:"cwd"`
McpServers []McpServer `json:"mcpServers"`
AdditionalDirectories []string `json:"additionalDirectories,omitempty"`
}
type NewSessionResult struct {
@@ -155,39 +140,6 @@ type LoadSessionResult struct {
Modes *SessionModeState `json:"modes,omitempty"`
}
// ListSessionsParams and the following types implement ACP v1 session/list.
// Transcript contents stay behind session/load; this method returns metadata
// only and leaves the optional pagination cursor opaque.
type ListSessionsParams struct {
Cwd string `json:"cwd,omitempty"`
Cursor string `json:"cursor,omitempty"`
}
type SessionInfoMeta struct {
ModelID string `json:"modelId,omitempty"`
CreatedAt string `json:"createdAt,omitempty"`
}
type SessionInfo struct {
SessionID string `json:"sessionId"`
Cwd string `json:"cwd"`
Title string `json:"title,omitempty"`
UpdatedAt string `json:"updatedAt,omitempty"`
Meta *SessionInfoMeta `json:"_meta,omitempty"`
}
type ListSessionsResult struct {
Sessions []SessionInfo `json:"sessions"`
NextCursor string `json:"nextCursor,omitempty"`
}
// session/resume takes session/load's wire shape, so its cwd is a plain string
// too and an omitted one decodes exactly like an empty one. Sharing the type is
// safe only because activatePersistedSession requires an absolute cwd from the
// request itself — nothing downstream may supply a default for either method.
type ResumeSessionParams = LoadSessionParams
type ResumeSessionResult = LoadSessionResult
// ---- prompt turn ----
type PromptParams struct {
@@ -222,7 +174,6 @@ type SessionNotification struct {
// ContentBlock under "content"; the variant is set via SessionUpdate.
type ContentChunk struct {
SessionUpdate string `json:"sessionUpdate"`
MessageID string `json:"messageId,omitempty"`
Content ContentBlock `json:"content"`
}
+20 -2
View File
@@ -33,8 +33,14 @@ type WindowsACLEntry struct {
// down onto the target's EXISTING descendants (not just new ones it
// creates going forward), which is why direct-only denies must set this
// flag rather than rely on inheritance.
NoInherit bool `json:"noInherit,omitempty"`
Materialize bool `json:"materialize,omitempty"`
NoInherit bool `json:"noInherit,omitempty"`
// Anchor is the write root Path was DERIVED from, for the paths this
// package constructs rather than the operator naming. The apply requires
// the object it finally opens to still live under it, so a reparse point
// planted on the derived tail cannot walk an elevated ACL write out of the
// sandbox. Empty for an operator-named path, which has no owned tail.
Anchor string `json:"anchor,omitempty"`
Materialize bool `json:"materialize,omitempty"`
}
type WindowsACLPlan struct {
@@ -57,10 +63,22 @@ func BuildWindowsACLPlan(config WindowsSandboxCommandConfig) (WindowsACLPlan, er
Capability: capability.SID,
})
for _, path := range capability.ProtectedWriteDenyPaths {
// ANCHORED ONLY WHERE THE ANCHOR IS TRUE. Most of these are derived
// from the root and sit under it, and holding those inside it is the
// point. But ReadOnlySubpaths is a profile field an operator can set to
// any path, and one deliberately placed outside the root is a
// configuration that works today; anchoring it would turn that into a
// containment refusal. A path that is not under the root gets no
// anchor and keeps the final-component guard it always had.
anchor := ""
if pathWithinRoot(capability.Root, path) {
anchor = capability.Root
}
entries = append(entries, WindowsACLEntry{
Action: WindowsACLDenyWrite,
Path: path,
Capability: capability.SID,
Anchor: anchor,
})
}
}
+26 -1
View File
@@ -23,7 +23,10 @@ const (
)
type windowsACLPathGroup struct {
Path string
Path string
// Anchor is the write root Path was derived from, empty when the operator
// named the path. See verifyWindowsACLHandleUnderAnchor.
Anchor string
Entries []WindowsACLEntry
Materialize bool
}
@@ -69,6 +72,12 @@ func groupWindowsACLPlanByPath(plan WindowsACLPlan) []windowsACLPathGroup {
}
group.Entries = append(group.Entries, entry)
group.Materialize = group.Materialize || entry.Materialize
// One path is derived from at most one root, so the first anchor seen is
// the anchor. Taking it rather than overwriting keeps a later
// operator-named duplicate of the same path from clearing it.
if group.Anchor == "" {
group.Anchor = entry.Anchor
}
}
out := make([]windowsACLPathGroup, 0, len(byPath))
for _, group := range byPath {
@@ -104,6 +113,13 @@ func applyWindowsACLPathGroup(group windowsACLPathGroup) (windowsACLSnapshot, bo
}
return windowsACLSnapshot{}, false, nil
}
// Before creating anything: os.MkdirAll walks a pathname and follows every
// reparse point on it, so a junction on the derived tail would have this
// elevated setup create the directory outside the write root and only the
// containment check below would notice, after the fact.
if err := verifyWindowsACLPathUnderAnchor(group.Anchor, path); err != nil {
return windowsACLSnapshot{}, false, err
}
if err := os.MkdirAll(path, 0o700); err != nil {
return windowsACLSnapshot{}, false, fmt.Errorf("materialize windows ACL target %s: %w", path, err)
}
@@ -123,6 +139,15 @@ func applyWindowsACLPathGroup(group windowsACLPathGroup) (windowsACLSnapshot, bo
}
return windowsACLSnapshot{}, false, err
}
// THE OBJECT HAS TO BE WHERE THE PLAN SAID IT WOULD BE. The open above
// refuses a reparse point at the final component and resolves every one
// above it, which is right for a path the operator named and not enough for
// one this package derived from a write root: a junction on the derived tail
// redirects the handle out of the sandbox with nothing about the final
// object looking wrong.
if err := verifyWindowsACLHandleUnderAnchor(handle, group.Anchor, path); err != nil {
return fail(err)
}
descriptor, err := windows.GetSecurityInfo(handle, windows.SE_FILE_OBJECT, windows.DACL_SECURITY_INFORMATION)
if err != nil {
return fail(fmt.Errorf("read windows ACL for %s: %w", path, err))
@@ -0,0 +1,228 @@
//go:build windows
package sandbox
import (
"errors"
"fmt"
"os"
"path/filepath"
"strings"
"golang.org/x/sys/windows"
)
// Keeping an elevated ACL write inside the tree it was planned for.
//
// FILE_FLAG_OPEN_REPARSE_POINT GUARDS ONE COMPONENT, AND A PATH HAS MANY. The
// apply opens its target with that flag, which stops the FINAL component from
// being followed and refuses the handle when it is a reparse point. Every
// component above it is resolved normally, as it must be for any absolute path
// to work at all.
//
// That is enough for a target the operator named, and not enough for one the
// sandbox derived. The write-root carveouts are derived: <root>/.git/hooks and
// <root>/.git/config are constructed from a root, and <root>/.git is a name an
// unprivileged workspace writer can create. `mklink /J` needs no privilege, so
// a junction there makes the apply open <junction-target>/hooks instead, and
// `zero sandbox setup` writes a deny-write ACE, as Administrator, on an object
// outside the workspace. The final-component flag never fires: the object at
// the end of the walk is an ordinary directory.
//
// So the derived targets carry the root they were derived FROM, and the object
// finally opened has to still be inside it. The check is on the handle, not on
// the name: GetFinalPathNameByHandle answers where the object this process is
// holding actually lives, and both sides of the comparison go through it so the
// spelling is normalized the same way (\\?\ prefix, long names, drive letter).
// A junction anywhere along the derived tail moves the answer out of the root
// and the apply refuses.
//
// The rule is deliberately not "no reparse point anywhere in the path". Above
// the write root the path belongs to the operator, who may legitimately keep a
// workspace under a junction or a mapped directory, and refusing that would
// break setups this has nothing to say about. Only the tail the sandbox itself
// appended is held strict.
// windowsFinalPathByHandle reports where an open handle's object actually lives.
func windowsFinalPathByHandle(handle windows.Handle) (string, error) {
buffer := make([]uint16, windows.MAX_LONG_PATH)
n, err := windows.GetFinalPathNameByHandle(handle, &buffer[0], uint32(len(buffer)), 0)
if err != nil {
return "", err
}
if int(n) > len(buffer) {
buffer = make([]uint16, n)
n, err = windows.GetFinalPathNameByHandle(handle, &buffer[0], uint32(len(buffer)), 0)
if err != nil {
return "", err
}
}
return windows.UTF16ToString(buffer[:n]), nil
}
// windowsFinalPathOf opens path without following a final-component reparse
// point and reports where the object lives. A missing path is surfaced as
// os.ErrNotExist so callers can walk up to a parent that exists.
func windowsFinalPathOf(path string) (string, error) {
utf16Path, err := windows.UTF16PtrFromString(path)
if err != nil {
return "", fmt.Errorf("encode windows path %s: %w", path, err)
}
handle, err := windows.CreateFile(
utf16Path,
windows.READ_CONTROL,
windows.FILE_SHARE_READ|windows.FILE_SHARE_WRITE|windows.FILE_SHARE_DELETE,
nil,
windows.OPEN_EXISTING,
windows.FILE_FLAG_BACKUP_SEMANTICS|windows.FILE_FLAG_OPEN_REPARSE_POINT,
0,
)
if err != nil {
return "", fmt.Errorf("open windows path %s: %w", path, err)
}
defer func() { _ = windows.CloseHandle(handle) }()
final, err := windowsFinalPathByHandle(handle)
if err != nil {
return "", fmt.Errorf("resolve windows path %s: %w", path, err)
}
return final, nil
}
// windowsACLTailUnderAnchor returns the components target adds to anchor, and
// whether target is under anchor lexically at all.
func windowsACLTailUnderAnchor(anchor, target string) (string, bool) {
relative, err := filepath.Rel(filepath.Clean(anchor), filepath.Clean(target))
if err != nil {
return "", false
}
if relative == "." {
return "", true
}
if relative == ".." || strings.HasPrefix(relative, ".."+string(filepath.Separator)) {
return "", false
}
return relative, true
}
// windowsACLContainmentError reports a derived target that resolved outside the
// root it was derived from.
type windowsACLContainmentError struct {
Anchor string
Target string
Actual string
}
func (err windowsACLContainmentError) Error() string {
return fmt.Sprintf(
"refusing to apply ACL to %s: it resolves to %s, outside the write root %s; "+
"a reparse point on the way down redirects an elevated ACL write out of the sandbox",
err.Target, err.Actual, err.Anchor)
}
// verifyWindowsACLHandleUnderAnchor refuses a handle whose object does not live
// where the plan said it would.
//
// anchor empty means the target was named rather than derived, and there is no
// owned tail to hold strict; the final-component check the caller already did is
// the whole guard for those.
func verifyWindowsACLHandleUnderAnchor(handle windows.Handle, anchor, target string) error {
if strings.TrimSpace(anchor) == "" {
return nil
}
tail, ok := windowsACLTailUnderAnchor(anchor, target)
if !ok {
return windowsACLContainmentError{Anchor: anchor, Target: target, Actual: target}
}
anchorFinal, err := windowsFinalPathOf(anchor)
if err != nil {
return fmt.Errorf("resolve write root %s: %w", anchor, err)
}
targetFinal, err := windowsFinalPathByHandle(handle)
if err != nil {
return fmt.Errorf("resolve windows ACL target %s: %w", target, err)
}
expected := anchorFinal
if tail != "" {
expected = filepath.Join(anchorFinal, tail)
}
if !strings.EqualFold(filepath.Clean(targetFinal), filepath.Clean(expected)) {
return windowsACLContainmentError{Anchor: anchor, Target: target, Actual: targetFinal}
}
return nil
}
// verifyWindowsACLPathUnderAnchor makes the same check against the deepest
// component of target that exists, before anything is created.
//
// Materialization creates a missing target with os.MkdirAll, which walks a
// pathname and follows every reparse point on it, so without this an elevated
// setup creates the directory on the far side of a junction and only the
// containment check afterwards notices. The window between this check and that
// create is not closed here; closing it needs the components created relative to
// retained handles, which is the rooted descent tracked in #808.
func verifyWindowsACLPathUnderAnchor(anchor, target string) error {
if strings.TrimSpace(anchor) == "" {
return nil
}
if _, ok := windowsACLTailUnderAnchor(anchor, target); !ok {
return windowsACLContainmentError{Anchor: anchor, Target: target, Actual: target}
}
existing := filepath.Clean(target)
for {
if _, err := os.Lstat(existing); err == nil {
break
} else if !errors.Is(err, os.ErrNotExist) {
return fmt.Errorf("inspect windows ACL target %s: %w", existing, err)
}
parent := filepath.Dir(existing)
if parent == existing {
return nil
}
existing = parent
}
tail, ok := windowsACLTailUnderAnchor(anchor, existing)
if !ok || tail == "" {
// Above or at the write root, which is the operator's to arrange.
return nil
}
handle, err := windowsACLOpenForContainment(existing)
if err != nil {
return err
}
defer func() { _ = windows.CloseHandle(handle) }()
// THE ANCESTOR ITSELF IS THE REDIRECTION, WHICH IS WHY ITS NAME LOOKS RIGHT.
//
// Asking only where this object lives answers "exactly where you asked": the
// open above does not follow a final-component reparse point, so a junction
// at <root>/.git comes back as <root>/.git. os.MkdirAll does follow it, and
// creates the missing components on the far side. So what disqualifies this
// ancestor is that it IS a reparse point, not where it reports living.
var info windows.ByHandleFileInformation
if err := windows.GetFileInformationByHandle(handle, &info); err != nil {
return fmt.Errorf("inspect windows ACL target %s: %w", existing, err)
}
if info.FileAttributes&windows.FILE_ATTRIBUTE_REPARSE_POINT != 0 {
return windowsACLContainmentError{Anchor: anchor, Target: target, Actual: existing + " (a reparse point)"}
}
return verifyWindowsACLHandleUnderAnchor(handle, anchor, existing)
}
func windowsACLOpenForContainment(path string) (windows.Handle, error) {
utf16Path, err := windows.UTF16PtrFromString(path)
if err != nil {
return 0, fmt.Errorf("encode windows ACL target %s: %w", path, err)
}
handle, err := windows.CreateFile(
utf16Path,
windows.READ_CONTROL,
windows.FILE_SHARE_READ|windows.FILE_SHARE_WRITE|windows.FILE_SHARE_DELETE,
nil,
windows.OPEN_EXISTING,
windows.FILE_FLAG_BACKUP_SEMANTICS|windows.FILE_FLAG_OPEN_REPARSE_POINT,
0,
)
if err != nil {
return 0, fmt.Errorf("open windows ACL target %s: %w", path, err)
}
return handle, nil
}
@@ -0,0 +1,231 @@
//go:build windows
package sandbox
import (
"errors"
"os"
"os/exec"
"path/filepath"
"strings"
"testing"
)
// windowsACLTestDenySID is a well-known group this process is not a member of.
//
// A deny ACE is what the carveouts actually are, and denying a group the test
// runs as revokes its own WRITE_DAC: the apply succeeds and then nothing can
// put the DACL back, so the temporary tree cannot be removed. Guests keeps the
// deny semantics without locking the test out of its own fixture.
const windowsACLTestDenySID = "S-1-5-32-546"
// makeACLJunction plants a directory junction at link pointing at target.
//
// A junction rather than a symlink on purpose: mklink /J needs no privilege, so
// it is reachable by exactly the unprivileged workspace writer this guards
// against, and the test runs on an ordinary developer or CI account.
func makeACLJunction(t *testing.T, link, target string) {
t.Helper()
output, err := exec.Command("cmd.exe", "/c", "mklink", "/J", link, target).CombinedOutput()
if err != nil {
t.Skipf("mklink /J unavailable here: %v (%s)", err, output)
}
}
// AN ELEVATED ACL WRITE MUST NOT LEAVE THE WRITE ROOT THROUGH A JUNCTION.
//
// The apply opens its target with FILE_FLAG_OPEN_REPARSE_POINT, which refuses a
// reparse point at the FINAL component and resolves every component above it.
// The write-root carveouts are derived rather than named: <root>/.git/hooks and
// <root>/.git/config are constructed from the root, and <root>/.git is a name an
// unprivileged workspace writer can create before setup runs. A junction there
// makes the apply open <junction-target>/hooks, an ordinary directory with
// nothing wrong about its final component, and `zero sandbox setup` writes a
// deny-write ACE on it as Administrator, outside the workspace.
func TestApplyWindowsACLRefusesACarveoutRedirectedByAJunction(t *testing.T) {
root := t.TempDir()
outside := filepath.Join(t.TempDir(), "elsewhere")
if err := os.MkdirAll(filepath.Join(outside, "hooks"), 0o700); err != nil {
t.Fatal(err)
}
makeACLJunction(t, filepath.Join(root, ".git"), outside)
carveout := filepath.Join(root, ".git", "hooks")
group := windowsACLPathGroup{
Path: carveout,
Anchor: root,
Entries: []WindowsACLEntry{{
Action: WindowsACLDenyWrite,
Path: carveout,
Capability: windowsACLTestDenySID,
Anchor: root,
}},
}
_, applied, err := applyWindowsACLPathGroup(group)
if applied {
t.Error("applied a deny-write ACE through a junction, so an elevated setup rewrote the DACL of an object outside the workspace")
}
var containment windowsACLContainmentError
if !errors.As(err, &containment) {
t.Fatalf("err = %v, want a containment refusal", err)
}
if !strings.Contains(containment.Actual, "elsewhere") {
t.Errorf("the refusal does not name where the target actually resolved: %v", containment)
}
}
// The same carveout inside an ordinary workspace still gets its ACE, or the
// refusal above would prove only that the apply fails.
func TestApplyWindowsACLStillAppliesAnOrdinaryCarveout(t *testing.T) {
root := t.TempDir()
carveout := filepath.Join(root, ".git", "hooks")
if err := os.MkdirAll(carveout, 0o700); err != nil {
t.Fatal(err)
}
group := windowsACLPathGroup{
Path: carveout,
Anchor: root,
Entries: []WindowsACLEntry{{
Action: WindowsACLDenyWrite,
Path: carveout,
Capability: windowsACLTestDenySID,
Anchor: root,
}},
}
snapshot, applied, err := applyWindowsACLPathGroup(group)
if err != nil {
t.Fatalf("apply: %v", err)
}
// PUT IT BACK, OR THE TEMP TREE CANNOT BE REMOVED. The ACE just written
// denies write to the group this test runs as, and t.TempDir removes the
// tree when the test ends.
if applied {
t.Cleanup(func() {
if err := rollbackWindowsACLSnapshots([]windowsACLSnapshot{snapshot}); err != nil {
t.Errorf("rollback: %v", err)
}
})
}
if !applied {
t.Fatal("an ordinary carveout inside the write root was not applied")
}
if snapshot.Path != carveout {
t.Fatalf("snapshot path = %q, want %q", snapshot.Path, carveout)
}
}
// MATERIALIZATION MUST NOT CREATE THROUGH THE JUNCTION EITHER. os.MkdirAll
// walks a pathname and follows every reparse point on it, so without the
// pre-create check an elevated setup makes the directory on the far side and
// only notices afterwards.
func TestApplyWindowsACLRefusesToMaterializeThroughAJunction(t *testing.T) {
root := t.TempDir()
outside := filepath.Join(t.TempDir(), "elsewhere")
if err := os.MkdirAll(outside, 0o700); err != nil {
t.Fatal(err)
}
makeACLJunction(t, filepath.Join(root, ".git"), outside)
carveout := filepath.Join(root, ".git", "hooks")
if _, err := os.Lstat(filepath.Join(outside, "hooks")); !errors.Is(err, os.ErrNotExist) {
t.Fatalf("SETUP INVALID: the target already holds hooks: %v", err)
}
group := windowsACLPathGroup{
Path: carveout,
Anchor: root,
Materialize: true,
Entries: []WindowsACLEntry{{
Action: WindowsACLDenyRead,
Path: carveout,
Capability: windowsACLTestDenySID,
Anchor: root,
Materialize: true,
}},
}
if _, applied, err := applyWindowsACLPathGroup(group); applied || err == nil {
t.Errorf("materialized through a junction: applied=%v err=%v", applied, err)
}
if _, err := os.Lstat(filepath.Join(outside, "hooks")); !errors.Is(err, os.ErrNotExist) {
t.Errorf("an elevated create landed outside the write root: %v", err)
}
}
// An operator-named path carries no anchor and keeps the behaviour it had: the
// final component is still refused when it is a reparse point, and nothing above
// it is second-guessed, because above the write root the path belongs to the
// operator, who may keep a workspace under a junction.
func TestApplyWindowsACLLeavesAnUnanchoredPathAlone(t *testing.T) {
base := t.TempDir()
outside := filepath.Join(t.TempDir(), "elsewhere")
if err := os.MkdirAll(filepath.Join(outside, "leaf"), 0o700); err != nil {
t.Fatal(err)
}
makeACLJunction(t, filepath.Join(base, "link"), outside)
named := filepath.Join(base, "link", "leaf")
group := windowsACLPathGroup{
Path: named,
Entries: []WindowsACLEntry{{
Action: WindowsACLDenyWrite,
Path: named,
Capability: windowsACLTestDenySID,
}},
}
snapshot, applied, err := applyWindowsACLPathGroup(group)
if err != nil {
t.Fatalf("an operator-named path below a junction was refused: %v", err)
}
if applied {
t.Cleanup(func() {
if err := rollbackWindowsACLSnapshots([]windowsACLSnapshot{snapshot}); err != nil {
t.Errorf("rollback: %v", err)
}
})
}
if !applied {
t.Fatal("an operator-named path below a junction was not applied")
}
}
// AND THE GUARD ON THE CREATE HAS TO BE PINNED WHERE IT CAN BE SEEN.
//
// Through the whole apply, dropping the pre-create check is invisible: the
// directory is made on the far side of the junction, the containment check on
// the handle then refuses, and the failure path removes what it made, so the
// filesystem afterwards looks the same either way. The transient elevated create
// outside the write root is the thing being prevented, so it is checked here,
// against the function that prevents it.
func TestVerifyWindowsACLPathUnderAnchorRefusesAMissingTargetBehindAJunction(t *testing.T) {
root := t.TempDir()
outside := filepath.Join(t.TempDir(), "elsewhere")
if err := os.MkdirAll(outside, 0o700); err != nil {
t.Fatal(err)
}
makeACLJunction(t, filepath.Join(root, ".git"), outside)
carveout := filepath.Join(root, ".git", "hooks")
if _, err := os.Lstat(carveout); !errors.Is(err, os.ErrNotExist) {
t.Fatalf("SETUP INVALID: the carveout already exists, so nothing would be created: %v", err)
}
err := verifyWindowsACLPathUnderAnchor(root, carveout)
var containment windowsACLContainmentError
if !errors.As(err, &containment) {
t.Fatalf("err = %v, want a containment refusal before anything is created", err)
}
// The ordinary case still says yes, or the refusal proves only that the
// function rejects things.
ordinary := t.TempDir()
if err := verifyWindowsACLPathUnderAnchor(ordinary, filepath.Join(ordinary, ".git", "hooks")); err != nil {
t.Errorf("an ordinary missing carveout was refused: %v", err)
}
// And an unanchored path is not this function's business.
if err := verifyWindowsACLPathUnderAnchor("", carveout); err != nil {
t.Errorf("an unanchored path was refused: %v", err)
}
}
+114
View File
@@ -240,3 +240,117 @@ func TestDedupeWindowsACLEntriesKeepsInheritanceVariants(t *testing.T) {
t.Fatalf("dedupe order/shape = %#v, want first NoInherit then inheritable", out)
}
}
// THE PLAN HAS TO CARRY THE ANCHOR, OR THE APPLY HAS NOTHING TO ENFORCE.
//
// The apply-side tests hand applyWindowsACLPathGroup an anchor directly, so they
// pass whether or not anything ever sets one. This drives the real builder and
// checks which entries come out anchored: the carveouts derived from a write
// root, and not the paths the operator named, which have no owned tail and whose
// intermediates are the operator's own business.
//
// NATIVE PATHS, NOT WINDOWS LITERALS. The builder is cross-platform and so is
// this test, and the anchoring decision runs through pathWithinRoot, which is
// filepath.Rel underneath. On Linux a backslash is an ordinary character, so
// `C:\workspace\.git\hooks` is one component and is not under `C:\workspace`;
// the neighbouring tests get away with such literals because they only compare
// strings they built the same way and never ask whether one contains another.
func TestBuildWindowsACLPlanAnchorsDerivedCarveouts(t *testing.T) {
home := t.TempDir()
workspace := t.TempDir()
carveout := filepath.Join(workspace, ".git", "hooks")
metadata := filepath.Join(workspace, ".zero")
named := filepath.Join(t.TempDir(), "named-by-the-operator")
config := WindowsSandboxCommandConfig{
SandboxHome: home,
WorkspaceRoots: []string{workspace},
PermissionProfile: PermissionProfile{
FileSystem: FileSystemPolicy{
Kind: FileSystemRestricted,
WriteRoots: []WritableRoot{{
Root: workspace,
ReadOnlySubpaths: []string{carveout},
ProtectedMetadataNames: []string{".zero"},
}},
DenyWrite: []string{named},
},
Network: NetworkPolicy{Mode: NetworkDeny},
},
}
plan, err := BuildWindowsACLPlan(config)
if err != nil {
t.Fatalf("BuildWindowsACLPlan: %v", err)
}
anchors := map[string]string{}
for _, entry := range plan.Entries {
anchors[strings.ToLower(entry.Path)] = entry.Anchor
}
for _, derived := range []string{carveout, metadata} {
anchor, present := anchors[strings.ToLower(derived)]
if !present {
t.Fatalf("SETUP INVALID: the plan has no entry for the derived carveout %s", derived)
}
if anchor != workspace {
t.Errorf("derived carveout %s carries anchor %q, want the write root it came from", derived, anchor)
}
}
operatorAnchor, present := anchors[strings.ToLower(named)]
if !present {
t.Fatal("SETUP INVALID: the plan has no entry for the operator-named deny path")
}
if operatorAnchor != "" {
t.Errorf("operator-named path carries anchor %q, want none: its intermediates are not the sandbox's to police", operatorAnchor)
}
}
// AND AN OUT-OF-ROOT READ-ONLY SUBPATH IS NOT ANCHORED AT ALL.
//
// ReadOnlySubpaths is a profile field an operator can set to any path. One
// placed outside the write root is a configuration that works today, and
// anchoring it to that root would turn it into a containment refusal: the
// object is not under the root and never was. Only the paths that are actually
// under it are held to it.
func TestBuildWindowsACLPlanLeavesAnOutOfRootSubpathUnanchored(t *testing.T) {
home := t.TempDir()
workspace := t.TempDir()
inside := filepath.Join(workspace, "vendor")
outside := filepath.Join(t.TempDir(), "shared")
plan, err := BuildWindowsACLPlan(WindowsSandboxCommandConfig{
SandboxHome: home,
WorkspaceRoots: []string{workspace},
PermissionProfile: PermissionProfile{
FileSystem: FileSystemPolicy{
Kind: FileSystemRestricted,
WriteRoots: []WritableRoot{{
Root: workspace,
ReadOnlySubpaths: []string{inside, outside},
}},
},
Network: NetworkPolicy{Mode: NetworkDeny},
},
})
if err != nil {
t.Fatalf("BuildWindowsACLPlan: %v", err)
}
seen := map[string]string{}
for _, entry := range plan.Entries {
seen[strings.ToLower(entry.Path)] = entry.Anchor
}
insideAnchor, present := seen[strings.ToLower(inside)]
if !present {
t.Fatal("SETUP INVALID: the plan has no entry for the in-root subpath")
}
if insideAnchor != workspace {
t.Errorf("in-root subpath carries anchor %q, want the write root", insideAnchor)
}
outsideAnchor, present := seen[strings.ToLower(outside)]
if !present {
t.Fatal("SETUP INVALID: the plan has no entry for the out-of-root subpath")
}
if outsideAnchor != "" {
t.Errorf("out-of-root subpath carries anchor %q, want none: the apply would refuse a config that works today", outsideAnchor)
}
}
+3 -12
View File
@@ -226,20 +226,11 @@ func CompactionPayloadFromPlan(summary string, plan CompactionPlan) (CompactionP
}
func (store *Store) ReadRehydratedEvents(sessionID string) ([]Event, error) {
events, _, err := store.ReadRehydratedEventsWithPresence(sessionID)
return events, err
}
// ReadRehydratedEventsWithPresence carries the underlying event-log presence
// through compaction projection without changing ReadRehydratedEvents' existing
// empty-on-missing contract.
func (store *Store) ReadRehydratedEventsWithPresence(sessionID string) ([]Event, bool, error) {
events, present, err := store.ReadEventsWithPresence(sessionID)
events, err := store.ReadEvents(sessionID)
if err != nil {
return nil, present, err
return nil, err
}
rehydrated, err := RehydrateEvents(events)
return rehydrated, present, err
return RehydrateEvents(events)
}
func (store *Store) ReadReplayEvents(sessionID string) ([]Event, error) {
+5 -14
View File
@@ -813,24 +813,15 @@ func (store *Store) UpdateModel(sessionID string, modelID string) (Metadata, err
}
func (store *Store) ReadEvents(sessionID string) ([]Event, error) {
events, _, err := store.ReadEventsWithPresence(sessionID)
return events, err
}
// ReadEventsWithPresence preserves the distinction between an intentionally
// empty event log and a missing one. Most readers accept both as empty history,
// but activation protocols that promise to restore previously populated
// context need the stronger signal.
func (store *Store) ReadEventsWithPresence(sessionID string) ([]Event, bool, error) {
if !ValidSessionID(sessionID) {
return nil, false, fmt.Errorf("invalid zero session id %q", sessionID)
return nil, fmt.Errorf("invalid zero session id %q", sessionID)
}
data, err := os.ReadFile(store.eventsPath(sessionID))
if err != nil {
if errors.Is(err, os.ErrNotExist) {
return []Event{}, false, nil
return []Event{}, nil
}
return nil, false, fmt.Errorf("read zero session events: %w", err)
return nil, fmt.Errorf("read zero session events: %w", err)
}
// A genuine torn tail is an INCOMPLETE final write — a crash mid-append leaves
// the last line without its terminating newline. If the file ends with a
@@ -860,11 +851,11 @@ func (store *Store) ReadEventsWithPresence(sessionID string) ([]Event, bool, err
if index == lastNonEmpty && tornTailPossible {
break
}
return nil, true, fmt.Errorf("invalid json in zero session %s %s at line %d: %w", sessionID, EventsFile, index+1, err)
return nil, fmt.Errorf("invalid json in zero session %s %s at line %d: %w", sessionID, EventsFile, index+1, err)
}
events = append(events, event)
}
return events, true, nil
return events, nil
}
func (store *Store) timestamp() string {