fix: idle Factory polls read only job documents that changed

cmd/codeaf/factory_stage.go

@@ -12,6 +12,7 @@ import (
 	"time"
 
 	"github.com/Agent-Field/codeaf/internal/approval"
+	"github.com/Agent-Field/codeaf/internal/config"
 	"github.com/Agent-Field/codeaf/internal/env"
 	"github.com/Agent-Field/codeaf/internal/factory"
 	factoryrun "github.com/Agent-Field/codeaf/internal/factory/run"
@@ -295,13 +296,21 @@ func factoryStagePosture() string {
 // name) still hold, because they sit under every policy and they are not a
 // person's preference. A person who wants a stage narrower says so with
 // [factoryPostureEnv].
+// An unreadable profile still requires a person: its saved denials are unknown.
 type stageApprovalGate struct {
 	rows    *v3ApprovalGate
 	posture string
 }
 
 // Build is the gate for one posture.
 func (g stageApprovalGate) Build(posture string) (*approval.Policy, bool, error) {
+	if g.rows != nil {
+		_, err := config.ReadProfileApprovalConfig(g.rows.profileDir)
+		var pathErr *os.PathError
+		if errors.As(err, &pathErr) {
+			return &approval.Policy{ProfileReadError: err.Error()}, false, nil
+		}
+	}
 	if posture != session.PostureAllow {
 		return g.rows.Build(posture)
 	}
@@ -486,6 +495,10 @@ func (c *stageConversation) drain(events <-chan session.Event) {
 			}
 			c.say(factory.ToolCaptionPrefix + factory.ToolLogText(hint, c.dir))
 		case session.EventToolFailed:
+			var refused *approval.ProfileReadFailure
+			if errors.As(event.Err, &refused) {
+				c.fail(refused)
+			}
 			reason := event.Hint
 			if reason == "" && event.Err != nil {
 				reason = event.Err.Error()
@@ -519,7 +532,10 @@ func (c *stageConversation) fail(err error) {
 	}
 	c.mu.Lock()
 	defer c.mu.Unlock()
-	c.failure = err
+	var refused *approval.ProfileReadFailure
+	if !errors.As(c.failure, &refused) {
+		c.failure = err
+	}
 }
 
 // say puts one caption on the account, dropping it rather than holding the
@@ -631,6 +647,10 @@ func (c *stageConversation) Wait(ctx context.Context) error {
 	case <-done:
 		c.mu.Lock()
 		defer c.mu.Unlock()
+		var refused *approval.ProfileReadFailure
+		if errors.As(c.failure, &refused) {
+			return refused
+		}
 		if c.budget != nil {
 			if why := c.budget.Refusal(); why != "" {
 				// A daily refusal already recorded was not caused by memory, so

cmd/codeaf/factory_stage_profile_job_test.go

@@ -0,0 +1,148 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"github.com/Agent-Field/codeaf/internal/config"
+	"github.com/Agent-Field/codeaf/internal/factory"
+	factoryrun "github.com/Agent-Field/codeaf/internal/factory/run"
+	"github.com/Agent-Field/codeaf/internal/factory/store"
+	"github.com/Agent-Field/codeaf/internal/home"
+	"github.com/Agent-Field/codeaf/internal/remote"
+	"io"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+func TestUnreadableStageJobKeepsConfigReason(t *testing.T) {
+	testUnreadableStageJobKeepsConfigReason(t, false)
+}
+
+func TestInitiallyUnreadableStageJobKeepsConfigReason(t *testing.T) {
+	testUnreadableStageJobKeepsConfigReason(t, true)
+}
+
+func testUnreadableStageJobKeepsConfigReason(t *testing.T, initiallyUnreadable bool) {
+	t.Helper()
+	if os.Geteuid() == 0 {
+		t.Skip("mode 000 requires nonroot")
+	}
+	root := resolvedTempDir(t)
+	t.Setenv("HOME", filepath.Join(root, "home"))
+	t.Setenv(home.EnvVar, filepath.Join(root, "state"))
+	t.Setenv(config.ProfileDirEnv, "")
+	t.Setenv("OPENROUTER_API_KEY", "test-key")
+	t.Setenv(factoryPostureEnv, "")
+	workspace := filepath.Join(root, "work")
+	if err := os.MkdirAll(workspace, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	target := filepath.Join(workspace, "NOTES.md")
+	writeArgs, _ := json.Marshal(map[string]string{"path": target, "content": "written by the stage\n"})
+	model := newScriptedStageModel(t, stageCall{"write", string(writeArgs)}, stageCall{"stage_result", stageModelReport})
+	t.Setenv("CODEAF_BASE_URL", model.server.URL)
+	t.Chdir(workspace)
+	freshEngineProcess(t)
+	stageSettle = 20 * time.Millisecond
+	t.Cleanup(func() { stageSettle = unattendedRunSettleTick })
+
+	proc, err := openEngineProcess()
+	if err != nil {
+		t.Fatal(err)
+	}
+	launch, err := openV3Launch(proc, engineLaunchOptions(remote.Hello{Version: remote.Version, Workspace: workspace}, workspace, ""))
+	if err != nil {
+		t.Fatal(err)
+	}
+	st, err := store.Open(filepath.Join(root, "factory"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	id, err := st.Add(context.Background(), factory.Item{
+		Title: "write the notes", Repo: "work", Tier: factory.TierOwner,
+		Stages: []factory.Stage{{Name: "write", Kind: factory.StageChat, Ask: "write the notes", Until: "done", On: true}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := st.Update(id, func(it *factory.Item) error { it.Stream = &factory.Stream{}; return nil }); err != nil {
+		t.Fatal(err)
+	}
+	it, _ := st.Get(id)
+
+	profile := ""
+	file := config.BudgetConfigPath(profile)
+	if err := os.MkdirAll(filepath.Dir(file), 0700); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(file, []byte(`{"tools.approvals":"write:deny"}`), 0600); err != nil {
+		t.Fatal(err)
+	}
+	reason := "saved approval rules could not be read"
+	if initiallyUnreadable {
+		if err := os.Chmod(file, 0); err != nil {
+			t.Fatal(err)
+		}
+		reason = "permission denied"
+	} else {
+		oldHandler := model.server.Config.Handler
+		var once sync.Once
+		model.server.Config.Handler = http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			body, _ := io.ReadAll(r.Body)
+			r.Body = io.NopCloser(bytes.NewReader(body))
+			if bytes.Contains(body, []byte(`"stage_result"`)) {
+				once.Do(func() {
+					if err := os.Chmod(file, 0); err != nil {
+						t.Error(err)
+					}
+				})
+			}
+			oldHandler.ServeHTTP(w, r)
+		})
+	}
+	t.Cleanup(func() { _ = os.Chmod(file, 0600) })
+	maker := factoryStageMaker(st, workspace, profile, launch.Config)
+	exec := factoryChatStage(st, profile, factoryrun.NewChatExecutor(maker))
+	ctx, cancel := context.WithTimeout(context.Background(), 12*time.Second)
+	defer cancel()
+	res, err := exec.Run(ctx, factoryrun.Job{Item: it, Stage: it.Stages[0], Index: 0, Round: 1, Dir: workspace, Log: func(line string) {
+		t.Logf("JOB LOG: %s", line)
+		if strings.Contains(line, "reminded") {
+			t.Error("terminal profile refusal must not retry reporting")
+		}
+	}})
+	if err != nil {
+		t.Fatalf("ROUND ERROR: %v", err)
+	}
+	t.Logf("ROUND RESULT: %+v", res)
+	if res.Done || res.Exit != 1 || !strings.Contains(res.Output, file) || !strings.Contains(res.Output, reason) {
+		t.Errorf("job lost the terminal profile-read reason: %+v", res)
+	}
+	model.mu.Lock()
+	defer model.mu.Unlock()
+	if initiallyUnreadable {
+		t.Logf("STAGE MODEL REQUESTS: %d", len(model.briefs))
+		if len(model.briefs) != 0 || res.Spent != 0 || res.Reports != 0 {
+			t.Fatalf("refused admission did work: requests=%d spent=%v reports=%d", len(model.briefs), res.Spent, res.Reports)
+		}
+	}
+	for _, brief := range model.briefs {
+		if i := strings.Index(brief, "no resolver is attached"); i >= 0 {
+			t.Logf("MODEL RECEIVED REFUSAL: %.650s", brief[i:])
+		}
+	}
+	got, readErr := os.ReadFile(target)
+	t.Logf("TARGET: content=%q err=%v", got, readErr)
+	if !os.IsNotExist(readErr) {
+		t.Fatalf("unreadable stage wrote file")
+	}
+	if ctx.Err() != nil {
+		t.Fatalf("stage failed to stop before timeout: %v", ctx.Err())
+	}
+}

cmd/codeaf/factory_stage_unreadable_test.go

@@ -0,0 +1,95 @@
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/Agent-Field/codeaf/internal/approval"
+	"github.com/Agent-Field/codeaf/internal/config"
+	"github.com/Agent-Field/codeaf/internal/session"
+)
+
+func TestStageUnreadableProfileRequiresPerson(t *testing.T) {
+	if os.Geteuid() == 0 {
+		t.Skip("mode 000 requires nonroot")
+	}
+	profile := v3Profile(t, map[string]any{config.KeyToolApprovalMode: "prompt", config.KeyToolApprovals: "write:deny"})
+	file := filepath.Join(profile, "config.json")
+	if err := os.Chmod(file, 0); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = os.Chmod(file, 0o600) })
+	rows := &v3ApprovalGate{workspace: t.TempDir(), profileDir: profile, headless: true}
+	for _, posture := range []string{session.PostureAllow, session.PostureAsk, session.PostureDeny} {
+		t.Run(posture, func(t *testing.T) {
+			var gate session.ApprovalGate = stageApprovalGate{rows: rows, posture: posture}
+			policy, guardian, err := gate.Build(posture)
+			if err != nil {
+				t.Fatal(err)
+			}
+			decision := policy.Check("write", json.RawMessage(`{"path":"x","content":"x"}`))
+			t.Logf("stage write: %s", decision)
+			if decision.Action != approval.ActionPrompt || guardian || !strings.Contains(decision.Rule, file) || !strings.Contains(decision.Rule, "could not be read") {
+				t.Fatalf("unreadable saved denial must require a person with a clear reason: decision=%+v guardian=%v", decision, guardian)
+			}
+		})
+	}
+}
+
+func TestStageProfileReadFailureWinsOverProjectParseError(t *testing.T) {
+	profile, workspace := t.TempDir(), t.TempDir()
+	file := filepath.Join(profile, "config.json")
+	if err := os.Mkdir(file, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	_, readErr := config.ReadProfileApprovalConfig(profile)
+	if readErr == nil {
+		t.Fatal("expected a filesystem read failure")
+	}
+	if err := os.Mkdir(filepath.Join(workspace, ".codeaf"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(workspace, ".codeaf", "config.json"), []byte("{"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	for _, posture := range []string{session.PostureAllow, session.PostureAsk, session.PostureDeny, session.PostureGuardian, session.PostureAuto} {
+		t.Run(posture, func(t *testing.T) {
+			var gate session.ApprovalGate = stageApprovalGate{rows: &v3ApprovalGate{workspace: workspace, profileDir: profile}, posture: posture}
+			policy, guardian, err := gate.Build(posture)
+			if err != nil {
+				t.Fatalf("known profile failure lost to project error: %v", err)
+			}
+			decision := policy.Check("write", json.RawMessage(`{"path":"notes.txt","content":"no"}`))
+			if guardian || decision.Action != approval.ActionPrompt || !strings.Contains(decision.ProfileReadError, file) {
+				t.Fatalf("lost profile failure: guardian=%v decision=%+v", guardian, decision)
+			}
+		})
+	}
+}
+
+func TestStageReadableMalformedProfileKeepsDefaultAllowance(t *testing.T) {
+	for _, contents := range []string{"{", "null", "[]"} {
+		t.Run(contents, func(t *testing.T) {
+			profile := t.TempDir()
+			if err := os.WriteFile(filepath.Join(profile, "config.json"), []byte(contents), 0o600); err != nil {
+				t.Fatal(err)
+			}
+			rows := &v3ApprovalGate{workspace: t.TempDir(), profileDir: profile}
+			var gate session.ApprovalGate = stageApprovalGate{rows: rows, posture: session.PostureAllow}
+			policy, guardian, err := gate.Build(session.PostureAllow)
+			if err != nil {
+				t.Fatal(err)
+			}
+			decision := policy.Check("write", json.RawMessage(`{"path":"notes.txt","content":"ok"}`))
+			if guardian || decision.Action != approval.ActionAllow || decision.ProfileReadError != "" {
+				t.Fatalf("readable default-stage bypass changed: guardian=%v decision=%+v", guardian, decision)
+			}
+			if _, _, err := gate.Build(session.PostureAsk); err == nil {
+				t.Fatal("readable malformed JSON must still fail ordinary row construction")
+			}
+		})
+	}
+}

docs/changes/unreleased/2325-unreadable-config-limits.md

@@ -0,0 +1,25 @@
+---
+kind: fixed
+title: codeaf keeps unreadable profile approval limits in factory stages and task children
+pr: 2325
+surface: [engine, factory]
+invalidates:
+  - "A factory stage's allow posture bypassed an unreadable profile's saved approval rules. It now requires a person when those rules cannot be read."
+  - "Task workers, repair rounds and auditors dropped their parent's profile read failure. They now consult that signal at each decision and refuse calls that need a person when nobody can answer."
+  - "A project parse error could hide a known profile read failure. The stage now returns the failure already in hand before reading project rules."
+  - "Standalone run workers discarded profile read failures. They now check their run's profile before each call, including after work starts."
+  - "A refused job could lose the config-file reason during report recovery or first-turn budget admission. Its result now keeps that reason without futile report retries."
+---
+
+A readable profile keeps its existing behavior, including a factory stage's own
+allow posture. An unreadable profile cannot silently discard a saved denial:
+the existing approval gate reports the failed file read and requires a person.
+
+The named readable default-stage bypass includes readable malformed JSON.
+Other postures retain their existing parse-error behavior. Filesystem read
+failures remain distinct from JSON parse failures. Task, repair, checker and
+fleet children keep their own normal allowance while following the parent's
+current failed-read signal; recovery clears that inherited signal too.
+
+First-turn daily-limit refusals retain the config filename and read error.
+Admission decisions, spending counts and charges are unchanged.

docs/changes/unreleased/2354-agents-cannot-read-credentials.md

@@ -0,0 +1,10 @@
+---
+kind: changed
+title: Agents cannot read credentials in any approval mode
+pr: 2354
+surface: [chat, engine, app, factory]
+invalidates:
+  - "The conversation could run gh auth token and read credential files under the default allow; only a task was refused. Every agent is now refused, in every approval mode, and only a person's own ! command is exempt."
+  - "Agent shells inherited GH_TOKEN, GITHUB_TOKEN and other variables named like credentials, though provider keys were stripped. They are now removed too, and CODEAF_ALLOW_PROVIDER_KEYS_IN_SHELL keeps both."
+  - "The watch tool's commands and principal audit checks inherited the whole environment. They now get the same environment as any other agent shell."
+---

docs/changes/unreleased/2361-floor-reads.md

@@ -0,0 +1,18 @@
+---
+kind: fixed
+title: Idle factory polls read only job documents that changed
+pr: 2361
+surface: [engine]
+invalidates:
+  - "Idle lease pulls, settlement and runner capacity checks decoded every job document on each look. They now reuse unchanged documents after checking every file's identity, size and modification time."
+---
+
+Each List result owns its mutable data. Repository migration and output retention
+still run on each look. Writes from another store and replacements with unchanged
+size and modification time are visible on the next look. Every document is read
+again at least once a minute, so an in-place edit that keeps its size and time is
+seen within that bound.
+
+Lease grants and execution gates still read the job under its lock. A Stop that
+lands during selection permits only a thread-only lease for a pending reply,
+without restarting job steps.

docs/changes/unreleased/2365-usage-writer-drain.md

@@ -0,0 +1,14 @@
+---
+kind: fixed
+title: Closing a conversation waits for queued spending records
+pr: 2365
+surface: [engine, chat]
+invalidates:
+  - Closing a session previously left queued usage rows behind; it now drains that ledger under a bounded wait.
+---
+
+Closing a conversation now waits for its queued spending records, including the
+last call, before returning. A stalled ledger can still leave rows pending after
+the two-second limit on each close; that leaves one line in the log and counts
+those rows as not written down until they are. Other conversations sharing the ledger
+can keep recording spending.