fix: 7 bugs from discussion #149 smoke walk (envelope, spawn, CDC, observer) (#153)

* fix(cdc): emit pr_review_thread_resolved on replace polls (#152 bug 5)

writePRRows was DELETE-then-UPSERT on the Replace path, so every poll's
upserts hit the INSERT branch and the AFTER UPDATE trigger that emits
pr_review_thread_resolved never fired in production. Replaces the
blanket delete with a set-diff: upsert observed threads first (so
unchanged thread_ids go through ON CONFLICT DO UPDATE and fire the
UPDATE trigger when resolved flips), then delete orphans whose
thread_id is not in the observed set, all inside the existing tx.

Adds DeletePRReviewThread query (sqlc-generated form hand-edited; no
sqlc binary available locally — sqlc generate from backend/ produces an
identical file).

Tests: TestPRReviewThreadsCDC_EmitsResolvedOnReplacePoll (regression —
fails without fix) and TestPRReviewThreadsReplace_PrunesOrphansWithoutReinserting.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* fix(observe): emit scm-disabled log on startup with no subjects (#152 bug 7)

checkCredentials lived only inside Poll, which short-circuits when
discoverSubjects is empty. On a fresh daemon with no tracked PRs the
documented "scm observer disabled: provider credentials unavailable"
warn was unreachable, leaving users with no signal that the SCM
observer was a no-op.

Calls checkCredentials once in Observer.loop before the first Poll.
The existing credentialsChecked guard preserves once-per-process
semantics; provider construction still uses SkipTokenPreflight so
daemon readiness doesn't block on gh.

Test: TestStart_LogsDisabledWarningWhenNoTokenAndNoSubjects with a
race-safe syncBuffer for capturing slog from the observer goroutine.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* fix(api,spawn): typed errors + project/branch/binary preflight (#152 bugs 1-4,6)

Closes the long tail of opaque-500-and-orphan-row failures that
discussion #149's smoke walk surfaced. The common shape: spawn created
the session row before validating preconditions, and the underlying
errors weren't typed, so toAPIError defaulted to INTERNAL_ERROR.

Bug 1 (orphan row + opaque 500 on unknown projectId):
Service.Spawn / SpawnOrchestrator now call store.GetProject first and
return apierr.NotFound("PROJECT_NOT_FOUND", ...) before manager.Spawn,
eliminating the create-row-then-fail-workspace ordering.

Bug 2 (Restore opaque 500 on half-spawned/terminated session):
Manager.Restore gained the ErrIncompleteHandle guard that Kill has at
manager.go:189-193. toAPIError now maps both restore and kill to the
same SESSION_INCOMPLETE_HANDLE 409 envelope.

Bug 3 (--branch unfetched / checked-out-elsewhere → opaque 500):
gitworktree pre-checks listRecords for branch-in-other-worktree, falls
back to refs/tags on missing local/remote head, and emits two new port
sentinels (ErrWorkspaceBranchCheckedOutElsewhere,
ErrWorkspaceBranchNotFetched) mapped to BRANCH_CHECKED_OUT_ELSEWHERE
(409) and BRANCH_NOT_FETCHED (400).

Bug 4 (orphan terminated row on claim-pr rollback):
Adds Store.DeleteSession gated to seed-state rows only (preserves the
no-resurrection guarantee for live sessions), transactional change_log
cleanup, Manager.RollbackSpawn (delete-then-fallback-to-kill), a new
POST /sessions/{id}/rollback endpoint, and rewires
cli/spawn.rollbackSpawnedSession to use it. The exit-0 sub-symptom was
unreproducible from current source and is left unaddressed.

Bug 6 (agent binary not on PATH → silent idle session):
Drops the "return name, nil" anti-pattern from all 21 agent adapters
and returns the new ports.ErrAgentBinaryNotFound on exec.LookPath miss.
Manager.Spawn gained a validateAgentBinary pre-flight (with injectable
LookPath so tests don't need real binaries on PATH) that aborts before
runtime.Create. Mapped to AGENT_BINARY_NOT_FOUND (400). Integration
tests in internal/integration/ stub LookPath to /usr/bin/true.

Tests cover each bug end-to-end. OpenAPI regenerated for /rollback.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* chore: gofmt + regen frontend schema.ts for /rollback

CI fixes for #153:
- gofmt/goimports on kilocode and kiro adapters that the bug 6 audit
  left mis-grouped.
- openapi-typescript regen against the new /rollback endpoint added in
  the Lane A commit.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* fix(store): guard change_log delete behind seed probe + regen sqlc (#152, PR #153 review)

Addresses @greptile-apps P1 and P2 review feedback on PR #153.

P1 (CDC events deleted for live sessions in rollback fallback):
DeleteChangeLogForSession ran unconditionally inside the transaction
before DeleteSeedSession's seed-state predicates filtered the session
delete to a no-op. For a live session reaching DeleteSession (the
delete-then-kill fallback path inside RollbackSpawn), the seed delete
returned 0 rows but the session_created/session_updated CDC events
had already been purged. Now probes via a new SessionIsSeed query
first and short-circuits the whole tx — including the change_log
cleanup — when the row is not in seed state.

P2 (regen sqlc): installed sqlc 1.31.1 and ran `sqlc generate` from
backend/, replacing the hand-edited pr_review_threads.sql.go (and
producing minor format-only churn in models.go, pr.sql.go,
sessions.sql.go, changelog.sql.go).

The regen surfaced two issues:

1. GetPR / ListPRsBySession had their return types hand-changed to
   gen.PR by the previous PR; sqlc actually emits GetPRRow /
   ListPRsBySessionRow when queries enumerate columns. Fixed by
   collapsing those two queries to `SELECT * FROM pr` so sqlc returns
   gen.PR (which is what the store's prRowFromGen converter expects),
   and pr.last_nudge_signature now lands in the result alongside the
   existing 37 columns.

2. sqlc 1.31.1's SQLite parser silently strips trailing `?`
   placeholders and string literals from DELETE statements (reproduced
   with sqlc.arg, IFNULL, rowid subquery, and second-predicate
   workarounds — all eaten). DeleteSeedSession and
   DeleteChangeLogForSession both tripped it. They are now run as
   plain tx.ExecContext calls inside Store.DeleteSession, inside the
   same write transaction as SessionIsSeed; both queries are removed
   from the queries/ directory and the workaround context is
   documented inline in queries/sessions.sql and queries/changelog.sql
   to keep future contributors from re-adding them.

Verified: go build ./... clean, go test -race ./... 1097/1097 pass.

backend/internal/adapters/agent/agy/agy.go

@@ -5,6 +5,7 @@ package agy
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -182,7 +183,7 @@ func ResolveAgyBinary(ctx context.Context) (string, error) {
 			}
 		}
 
-		return "agy", nil
+		return "", fmt.Errorf("agy: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("agy"); err == nil && path != "" {
@@ -210,7 +211,7 @@ func ResolveAgyBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "agy", nil
+	return "", fmt.Errorf("agy: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) agyBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/aider/aider.go

@@ -10,6 +10,7 @@ package aider
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -173,7 +174,7 @@ func ResolveAiderBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "aider", nil
+		return "", fmt.Errorf("aider: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("aider"); err == nil && path != "" {
@@ -197,7 +198,7 @@ func ResolveAiderBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "aider", nil
+	return "", fmt.Errorf("aider: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) aiderBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/aider/aider_test.go

@@ -281,11 +281,17 @@ func TestContextCancellation(t *testing.T) {
 }
 
 func TestResolveAiderBinaryFallback(t *testing.T) {
+	// When the binary is not on PATH or any well-known location, the resolver
+	// MUST surface ports.ErrAgentBinaryNotFound rather than a silent string
+	// fallback that lets a missing CLI launch into an empty zellij pane.
 	bin, err := ResolveAiderBinary(context.Background())
 	if err != nil {
-		t.Fatalf("err: %v", err)
+		if !errors.Is(err, ports.ErrAgentBinaryNotFound) {
+			t.Fatalf("err = %v, want ports.ErrAgentBinaryNotFound", err)
+		}
+		return
 	}
 	if bin == "" {
-		t.Fatal("ResolveAiderBinary returned empty string")
+		t.Fatal("ResolveAiderBinary returned empty string with no error")
 	}
 }

backend/internal/adapters/agent/amp/amp.go

@@ -8,6 +8,7 @@ package amp
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -176,7 +177,7 @@ func ResolveAmpBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "amp", nil
+		return "", fmt.Errorf("amp: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("amp"); err == nil && path != "" {
@@ -203,7 +204,7 @@ func ResolveAmpBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "amp", nil
+	return "", fmt.Errorf("amp: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) ampBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/auggie/auggie.go

@@ -38,6 +38,7 @@ package auggie
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -201,7 +202,7 @@ func ResolveAuggieBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "auggie", nil
+		return "", fmt.Errorf("auggie: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("auggie"); err == nil && path != "" {
@@ -229,7 +230,7 @@ func ResolveAuggieBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "auggie", nil
+	return "", fmt.Errorf("auggie: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) auggieBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/auggie/auggie_test.go

@@ -184,14 +184,18 @@ func TestSessionInfoNoOp(t *testing.T) {
 }
 
 func TestResolveAuggieBinaryFallback(t *testing.T) {
-	// With a cancelled context the resolver returns the context error rather than
+	// When the binary is not on PATH or any well-known location, the resolver
-	// a binary path; with a live context it always yields a non-empty path.
+	// MUST surface ports.ErrAgentBinaryNotFound rather than a silent string
+	// fallback that lets a missing CLI launch into an empty zellij pane.
 	bin, err := ResolveAuggieBinary(context.Background())
 	if err != nil {
-		t.Fatalf("ResolveAuggieBinary err = %v", err)
+		if !errors.Is(err, ports.ErrAgentBinaryNotFound) {
+			t.Fatalf("err = %v, want ports.ErrAgentBinaryNotFound", err)
+		}
+		return
 	}
 	if bin == "" {
-		t.Fatal("ResolveAuggieBinary returned empty path")
+		t.Fatal("ResolveAuggieBinary returned empty path with no error")
 	}
 }
 

backend/internal/adapters/agent/autohand/autohand.go

@@ -12,6 +12,7 @@ package autohand
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -225,7 +226,7 @@ func ResolveAutohandBinary(ctx context.Context) (string, error) {
 			}
 		}
 
-		return "autohand", nil
+		return "", fmt.Errorf("autohand: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("autohand"); err == nil && path != "" {
@@ -252,7 +253,7 @@ func ResolveAutohandBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "autohand", nil
+	return "", fmt.Errorf("autohand: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) autohandBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/claudecode/claudecode.go

@@ -306,7 +306,7 @@ func ResolveClaudeBinary(ctx context.Context) (string, error) {
 				return candidate, nil
 			}
 		}
-		return "claude", nil
+		return "", fmt.Errorf("claude: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("claude"); err == nil && path != "" {
@@ -333,7 +333,7 @@ func ResolveClaudeBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "claude", nil
+	return "", fmt.Errorf("claude: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) claudeBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/cline/cline.go

@@ -14,6 +14,7 @@ package cline
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -178,7 +179,7 @@ func ResolveClineBinary(ctx context.Context) (string, error) {
 			}
 		}
 
-		return "cline", nil
+		return "", fmt.Errorf("cline: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("cline"); err == nil && path != "" {
@@ -206,7 +207,7 @@ func ResolveClineBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "cline", nil
+	return "", fmt.Errorf("cline: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) clineBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/codex/codex.go

@@ -8,6 +8,7 @@ package codex
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -175,7 +176,7 @@ func ResolveCodexBinary(ctx context.Context) (string, error) {
 			}
 		}
 
-		return "codex", nil
+		return "", fmt.Errorf("codex: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("codex"); err == nil && path != "" {
@@ -202,7 +203,7 @@ func ResolveCodexBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "codex", nil
+	return "", fmt.Errorf("codex: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) codexBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/continueagent/continueagent.go

@@ -22,6 +22,7 @@ package continueagent
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -197,7 +198,7 @@ func ResolveContinueBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "cn", nil
+		return "", fmt.Errorf("cn: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("cn"); err == nil && path != "" {
@@ -225,7 +226,7 @@ func ResolveContinueBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "cn", nil
+	return "", fmt.Errorf("cn: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) continueBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/copilot/copilot.go

@@ -19,6 +19,7 @@ package copilot
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -189,7 +190,7 @@ func ResolveCopilotBinary(ctx context.Context) (string, error) {
 			}
 		}
 
-		return "copilot", nil
+		return "", fmt.Errorf("copilot: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("copilot"); err == nil && path != "" {
@@ -217,7 +218,7 @@ func ResolveCopilotBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "copilot", nil
+	return "", fmt.Errorf("copilot: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) copilotBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/crush/crush.go

@@ -8,6 +8,7 @@ package crush
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -191,7 +192,7 @@ func ResolveCrushBinary(ctx context.Context) (string, error) {
 			}
 		}
 
-		return "crush", nil
+		return "", fmt.Errorf("crush: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("crush"); err == nil && path != "" {
@@ -218,7 +219,7 @@ func ResolveCrushBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "crush", nil
+	return "", fmt.Errorf("crush: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) crushBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/cursor/cursor.go

@@ -9,6 +9,7 @@ package cursor
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -165,7 +166,7 @@ func ResolveCursorBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "cursor-agent", nil
+		return "", fmt.Errorf("cursor: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("cursor-agent"); err == nil && path != "" {
@@ -190,7 +191,7 @@ func ResolveCursorBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "cursor-agent", nil
+	return "", fmt.Errorf("cursor: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) cursorBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/devin/devin.go

@@ -24,6 +24,7 @@ package devin
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -200,7 +201,7 @@ func ResolveDevinBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "devin", nil
+		return "", fmt.Errorf("devin: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("devin"); err == nil && path != "" {
@@ -227,7 +228,7 @@ func ResolveDevinBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "devin", nil
+	return "", fmt.Errorf("devin: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) devinBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/devin/devin_test.go

@@ -2,6 +2,7 @@ package devin
 
 import (
 	"context"
+	"errors"
 	"reflect"
 	"strings"
 	"testing"
@@ -254,14 +255,18 @@ func TestGetAgentHooksCtxCancelled(t *testing.T) {
 }
 
 func TestResolveDevinBinaryFallback(t *testing.T) {
-	// When devin is not on PATH or well-known locations, fall back to the bare
+	// When the binary is not on PATH or any well-known location, the resolver
-	// name so exec can still attempt to launch it.
+	// MUST surface ports.ErrAgentBinaryNotFound rather than a silent string
+	// fallback that lets a missing CLI launch into an empty zellij pane.
 	bin, err := ResolveDevinBinary(context.Background())
 	if err != nil {
-		t.Fatalf("err: %v", err)
+		if !errors.Is(err, ports.ErrAgentBinaryNotFound) {
+			t.Fatalf("err = %v, want ports.ErrAgentBinaryNotFound", err)
+		}
+		return
 	}
 	if bin == "" {
-		t.Fatal("empty binary path")
+		t.Fatal("ResolveDevinBinary returned empty path with no error")
 	}
 }
 

backend/internal/adapters/agent/droid/droid.go

@@ -288,7 +288,7 @@ func ResolveDroidBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "droid", nil
+		return "", fmt.Errorf("droid: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("droid"); err == nil && path != "" {
@@ -315,7 +315,7 @@ func ResolveDroidBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "droid", nil
+	return "", fmt.Errorf("droid: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) droidBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/goose/goose.go

@@ -273,7 +273,7 @@ func ResolveGooseBinary(ctx context.Context) (string, error) {
 			}
 		}
 
-		return "goose", nil
+		return "", fmt.Errorf("goose: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("goose"); err == nil && path != "" {
@@ -301,7 +301,7 @@ func ResolveGooseBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "goose", nil
+	return "", fmt.Errorf("goose: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) gooseBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/goose/goose_test.go

@@ -3,10 +3,10 @@ package goose
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"os"
 	"path/filepath"
 	"reflect"
-	"strings"
 	"testing"
 
 	"github.com/aoagents/agent-orchestrator/backend/internal/ports"
@@ -380,17 +380,18 @@ func TestSessionInfoFalseWhenNoHookMetadata(t *testing.T) {
 }
 
 func TestResolveGooseBinaryFallback(t *testing.T) {
-	// On a machine without goose on PATH or any well-known location, resolution
+	// When the binary is not on PATH or any well-known location, the resolver
-	// still returns a usable last-ditch "goose" name rather than empty.
+	// MUST surface ports.ErrAgentBinaryNotFound rather than a silent string
-	got, err := ResolveGooseBinary(context.Background())
+	// fallback that lets a missing CLI launch into an empty zellij pane.
+	bin, err := ResolveGooseBinary(context.Background())
 	if err != nil {
-		t.Fatalf("ResolveGooseBinary err = %v", err)
+		if !errors.Is(err, ports.ErrAgentBinaryNotFound) {
+			t.Fatalf("err = %v, want ports.ErrAgentBinaryNotFound", err)
+		}
+		return
 	}
-	if got == "" {
+	if bin == "" {
-		t.Fatal("ResolveGooseBinary returned empty binary")
+		t.Fatal("ResolveGooseBinary returned empty path with no error")
-	}
-	if !strings.Contains(got, "goose") {
-		t.Fatalf("ResolveGooseBinary = %q, want a path containing goose", got)
 	}
 }
 

backend/internal/adapters/agent/grok/grok.go

@@ -16,6 +16,7 @@ package grok
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -212,7 +213,7 @@ func ResolveGrokBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "grok", nil
+		return "", fmt.Errorf("grok: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("grok"); err == nil && path != "" {
@@ -239,7 +240,7 @@ func ResolveGrokBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "grok", nil
+	return "", fmt.Errorf("grok: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) grokBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/kilocode/kilocode.go

@@ -24,6 +24,7 @@ package kilocode
 import (
 	"context"
 	"encoding/json"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -269,7 +270,7 @@ func ResolveKilocodeBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "kilocode", nil
+		return "", fmt.Errorf("kilocode: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("kilocode"); err == nil && path != "" {
@@ -297,7 +298,7 @@ func ResolveKilocodeBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "kilocode", nil
+	return "", fmt.Errorf("kilocode: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) kilocodeBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/kimi/kimi.go

@@ -17,6 +17,7 @@ package kimi
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -218,7 +219,7 @@ func ResolveKimiBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "kimi", nil
+		return "", fmt.Errorf("kimi: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("kimi"); err == nil && path != "" {
@@ -245,7 +246,7 @@ func ResolveKimiBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "kimi", nil
+	return "", fmt.Errorf("kimi: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) kimiBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/kiro/kiro.go

@@ -20,6 +20,7 @@ package kiro
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -189,7 +190,7 @@ func ResolveKiroBinary(ctx context.Context) (string, error) {
 			}
 		}
 
-		return "kiro-cli", nil
+		return "", fmt.Errorf("kiro: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("kiro-cli"); err == nil && path != "" {
@@ -216,7 +217,7 @@ func ResolveKiroBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "kiro-cli", nil
+	return "", fmt.Errorf("kiro: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) kiroBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/kiro/kiro_test.go

@@ -3,6 +3,7 @@ package kiro
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"os"
 	"path/filepath"
 	"reflect"
@@ -384,12 +385,18 @@ func TestDeriveActivityState(t *testing.T) {
 }
 
 func TestResolveKiroBinaryFallback(t *testing.T) {
-	got, err := ResolveKiroBinary(context.Background())
+	// When the binary is not on PATH or any well-known location, the resolver
+	// MUST surface ports.ErrAgentBinaryNotFound rather than a silent string
+	// fallback that lets a missing CLI launch into an empty zellij pane.
+	bin, err := ResolveKiroBinary(context.Background())
 	if err != nil {
-		t.Fatal(err)
+		if !errors.Is(err, ports.ErrAgentBinaryNotFound) {
+			t.Fatalf("err = %v, want ports.ErrAgentBinaryNotFound", err)
+		}
+		return
 	}
-	if got == "" {
+	if bin == "" {
-		t.Fatal("ResolveKiroBinary returned empty path")
+		t.Fatal("ResolveKiroBinary returned empty path with no error")
 	}
 }
 

backend/internal/adapters/agent/pi/pi.go

@@ -33,6 +33,7 @@ package pi
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -190,7 +191,7 @@ func ResolvePiBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "pi", nil
+		return "", fmt.Errorf("pi: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("pi"); err == nil && path != "" {
@@ -218,7 +219,7 @@ func ResolvePiBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "pi", nil
+	return "", fmt.Errorf("pi: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) piBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/qwen/qwen.go

@@ -15,6 +15,7 @@ package qwen
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -144,8 +145,9 @@ func (p *Plugin) SessionInfo(ctx context.Context, session ports.SessionRef) (por
 
 // ResolveQwenBinary returns the path to the qwen binary on this machine,
 // searching PATH then a handful of well-known install locations (Homebrew, npm
-// global). Returns "qwen" as a last-ditch fallback so callers see a clear
+// global). Returns ports.ErrAgentBinaryNotFound when none of those find the
-// "command not found" rather than an empty argv.
+// binary — better than the previous silent `"qwen"` fallback, which let an
+// empty zellij pane masquerade as a live session.
 func ResolveQwenBinary(ctx context.Context) (string, error) {
 	if err := ctx.Err(); err != nil {
 		return "", err
@@ -178,7 +180,7 @@ func ResolveQwenBinary(ctx context.Context) (string, error) {
 			}
 		}
 
-		return "qwen", nil
+		return "", fmt.Errorf("qwen: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("qwen"); err == nil && path != "" {
@@ -206,7 +208,7 @@ func ResolveQwenBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "qwen", nil
+	return "", fmt.Errorf("qwen: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) qwenBinary(ctx context.Context) (string, error) {

backend/internal/adapters/agent/vibe/vibe.go

@@ -26,6 +26,7 @@ package vibe
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -197,7 +198,7 @@ func ResolveVibeBinary(ctx context.Context) (string, error) {
 				return "", err
 			}
 		}
-		return "vibe", nil
+		return "", fmt.Errorf("vibe: %w", ports.ErrAgentBinaryNotFound)
 	}
 
 	if path, err := exec.LookPath("vibe"); err == nil && path != "" {
@@ -224,7 +225,7 @@ func ResolveVibeBinary(ctx context.Context) (string, error) {
 		}
 	}
 
-	return "vibe", nil
+	return "", fmt.Errorf("vibe: %w", ports.ErrAgentBinaryNotFound)
 }
 
 func (p *Plugin) vibeBinary(ctx context.Context) (string, error) {

backend/internal/adapters/workspace/gitworktree/workspace.go

@@ -24,6 +24,16 @@ var (
 	ErrUnsafePath = errors.New("gitworktree: unsafe workspace path")
 )
 
+// ErrBranchCheckedOutElsewhere and ErrBranchNotFetched are adapter-local aliases
+// of the port-level sentinels: they preserve the gitworktree-prefixed message
+// while letting the service layer match on ports.ErrWorkspaceBranchCheckedOutElsewhere
+// / ports.ErrWorkspaceBranchNotFetched without importing this package. Tests
+// inside the adapter use these names; callers outside use the port sentinels.
+var (
+	ErrBranchCheckedOutElsewhere = ports.ErrWorkspaceBranchCheckedOutElsewhere
+	ErrBranchNotFetched          = ports.ErrWorkspaceBranchNotFetched
+)
+
 // RepoResolver maps a project to the absolute path of its source git repo.
 type RepoResolver interface {
 	RepoPath(projectID domain.ProjectID) (string, error)
@@ -195,6 +205,17 @@ func (w *Workspace) Restore(ctx context.Context, cfg ports.WorkspaceConfig) (por
 }
 
 func (w *Workspace) addWorktree(ctx context.Context, repo, path, branch string) error {
+	// Refuse early if the branch is already checked out in another worktree:
+	// `git worktree add` will fail, but its stderr leaks through as an opaque
+	// 500. A typed sentinel lets the HTTP layer surface a 409.
+	records, err := w.listRecords(ctx, repo)
+	if err != nil {
+		return err
+	}
+	if conflict, ok := findWorktreeByBranch(records, branch); ok && filepath.Clean(conflict.Path) != filepath.Clean(path) {
+		return fmt.Errorf("%w: %q is checked out at %q", ErrBranchCheckedOutElsewhere, branch, conflict.Path)
+	}
+
 	localBranch, err := w.refExists(ctx, repo, "refs/heads/"+branch)
 	if err != nil {
 		return err
@@ -205,8 +226,18 @@ func (w *Workspace) addWorktree(ctx context.Context, repo, path, branch string)
 		}
 		return nil
 	}
+
+	// `worktree add -b <branch> <path> <base>` creates a fresh local branch from
+	// <base>. resolveBaseRef tries `origin/<branch>` first, so a fetched-but-
+	// not-checked-out remote branch auto-tracks cleanly via that path. If
+	// neither origin/<branch>, the default branch, nor any tag is reachable,
+	// the branch genuinely has no base — surface ErrBranchNotFetched so callers
+	// can suggest `git fetch`.
 	baseRef, err := w.resolveBaseRef(ctx, repo, branch)
 	if err != nil {
+		if errors.Is(err, errNoBaseRef) {
+			return fmt.Errorf("%w: %q has no local head, no remote, and no tag — run `git fetch` then retry", ErrBranchNotFetched, branch)
+		}
 		return err
 	}
 	if _, err := w.run(ctx, w.binary, worktreeAddNewBranchArgs(repo, branch, path, baseRef)...); err != nil {
@@ -222,6 +253,10 @@ func (w *Workspace) validateBranch(ctx context.Context, repo, branch string) err
 	return nil
 }
 
+// errNoBaseRef is an internal sentinel: every candidate base ref is missing.
+// addWorktree translates it into ErrBranchNotFetched.
+var errNoBaseRef = errors.New("gitworktree: no base ref found")
+
 func (w *Workspace) resolveBaseRef(ctx context.Context, repo, branch string) (string, error) {
 	candidates := baseRefCandidates(branch, w.defaultBranch)
 	for _, ref := range candidates {
@@ -233,7 +268,17 @@ func (w *Workspace) resolveBaseRef(ctx context.Context, repo, branch string) (st
 			return ref, nil
 		}
 	}
-	return "", fmt.Errorf("gitworktree: no base ref found for branch %q (tried %s)", branch, strings.Join(candidates, ", "))
+	// Also probe a same-named tag so requests like `--branch v1.2.3` can
+	// auto-track when the tag is fetched but no branch ref exists.
+	tagRef := "refs/tags/" + branch
+	exists, err := w.refExists(ctx, repo, tagRef)
+	if err != nil {
+		return "", err
+	}
+	if exists {
+		return tagRef, nil
+	}
+	return "", fmt.Errorf("%w for branch %q (tried %s, %s)", errNoBaseRef, branch, strings.Join(candidates, ", "), tagRef)
 }
 
 func (w *Workspace) refExists(ctx context.Context, repo, ref string) (bool, error) {
@@ -382,6 +427,15 @@ func findWorktree(records []worktreeRecord, path string) (worktreeRecord, bool)
 	return worktreeRecord{}, false
 }
 
+func findWorktreeByBranch(records []worktreeRecord, branch string) (worktreeRecord, bool) {
+	for _, rec := range records {
+		if rec.Branch == branch {
+			return rec, true
+		}
+	}
+	return worktreeRecord{}, false
+}
+
 func pathExistsNonEmpty(path string) (bool, error) {
 	entries, err := os.ReadDir(path)
 	if err == nil {

backend/internal/adapters/workspace/gitworktree/workspace_test.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"reflect"
 	"strings"
@@ -228,6 +229,77 @@ func TestDestroyRefusesStillRegisteredPathAndPreservesDirectory(t *testing.T) {
 	}
 }
 
+// TestAddWorktreeRefusesBranchCheckedOutElsewhere covers Bug 3 (a): if the
+// requested branch is already checked out in another worktree of the same repo,
+// Create must surface ports.ErrWorkspaceBranchCheckedOutElsewhere so the HTTP
+// layer can render a typed 409 instead of leaking raw git stderr through a 500.
+func TestAddWorktreeRefusesBranchCheckedOutElsewhere(t *testing.T) {
+	root := t.TempDir()
+	repo := t.TempDir()
+	ws, err := New(Options{ManagedRoot: root, RepoResolver: StaticRepoResolver{"proj": repo}})
+	if err != nil {
+		t.Fatalf("new: %v", err)
+	}
+	otherPath := filepath.Join(root, "proj", "other")
+	ws.run = func(_ context.Context, _ string, args ...string) ([]byte, error) {
+		joined := strings.Join(args, " ")
+		switch {
+		case strings.Contains(joined, "check-ref-format"):
+			return nil, nil
+		case strings.Contains(joined, "worktree list --porcelain"):
+			return []byte("worktree " + otherPath + "\nbranch refs/heads/feature/x\n"), nil
+		case strings.Contains(joined, "rev-parse"):
+			return []byte("commit"), nil
+		default:
+			t.Fatalf("unexpected git invocation: %v", args)
+			return nil, nil
+		}
+	}
+	_, err = ws.Create(context.Background(), ports.WorkspaceConfig{ProjectID: "proj", SessionID: "sess", Branch: "feature/x"})
+	if !errors.Is(err, ports.ErrWorkspaceBranchCheckedOutElsewhere) {
+		t.Fatalf("err = %v, want ports.ErrWorkspaceBranchCheckedOutElsewhere", err)
+	}
+	if !strings.Contains(err.Error(), otherPath) {
+		t.Fatalf("err = %v, want message to include conflicting path %q", err, otherPath)
+	}
+}
+
+// TestAddWorktreeReportsBranchNotFetched covers Bug 3 (b): if no local head,
+// no origin remote-tracking branch, no default branch ref, and no tag of the
+// same name is reachable, Create must surface ports.ErrWorkspaceBranchNotFetched
+// so the HTTP layer can render a typed 400 with a `git fetch` suggestion.
+func TestAddWorktreeReportsBranchNotFetched(t *testing.T) {
+	root := t.TempDir()
+	repo := t.TempDir()
+	ws, err := New(Options{ManagedRoot: root, RepoResolver: StaticRepoResolver{"proj": repo}})
+	if err != nil {
+		t.Fatalf("new: %v", err)
+	}
+	// Build a real exit-1 error so refExists treats every probe as "absent".
+	exitOne := func() error {
+		cmd := exec.Command("sh", "-c", "exit 1")
+		return cmd.Run()
+	}()
+	ws.run = func(_ context.Context, _ string, args ...string) ([]byte, error) {
+		joined := strings.Join(args, " ")
+		switch {
+		case strings.Contains(joined, "check-ref-format"):
+			return nil, nil
+		case strings.Contains(joined, "worktree list --porcelain"):
+			return nil, nil
+		case strings.Contains(joined, "rev-parse"):
+			return nil, commandError{args: args, err: exitOne}
+		default:
+			t.Fatalf("unexpected git invocation: %v", args)
+			return nil, nil
+		}
+	}
+	_, err = ws.Create(context.Background(), ports.WorkspaceConfig{ProjectID: "proj", SessionID: "sess", Branch: "feature/missing"})
+	if !errors.Is(err, ports.ErrWorkspaceBranchNotFetched) {
+		t.Fatalf("err = %v, want ports.ErrWorkspaceBranchNotFetched", err)
+	}
+}
+
 func mkdirFile(dir, name string) error {
 	if err := os.MkdirAll(dir, 0o755); err != nil {
 		return err

backend/internal/cli/dto_drift_e2e_test.go

@@ -78,6 +78,10 @@ func (f *fakeSessionService) Kill(context.Context, domain.SessionID) (bool, erro
 	return false, nil
 }
 
+func (f *fakeSessionService) RollbackSpawn(context.Context, domain.SessionID) (sessionsvc.RollbackOutcome, error) {
+	return sessionsvc.RollbackOutcome{}, nil
+}
+
 func (f *fakeSessionService) Cleanup(context.Context, domain.ProjectID) ([]domain.SessionID, error) {
 	return nil, nil
 }

backend/internal/cli/spawn.go

@@ -133,7 +133,21 @@ func newSpawnCommand(ctx *commandContext) *cobra.Command {
 	return cmd
 }
 
+// rollbackSpawnedSession reverses a partial `spawn` whose out-of-band follow-up
+// (PR claim) failed. It calls the daemon's `/rollback` endpoint, which deletes
+// the seed-state row outright instead of marking it terminated — so the user
+// does not see an orphan terminated session under `--include-terminated`. If
+// spawn output has already landed (workspace + runtime), the daemon falls back
+// to a Kill on the server side so teardown still happens.
 func (c *commandContext) rollbackSpawnedSession(ctx context.Context, id string) error {
-	var res killSessionResponse
+	var res rollbackSessionResponse
-	return c.postJSON(ctx, "sessions/"+url.PathEscape(id)+"/kill", struct{}{}, &res)
+	return c.postJSON(ctx, "sessions/"+url.PathEscape(id)+"/rollback", struct{}{}, &res)
+}
+
+// rollbackSessionResponse mirrors the daemon's RollbackSessionResponse body.
+type rollbackSessionResponse struct {
+	OK        bool   `json:"ok"`
+	SessionID string `json:"sessionId"`
+	Deleted   bool   `json:"deleted,omitempty"`
+	Killed    bool   `json:"killed,omitempty"`
 }

backend/internal/cli/spawn_test.go

@@ -98,9 +98,9 @@ func TestSpawnClaimPRFailureRollsBackSession(t *testing.T) {
 			}
 			w.WriteHeader(http.StatusNotFound)
 			_, _ = io.WriteString(w, `{"error":"not_found","code":"PR_NOT_FOUND","message":"PR not found"}`)
-		case r.Method == http.MethodPost && r.URL.Path == "/api/v1/sessions/demo-10/kill":
+		case r.Method == http.MethodPost && r.URL.Path == "/api/v1/sessions/demo-10/rollback":
 			delete(sessions, "demo-10")
-			_, _ = io.WriteString(w, `{"ok":true,"sessionId":"demo-10","freed":true}`)
+			_, _ = io.WriteString(w, `{"ok":true,"sessionId":"demo-10","deleted":true}`)
 		default:
 			http.NotFound(w, r)
 		}
@@ -119,7 +119,7 @@ func TestSpawnClaimPRFailureRollsBackSession(t *testing.T) {
 	if sessions["demo-10"] {
 		t.Fatalf("spawned session still present after claim rollback: %#v", sessions)
 	}
-	want := []string{"GET /api/v1/projects/demo", "POST /api/v1/sessions", "POST /api/v1/sessions/demo-10/pr/claim", "POST /api/v1/sessions/demo-10/kill"}
+	want := []string{"GET /api/v1/projects/demo", "POST /api/v1/sessions", "POST /api/v1/sessions/demo-10/pr/claim", "POST /api/v1/sessions/demo-10/rollback"}
 	if !reflect.DeepEqual(requests, want) {
 		t.Fatalf("requests=%#v want %#v", requests, want)
 	}

backend/internal/httpd/apispec/openapi.yaml

@@ -775,6 +775,46 @@ paths:
       summary: Restore a terminated session
       tags:
       - sessions
+  /api/v1/sessions/{sessionId}/rollback:
+    post:
+      operationId: rollbackSession
+      parameters:
+      - description: Session identifier, e.g. project-1.
+        in: path
+        name: sessionId
+        required: true
+        schema:
+          description: Session identifier, e.g. project-1.
+          type: string
+      responses:
+        "200":
+          content:
+            application/json:
+              schema:
+                $ref: '#/components/schemas/RollbackSessionResponse'
+          description: OK
+        "404":
+          content:
+            application/json:
+              schema:
+                $ref: '#/components/schemas/APIError'
+          description: Not Found
+        "409":
+          content:
+            application/json:
+              schema:
+                $ref: '#/components/schemas/APIError'
+          description: Conflict
+        "500":
+          content:
+            application/json:
+              schema:
+                $ref: '#/components/schemas/APIError'
+          description: Internal Server Error
+      summary: Undo a partially-completed spawn (delete seed row, or kill if spawn
+        output exists)
+      tags:
+      - sessions
   /api/v1/sessions/{sessionId}/send:
     post:
       operationId: sendSessionMessage
@@ -1149,6 +1189,20 @@ components:
       - sessionId
       - session
       type: object
+    RollbackSessionResponse:
+      properties:
+        deleted:
+          type: boolean
+        killed:
+          type: boolean
+        ok:
+          type: boolean
+        sessionId:
+          type: string
+      required:
+      - ok
+      - sessionId
+      type: object
     SCMConfig:
       properties:
         package:

backend/internal/httpd/apispec/specgen/build.go

@@ -137,6 +137,7 @@ var schemaNames = map[string]string{
 	"ControllersRestoreSessionResponse":     "RestoreSessionResponse",
 	"ControllersCleanupSessionsResponse":    "CleanupSessionsResponse",
 	"ControllersKillSessionResponse":        "KillSessionResponse",
+	"ControllersRollbackSessionResponse":    "RollbackSessionResponse",
 	"ControllersSendSessionMessageRequest":  "SendSessionMessageRequest",
 	"ControllersSendSessionMessageResponse": "SendSessionMessageResponse",
 	"ControllersClaimPRResponse":            "ClaimPRResponse",
@@ -414,6 +415,17 @@ func sessionOperations() []operation {
 				{http.StatusInternalServerError, envelope.APIError{}},
 			},
 		},
+		{
+			method: http.MethodPost, path: "/api/v1/sessions/{sessionId}/rollback", id: "rollbackSession", tag: "sessions",
+			summary:    "Undo a partially-completed spawn (delete seed row, or kill if spawn output exists)",
+			pathParams: []any{controllers.SessionIDParam{}},
+			resps: []respUnit{
+				{http.StatusOK, controllers.RollbackSessionResponse{}},
+				{http.StatusNotFound, envelope.APIError{}},
+				{http.StatusConflict, envelope.APIError{}},
+				{http.StatusInternalServerError, envelope.APIError{}},
+			},
+		},
 		{
 			method: http.MethodPost, path: "/api/v1/sessions/{sessionId}/send", id: "sendSessionMessage", tag: "sessions",
 			summary:    "Send a message to a running session's agent",

backend/internal/httpd/controllers/dto.go

@@ -158,6 +158,16 @@ type KillSessionResponse struct {
 	Freed     bool             `json:"freed,omitempty"`
 }
 
+// RollbackSessionResponse is the body of POST /api/v1/sessions/{sessionId}/rollback.
+// Exactly one of Deleted/Killed is true on a successful rollback; both are
+// false when the session was already absent or already terminated (benign).
+type RollbackSessionResponse struct {
+	OK        bool             `json:"ok"`
+	SessionID domain.SessionID `json:"sessionId"`
+	Deleted   bool             `json:"deleted,omitempty"`
+	Killed    bool             `json:"killed,omitempty"`
+}
+
 // CleanupSessionsResponse is the body of POST /api/v1/sessions/cleanup.
 type CleanupSessionsResponse struct {
 	OK      bool               `json:"ok"`

backend/internal/httpd/controllers/sessions.go

@@ -30,6 +30,7 @@ type SessionService interface {
 	Get(ctx context.Context, id domain.SessionID) (domain.Session, error)
 	Restore(ctx context.Context, id domain.SessionID) (domain.Session, error)
 	Kill(ctx context.Context, id domain.SessionID) (bool, error)
+	RollbackSpawn(ctx context.Context, id domain.SessionID) (sessionsvc.RollbackOutcome, error)
 	Cleanup(ctx context.Context, project domain.ProjectID) ([]domain.SessionID, error)
 	Rename(ctx context.Context, id domain.SessionID, displayName string) error
 	Send(ctx context.Context, id domain.SessionID, message string) error
@@ -64,6 +65,7 @@ func (c *SessionsController) Register(r chi.Router) {
 	r.Patch("/sessions/{sessionId}", c.rename)
 	r.Post("/sessions/{sessionId}/restore", c.restore)
 	r.Post("/sessions/{sessionId}/kill", c.kill)
+	r.Post("/sessions/{sessionId}/rollback", c.rollback)
 	r.Post("/sessions/{sessionId}/send", c.send)
 	r.Post("/sessions/{sessionId}/activity", c.activity)
 	r.Get("/orchestrators", c.listOrchestrators)
@@ -218,6 +220,25 @@ func (c *SessionsController) kill(w http.ResponseWriter, r *http.Request) {
 	envelope.WriteJSON(w, http.StatusOK, KillSessionResponse{OK: true, SessionID: sessionID(r), Freed: freed})
 }
 
+// rollback undoes a partially-completed spawn: if the session row is still in
+// seed state (no workspace, no runtime handle yet), the row is deleted
+// outright. If anything observable has landed it falls back to Kill so the
+// runtime/workspace are torn down. Used by `ao spawn --claim-pr` to undo a
+// session whose claim step failed, avoiding the orphan terminated row a
+// plain Kill would leave behind.
+func (c *SessionsController) rollback(w http.ResponseWriter, r *http.Request) {
+	if c.Svc == nil {
+		apispec.NotImplemented(w, r, "POST", "/api/v1/sessions/{sessionId}/rollback")
+		return
+	}
+	out, err := c.Svc.RollbackSpawn(r.Context(), sessionID(r))
+	if err != nil {
+		envelope.WriteError(w, r, err)
+		return
+	}
+	envelope.WriteJSON(w, http.StatusOK, RollbackSessionResponse{OK: true, SessionID: sessionID(r), Deleted: out.Deleted, Killed: out.Killed})
+}
+
 func (c *SessionsController) cleanup(w http.ResponseWriter, r *http.Request) {
 	if c.Svc == nil {
 		apispec.NotImplemented(w, r, "POST", "/api/v1/sessions/cleanup")

backend/internal/httpd/controllers/sessions_test.go

@@ -92,6 +92,14 @@ func (f *fakeSessionService) Kill(_ context.Context, id domain.SessionID) (bool,
 	return true, nil
 }
 
+func (f *fakeSessionService) RollbackSpawn(_ context.Context, id domain.SessionID) (sessionsvc.RollbackOutcome, error) {
+	if _, ok := f.sessions[id]; ok {
+		delete(f.sessions, id)
+		return sessionsvc.RollbackOutcome{Deleted: true}, nil
+	}
+	return sessionsvc.RollbackOutcome{}, nil
+}
+
 func (f *fakeSessionService) Cleanup(_ context.Context, project domain.ProjectID) ([]domain.SessionID, error) {
 	f.cleanupProjects = append(f.cleanupProjects, project)
 	if f.cleanupResult != nil {

backend/internal/integration/lifecycle_sqlite_test.go

@@ -97,7 +97,7 @@ func newStack(t *testing.T) *stack {
 	prm := prsvc.New(prsvc.Deps{Writer: store, Lifecycle: lcm})
 	rt := &stubRuntime{}
 	ws := &stubWorkspace{}
-	mgr := sessionmanager.New(sessionmanager.Deps{Runtime: rt, Agents: stubAgents{}, Workspace: ws, Store: store, Messenger: msg, Lifecycle: lcm})
+	mgr := sessionmanager.New(sessionmanager.Deps{Runtime: rt, Agents: stubAgents{}, Workspace: ws, Store: store, Messenger: msg, Lifecycle: lcm, LookPath: func(string) (string, error) { return "/usr/bin/true", nil }})
 	sm := sessionsvc.New(mgr, store)
 	return &stack{store: store, sm: sm, lcm: lcm, prm: prm, rt: rt, ws: ws, msg: msg}
 }

backend/internal/observe/scm/observer.go

@@ -168,6 +168,16 @@ func (o *Observer) Start(ctx context.Context) <-chan struct{} {
 
 func (o *Observer) loop(ctx context.Context, done chan<- struct{}) {
 	defer close(done)
+	// Run the credential gate once before the first poll so the
+	// "scm observer disabled: provider credentials unavailable" warning is
+	// emitted on a fresh daemon even if discoverSubjects has no subjects yet
+	// (which would otherwise short-circuit Poll before checkCredentials).
+	// checkCredentials is guarded by credentialsChecked, so this remains
+	// once-per-process; a transient error here simply defers the check to the
+	// next Poll, matching existing behavior.
+	if _, err := o.checkCredentials(ctx); err != nil && !errors.Is(err, context.Canceled) {
+		o.logger.Error("scm observer: initial credential check failed", "err", err)
+	}
 	if err := o.Poll(ctx); err != nil && !errors.Is(err, context.Canceled) {
 		o.logger.Error("scm observer: initial poll failed", "err", err)
 	}

backend/internal/observe/scm/observer_test.go

@@ -5,12 +5,14 @@ package scm
 // review cadence, semantic hashes, and notification behavior stay provider-neutral.
 
 import (
+	"bytes"
 	"context"
 	"errors"
 	"fmt"
 	"io"
 	"log/slog"
 	"os/exec"
+	"strings"
 	"sync"
 	"testing"
 	"time"
@@ -310,6 +312,89 @@ func TestPoll_RetriesTransientCredentialErrors(t *testing.T) {
 	}
 }
 
+// syncBuffer is a goroutine-safe wrapper around bytes.Buffer for capturing
+// slog output emitted from the observer's background goroutine.
+type syncBuffer struct {
+	mu  sync.Mutex
+	buf bytes.Buffer
+}
+
+func (s *syncBuffer) Write(p []byte) (int, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.buf.Write(p)
+}
+
+func (s *syncBuffer) String() string {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.buf.String()
+}
+
+// TestStart_LogsDisabledWarningWhenNoTokenAndNoSubjects exercises the bug-7
+// regression: on a fresh daemon with no tracked sessions/PRs, discoverSubjects
+// returns empty and Poll short-circuits before reaching the credential gate.
+// The "scm observer disabled: provider credentials unavailable" warn line must
+// still fire exactly once from the observer loop's pre-Poll credential check.
+func TestStart_LogsDisabledWarningWhenNoTokenAndNoSubjects(t *testing.T) {
+	store := &fakeStore{
+		sessions: nil, // no sessions → discoverSubjects returns empty
+		projects: map[string]domain.ProjectRecord{},
+		prs:      map[domain.SessionID][]domain.PullRequest{},
+		checks:   map[string][]domain.PullRequestCheck{},
+	}
+	provider := &fakeProvider{
+		credentialGate: true,
+		credentialOK:   false,
+		repoGuards:     map[string]ports.SCMGuardResult{},
+		observations:   map[string]ports.SCMObservation{},
+	}
+
+	buf := &syncBuffer{}
+	logger := slog.New(slog.NewTextHandler(buf, &slog.HandlerOptions{Level: slog.LevelDebug}))
+	obs := New(provider, store, &fakeLifecycle{}, Config{
+		Clock:    func() time.Time { return time.Unix(1, 0).UTC() },
+		Tick:     time.Hour,
+		Logger:   logger,
+		CacheMax: 128,
+	})
+
+	ctx, cancel := context.WithCancel(context.Background())
+	done := obs.Start(ctx)
+	// Wait until the loop has emitted the expected warn line, or fail on timeout.
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		if strings.Contains(buf.String(), "scm observer disabled: provider credentials unavailable") {
+			break
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	cancel()
+	select {
+	case <-done:
+	case <-time.After(time.Second):
+		t.Fatal("observer did not exit after context cancellation")
+	}
+
+	logged := buf.String()
+	if !strings.Contains(logged, "scm observer disabled: provider credentials unavailable") {
+		t.Fatalf("expected disabled-credentials warn line in logs; got:\n%s", logged)
+	}
+	if got := strings.Count(logged, "scm observer disabled: provider credentials unavailable"); got != 1 {
+		t.Fatalf("warn line should fire exactly once, got %d occurrences:\n%s", got, logged)
+	}
+	if !obs.credentialsChecked || !obs.disabled {
+		t.Fatalf("observer state after pre-poll credential check: checked=%v disabled=%v", obs.credentialsChecked, obs.disabled)
+	}
+	if provider.credentialChecks != 1 {
+		t.Fatalf("credential checks = %d, want exactly one pre-poll check", provider.credentialChecks)
+	}
+	if provider.repoGuardCalls != 0 || provider.detectCalls != 0 || len(provider.fetchBatches) != 0 {
+		t.Fatalf("no provider API calls expected when disabled: guards=%d detects=%d batches=%d",
+			provider.repoGuardCalls, provider.detectCalls, len(provider.fetchBatches))
+	}
+}
+
 func TestPoll_RepoETag304SkipsDetectPR(t *testing.T) {
 	store := testStoreWithSession()
 	provider := &fakeProvider{repoGuards: map[string]ports.SCMGuardResult{prKey(testRepo, 0): {ETag: "v1", NotModified: true}}, observations: map[string]ports.SCMObservation{}}

backend/internal/ports/agent.go

@@ -2,10 +2,18 @@ package ports
 
 import (
 	"context"
+	"errors"
 
 	"github.com/aoagents/agent-orchestrator/backend/internal/domain"
 )
 
+// ErrAgentBinaryNotFound is returned by agent adapters when neither PATH nor
+// any well-known install location holds the agent's binary. The session
+// manager surfaces this BEFORE creating the runtime so a missing CLI doesn't
+// silently launch into an empty zellij pane that the reaper later mistakes
+// for a live session.
+var ErrAgentBinaryNotFound = errors.New("agent: binary not found on PATH")
+
 // Agent is the contract every CLI coding agent adapter (claude-code, codex, …)
 // must satisfy. It supplies the argv and process configuration the Session
 // Manager needs to launch, restore, and read back a native agent session.

backend/internal/ports/outbound.go

@@ -108,6 +108,18 @@ type Workspace interface {
 	Restore(ctx context.Context, cfg WorkspaceConfig) (WorkspaceInfo, error)
 }
 
+// Workspace-level sentinels surfaced through Create/Restore so the HTTP layer
+// can map them to typed errors rather than collapsing every adapter failure
+// into an opaque 500. Adapters wrap these via fmt.Errorf("...: %w", sentinel).
+var (
+	// ErrWorkspaceBranchCheckedOutElsewhere reports the requested branch is
+	// already checked out in another worktree of the same repo.
+	ErrWorkspaceBranchCheckedOutElsewhere = errors.New("workspace: branch is already checked out in another worktree")
+	// ErrWorkspaceBranchNotFetched reports the requested branch exists nowhere
+	// reachable (no local head, no remote-tracking branch, no tag).
+	ErrWorkspaceBranchNotFetched = errors.New("workspace: branch is not fetched")
+)
+
 // WorkspaceConfig is the spec for creating or restoring a session's workspace.
 type WorkspaceConfig struct {
 	ProjectID domain.ProjectID

backend/internal/service/session/service.go

@@ -41,6 +41,15 @@ type commander interface {
 	Kill(ctx context.Context, id domain.SessionID) (bool, error)
 	Send(ctx context.Context, id domain.SessionID, message string) error
 	Cleanup(ctx context.Context, project domain.ProjectID) ([]domain.SessionID, error)
+	RollbackSpawn(ctx context.Context, id domain.SessionID) (deleted, killed bool, err error)
+}
+
+// RollbackOutcome reports what happened in a rollback: either the seed row was
+// deleted, or the partially-spawned session was killed (runtime+workspace torn
+// down, row marked terminated).
+type RollbackOutcome struct {
+	Deleted bool `json:"deleted"`
+	Killed  bool `json:"killed"`
 }
 
 type scmProvider interface {
@@ -92,6 +101,9 @@ func NewWithDeps(d Deps) *Service {
 
 // Spawn creates a session and returns the API-facing read model.
 func (s *Service) Spawn(ctx context.Context, cfg ports.SpawnConfig) (domain.Session, error) {
+	if err := s.requireProject(ctx, cfg.ProjectID); err != nil {
+		return domain.Session{}, err
+	}
 	rec, err := s.manager.Spawn(ctx, cfg)
 	if err != nil {
 		return domain.Session{}, toAPIError(err)
@@ -99,11 +111,34 @@ func (s *Service) Spawn(ctx context.Context, cfg ports.SpawnConfig) (domain.Sess
 	return s.toSession(ctx, rec)
 }
 
+// requireProject verifies the project is registered before any spawn write
+// touches the session store, so an unknown projectId surfaces as a typed 404
+// rather than an opaque 500 with an orphan terminated row left behind.
+func (s *Service) requireProject(ctx context.Context, id domain.ProjectID) error {
+	if id == "" {
+		return apierr.Invalid("PROJECT_ID_REQUIRED", "projectId is required", nil)
+	}
+	if s.store == nil {
+		return nil
+	}
+	_, ok, err := s.store.GetProject(ctx, string(id))
+	if err != nil {
+		return fmt.Errorf("get project %s: %w", id, err)
+	}
+	if !ok {
+		return apierr.NotFound("PROJECT_NOT_FOUND", "Unknown project — register it with `ao project add`")
+	}
+	return nil
+}
+
 // SpawnOrchestrator spawns an orchestrator session for a project. When clean is
 // true it first tears down any active orchestrator(s) for that project so the new
 // one is the only live coordinator — a business rule that belongs here, not in the
 // HTTP controller.
 func (s *Service) SpawnOrchestrator(ctx context.Context, projectID domain.ProjectID, clean bool) (domain.Session, error) {
+	if err := s.requireProject(ctx, projectID); err != nil {
+		return domain.Session{}, err
+	}
 	if clean {
 		active := true
 		existing, err := s.List(ctx, ListFilter{ProjectID: projectID, Active: &active, OrchestratorOnly: true})
@@ -134,6 +169,18 @@ func (s *Service) Kill(ctx context.Context, id domain.SessionID) (bool, error) {
 	return freed, toAPIError(err)
 }
 
+// RollbackSpawn deletes a seed-state session row, or falls back to a Kill if
+// the session has spawn output. Used by the CLI to undo a `spawn --claim-pr`
+// when the claim step fails, avoiding the orphan terminated row that a plain
+// Kill would leave behind.
+func (s *Service) RollbackSpawn(ctx context.Context, id domain.SessionID) (RollbackOutcome, error) {
+	deleted, killed, err := s.manager.RollbackSpawn(ctx, id)
+	if err != nil {
+		return RollbackOutcome{}, toAPIError(err)
+	}
+	return RollbackOutcome{Deleted: deleted, Killed: killed}, nil
+}
+
 // Send delegates agent messaging to the internal manager.
 func (s *Service) Send(ctx context.Context, id domain.SessionID, message string) error {
 	return toAPIError(s.manager.Send(ctx, id, message))
@@ -237,6 +284,12 @@ func toAPIError(err error) error {
 		return apierr.Conflict("SESSION_INCOMPLETE_HANDLE", "Session is missing runtime or workspace handles", nil)
 	case errors.Is(err, sessionmanager.ErrProjectNotResolvable):
 		return apierr.Invalid("PROJECT_NOT_RESOLVABLE", "Project is not registered or has no repo — register it with `ao project add`", nil)
+	case errors.Is(err, ports.ErrWorkspaceBranchCheckedOutElsewhere):
+		return apierr.Conflict("BRANCH_CHECKED_OUT_ELSEWHERE", err.Error(), nil)
+	case errors.Is(err, ports.ErrWorkspaceBranchNotFetched):
+		return apierr.Invalid("BRANCH_NOT_FETCHED", err.Error(), nil)
+	case errors.Is(err, ports.ErrAgentBinaryNotFound):
+		return apierr.Invalid("AGENT_BINARY_NOT_FOUND", err.Error(), nil)
 	default:
 		return err
 	}

backend/internal/service/session/service_test.go

@@ -147,9 +147,13 @@ func (f *fakeCommander) Send(context.Context, domain.SessionID, string) error {
 func (f *fakeCommander) Cleanup(context.Context, domain.ProjectID) ([]domain.SessionID, error) {
 	return nil, nil
 }
+func (f *fakeCommander) RollbackSpawn(context.Context, domain.SessionID) (bool, bool, error) {
+	return false, false, nil
+}
 
 func TestSpawnOrchestratorCleanKillsActiveOrchestratorsBeforeSpawn(t *testing.T) {
 	st := newFakeStore()
+	st.projects["mer"] = domain.ProjectRecord{ID: "mer"}
 	// Two active orchestrators plus an unrelated worker and a terminated
 	// orchestrator that must be left alone.
 	st.sessions["mer-1"] = domain.SessionRecord{ID: "mer-1", ProjectID: "mer", Kind: domain.KindOrchestrator}
@@ -172,8 +176,70 @@ func TestSpawnOrchestratorCleanKillsActiveOrchestratorsBeforeSpawn(t *testing.T)
 	}
 }
 
+// TestSpawnUnknownProjectReturns404 covers Bug 1: an HTTP spawn for an
+// unregistered projectId must surface PROJECT_NOT_FOUND (apierr.NotFound)
+// BEFORE any session row is created, so no orphan terminated row is left
+// behind under `--include-terminated`.
+func TestSpawnUnknownProjectReturns404(t *testing.T) {
+	st := newFakeStore()
+	fc := &fakeCommander{}
+	svc := &Service{manager: fc, store: st}
+
+	_, err := svc.Spawn(context.Background(), ports.SpawnConfig{ProjectID: "ghost", Kind: domain.KindWorker})
+	var e *apierr.Error
+	if !errors.As(err, &e) || e.Kind != apierr.KindNotFound || e.Code != "PROJECT_NOT_FOUND" {
+		t.Fatalf("err = %v, want apierr.NotFound PROJECT_NOT_FOUND", err)
+	}
+	if fc.spawned {
+		t.Fatal("manager.Spawn must NOT be invoked for an unknown project")
+	}
+}
+
+// TestSpawnOrchestratorUnknownProjectReturns404 is the orchestrator-side guard
+// for Bug 1: same pre-validation, same typed envelope.
+func TestSpawnOrchestratorUnknownProjectReturns404(t *testing.T) {
+	st := newFakeStore()
+	fc := &fakeCommander{}
+	svc := &Service{manager: fc, store: st}
+
+	_, err := svc.SpawnOrchestrator(context.Background(), "ghost", false)
+	var e *apierr.Error
+	if !errors.As(err, &e) || e.Kind != apierr.KindNotFound || e.Code != "PROJECT_NOT_FOUND" {
+		t.Fatalf("err = %v, want apierr.NotFound PROJECT_NOT_FOUND", err)
+	}
+	if fc.spawned {
+		t.Fatal("manager.Spawn must NOT be invoked for an unknown project")
+	}
+}
+
+// TestToAPIErrorMapsWorkspaceBranchSentinels covers Bug 3: the workspace
+// adapter's typed branch errors map to typed envelope errors instead of
+// collapsing to a 500.
+func TestToAPIErrorMapsWorkspaceBranchSentinels(t *testing.T) {
+	cases := []struct {
+		name     string
+		err      error
+		wantKind apierr.Kind
+		wantCode string
+	}{
+		{"checked out elsewhere", fmt.Errorf("spawn mer-1: workspace: %w: \"x\" is checked out at \"/tmp\"", ports.ErrWorkspaceBranchCheckedOutElsewhere), apierr.KindConflict, "BRANCH_CHECKED_OUT_ELSEWHERE"},
+		{"not fetched", fmt.Errorf("spawn mer-1: workspace: %w: \"x\" has no local head", ports.ErrWorkspaceBranchNotFetched), apierr.KindInvalid, "BRANCH_NOT_FETCHED"},
+		{"agent binary not found", fmt.Errorf("spawn mer-1: %w", ports.ErrAgentBinaryNotFound), apierr.KindInvalid, "AGENT_BINARY_NOT_FOUND"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			mapped := toAPIError(tc.err)
+			var e *apierr.Error
+			if !errors.As(mapped, &e) || e.Kind != tc.wantKind || e.Code != tc.wantCode {
+				t.Fatalf("mapped = %v, want %s %s", mapped, tc.wantCode, e)
+			}
+		})
+	}
+}
+
 func TestSpawnOrchestratorNoCleanSkipsKills(t *testing.T) {
 	st := newFakeStore()
+	st.projects["mer"] = domain.ProjectRecord{ID: "mer"}
 	st.sessions["mer-1"] = domain.SessionRecord{ID: "mer-1", ProjectID: "mer", Kind: domain.KindOrchestrator}
 
 	fc := &fakeCommander{}

backend/internal/session_manager/manager.go

@@ -6,6 +6,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"os/exec"
 	"time"
 
 	"github.com/aoagents/agent-orchestrator/backend/internal/domain"
@@ -49,6 +50,12 @@ type Store interface {
 	GetSession(ctx context.Context, id domain.SessionID) (domain.SessionRecord, bool, error)
 	ListSessions(ctx context.Context, project domain.ProjectID) ([]domain.SessionRecord, error)
 	ListAllSessions(ctx context.Context) ([]domain.SessionRecord, error)
+	// DeleteSession removes a session row only if it is still in seed state
+	// (no workspace, runtime handle, agent session id, or prompt; not
+	// terminated). Returns deleted=true when removal happened; deleted=false
+	// when the row had already progressed past seed state — preserving the
+	// no-resurrection guarantee for live sessions.
+	DeleteSession(ctx context.Context, id domain.SessionID) (bool, error)
 }
 
 // Manager coordinates internal session spawn, restore, kill, and cleanup over
@@ -62,6 +69,10 @@ type Manager struct {
 	lcm       lifecycleRecorder
 	dataDir   string
 	clock     func() time.Time
+	// lookPath is exec.LookPath in production; tests substitute a stub so
+	// they don't need real binaries on PATH. Returns ports.ErrAgentBinaryNotFound
+	// when the binary is missing so the sentinel propagates through toAPIError.
+	lookPath func(string) (string, error)
 }
 
 // Deps are the collaborators a Session Manager needs; New wires them together.
@@ -76,6 +87,10 @@ type Deps struct {
 	// commands can open the same store.
 	DataDir string
 	Clock   func() time.Time
+	// LookPath overrides exec.LookPath for the pre-launch agent-binary check.
+	// Production wiring leaves this nil and the manager defaults to
+	// exec.LookPath; tests inject a stub so they need not seed real binaries.
+	LookPath func(string) (string, error)
 }
 
 // New builds a Session Manager from its dependencies, defaulting the clock to
@@ -90,10 +105,14 @@ func New(d Deps) *Manager {
 		lcm:       d.Lifecycle,
 		dataDir:   d.DataDir,
 		clock:     d.Clock,
+		lookPath:  d.LookPath,
 	}
 	if m.clock == nil {
 		m.clock = time.Now
 	}
+	if m.lookPath == nil {
+		m.lookPath = exec.LookPath
+	}
 	return m
 }
 
@@ -147,6 +166,15 @@ func (m *Manager) Spawn(ctx context.Context, cfg ports.SpawnConfig) (domain.Sess
 		m.markSpawnFailedTerminated(ctx, id)
 		return domain.SessionRecord{}, fmt.Errorf("spawn %s: launch command: %w", id, err)
 	}
+	// Pre-flight: confirm argv[0] actually exists on PATH (or as an absolute
+	// path the adapter returned) BEFORE handing the launch to the runtime.
+	// Zellij happily creates a session+pane around a missing command, so an
+	// unresolved binary would leak through as a "live" session that never ran.
+	if err := m.validateAgentBinary(argv); err != nil {
+		_ = m.workspace.Destroy(ctx, ws)
+		m.markSpawnFailedTerminated(ctx, id)
+		return domain.SessionRecord{}, fmt.Errorf("spawn %s: %w", id, err)
+	}
 	handle, err := m.runtime.Create(ctx, ports.RuntimeConfig{
 		SessionID:     id,
 		WorkspacePath: ws.Path,
@@ -170,11 +198,43 @@ func (m *Manager) Spawn(ctx context.Context, cfg ports.SpawnConfig) (domain.Sess
 }
 
 // markSpawnFailedTerminated best-effort parks an orphaned spawn as terminated.
-// The store has no delete; a phantom half-spawned row is worse than a terminal one.
+// A phantom half-spawned row is worse than a terminal one; we only delete the
+// row when nothing observable has landed yet (seed state) via rollbackSpawn.
 func (m *Manager) markSpawnFailedTerminated(ctx context.Context, id domain.SessionID) {
 	_ = m.lcm.MarkTerminated(ctx, id)
 }
 
+// rollbackSpawn deletes a session row when it is still in seed state — used
+// when an out-of-band step that happens AFTER `Spawn` returns (e.g. PR claim
+// over HTTP) has failed and the caller wants the partially-spawned session
+// gone without leaving a terminated orphan visible under `--include-terminated`.
+//
+// If the row has progressed past seed state (workspace exists, runtime created,
+// etc.), DeleteSession is a no-op and rollbackSpawn falls back to a Kill so the
+// runtime/workspace are torn down. Returns (deleted, killed):
+//   - deleted=true: the row was a seed row and has been removed
+//   - killed=true:  the row had spawn output and was torn down + terminated
+//   - both false:   the row was already terminated or absent — benign no-op
+func (m *Manager) rollbackSpawn(ctx context.Context, id domain.SessionID) (deleted, killed bool, err error) {
+	deleted, err = m.store.DeleteSession(ctx, id)
+	if err != nil {
+		return false, false, fmt.Errorf("rollback %s: %w", id, err)
+	}
+	if deleted {
+		return true, false, nil
+	}
+	killed, err = m.Kill(ctx, id)
+	if err != nil {
+		return false, false, err
+	}
+	return false, killed, nil
+}
+
+// RollbackSpawn is the public surface of rollbackSpawn for service-layer callers.
+func (m *Manager) RollbackSpawn(ctx context.Context, id domain.SessionID) (deleted, killed bool, err error) {
+	return m.rollbackSpawn(ctx, id)
+}
+
 // Kill records terminal intent with the LCM, then tears down the runtime and
 // workspace. A workspace teardown refused by the worktree-remove safety
 // (uncommitted work) surfaces as an error with freed=false and is never forced.
@@ -218,6 +278,13 @@ func (m *Manager) Restore(ctx context.Context, id domain.SessionID) (domain.Sess
 		return domain.SessionRecord{}, fmt.Errorf("restore %s: %w", id, ErrNotRestorable)
 	}
 	meta := rec.Metadata
+	// Mirror Kill's incomplete-handle guard: a session whose spawn failed before
+	// the workspace landed has neither WorkspacePath nor Branch, and there is
+	// nothing meaningful to restore from. Surface this as a typed 409 instead of
+	// letting workspace.Restore fail with an opaque wrapped error.
+	if meta.WorkspacePath == "" || meta.Branch == "" {
+		return domain.SessionRecord{}, fmt.Errorf("restore %s: %w", id, ErrIncompleteHandle)
+	}
 	if meta.AgentSessionID == "" && meta.Prompt == "" {
 		return domain.SessionRecord{}, fmt.Errorf("restore %s: nothing to resume from", id)
 	}
@@ -462,6 +529,22 @@ func restoreArgv(ctx context.Context, agent ports.Agent, id domain.SessionID, wo
 	return argv, nil
 }
 
+// validateAgentBinary checks that argv[0] resolves via the manager's
+// lookPath (exec.LookPath in prod) before any runtime work happens. Adapters
+// that can't resolve their binary now return ports.ErrAgentBinaryNotFound from
+// GetLaunchCommand directly; this guard is a defense-in-depth for adapters
+// that return an argv[0] like "claude" without verifying.
+func (m *Manager) validateAgentBinary(argv []string) error {
+	if len(argv) == 0 {
+		return fmt.Errorf("agent: empty launch argv: %w", ports.ErrAgentBinaryNotFound)
+	}
+	bin := argv[0]
+	if _, err := m.lookPath(bin); err != nil {
+		return fmt.Errorf("agent binary %q: %w", bin, ports.ErrAgentBinaryNotFound)
+	}
+	return nil
+}
+
 func runtimeHandle(meta domain.SessionMetadata) ports.RuntimeHandle {
 	return ports.RuntimeHandle{ID: meta.RuntimeHandleID}
 }

backend/internal/session_manager/manager_test.go

@@ -53,6 +53,18 @@ func (f *fakeStore) ListAllSessions(context.Context) ([]domain.SessionRecord, er
 	}
 	return out, nil
 }
+func (f *fakeStore) DeleteSession(_ context.Context, id domain.SessionID) (bool, error) {
+	rec, ok := f.sessions[id]
+	if !ok {
+		return false, nil
+	}
+	// Mirror the sqlite gate: only delete rows still in seed state.
+	if rec.IsTerminated || rec.Metadata.WorkspacePath != "" || rec.Metadata.RuntimeHandleID != "" || rec.Metadata.AgentSessionID != "" || rec.Metadata.Prompt != "" {
+		return false, nil
+	}
+	delete(f.sessions, id)
+	return true, nil
+}
 func (f *fakeStore) GetDisplayPRFactsForSession(_ context.Context, id domain.SessionID) (domain.PRFacts, bool, error) {
 	if pr := f.pr[id]; pr.URL != "" {
 		return pr, true, nil
@@ -150,7 +162,10 @@ func newManager() (*Manager, *fakeStore, *fakeRuntime, *fakeWorkspace) {
 	st := newFakeStore()
 	rt := &fakeRuntime{}
 	ws := &fakeWorkspace{}
-	m := New(Deps{Runtime: rt, Agents: fakeAgents{}, Workspace: ws, Store: st, Messenger: &fakeMessenger{}, Lifecycle: &fakeLCM{store: st}})
+	// Stub lookPath so the pre-launch agent-binary check passes; the fakeAgent
+	// returns argv ["launch"] which is not a real binary on PATH.
+	lookPath := func(string) (string, error) { return "/bin/true", nil }
+	m := New(Deps{Runtime: rt, Agents: fakeAgents{}, Workspace: ws, Store: st, Messenger: &fakeMessenger{}, Lifecycle: &fakeLCM{store: st}, LookPath: lookPath})
 	return m, st, rt, ws
 }
 func seedTerminal(st *fakeStore, id domain.SessionID, meta domain.SessionMetadata) {
@@ -315,6 +330,99 @@ func TestSpawnOrchestrator_UsesCoordinatorPrompt(t *testing.T) {
 	}
 }
 
+// TestRestore_RefusesIncompleteHandle covers Bug 2: a terminated row whose
+// spawn failed before the workspace landed (no WorkspacePath, no Branch) must
+// fail Restore with ErrIncompleteHandle — the same typed sentinel Kill returns
+// for the same shape — so the HTTP layer surfaces a typed 409 instead of an
+// opaque 500.
+func TestRestore_RefusesIncompleteHandle(t *testing.T) {
+	m, st, _, _ := newManager()
+	// Seed a terminated row with no workspace and no branch (the post-failure
+	// shape of a Spawn that died before workspace.Create succeeded).
+	st.sessions["mer-1"] = domain.SessionRecord{
+		ID:           "mer-1",
+		ProjectID:    "mer",
+		IsTerminated: true,
+		Metadata:     domain.SessionMetadata{Prompt: "do it"},
+	}
+	if _, err := m.Restore(ctx, "mer-1"); !errors.Is(err, ErrIncompleteHandle) {
+		t.Fatalf("want ErrIncompleteHandle, got %v", err)
+	}
+}
+
+// TestRollbackSpawn_DeletesSeedRow covers Bug 4: a session row in seed state
+// (no workspace, no runtime, no agent session id, not terminated) is deleted
+// outright by RollbackSpawn so the user never sees an orphan terminated row.
+func TestRollbackSpawn_DeletesSeedRow(t *testing.T) {
+	m, st, _, _ := newManager()
+	// Seed row matches what CreateSession produces — no Metadata at all.
+	st.sessions["mer-1"] = domain.SessionRecord{
+		ID:        "mer-1",
+		ProjectID: "mer",
+		Activity:  domain.Activity{State: domain.ActivityIdle},
+	}
+	deleted, killed, err := m.RollbackSpawn(ctx, "mer-1")
+	if err != nil {
+		t.Fatalf("rollback err = %v", err)
+	}
+	if !deleted || killed {
+		t.Fatalf("deleted=%v killed=%v, want deleted=true killed=false", deleted, killed)
+	}
+	if _, present := st.sessions["mer-1"]; present {
+		t.Fatal("seed row must be removed from the store, not left as terminated")
+	}
+}
+
+// TestRollbackSpawn_FallsBackToKillForLiveRow asserts the no-resurrection
+// guarantee from Bug 4's RCA: once a row has observable spawn output (workspace
+// + runtime handle), DeleteSession is a no-op and rollback falls back to Kill
+// so the runtime + workspace are torn down rather than abandoned.
+func TestRollbackSpawn_FallsBackToKillForLiveRow(t *testing.T) {
+	m, st, rt, ws := newManager()
+	st.sessions["mer-1"] = mkLive("mer-1")
+	deleted, killed, err := m.RollbackSpawn(ctx, "mer-1")
+	if err != nil {
+		t.Fatalf("rollback err = %v", err)
+	}
+	if deleted || !killed {
+		t.Fatalf("deleted=%v killed=%v, want deleted=false killed=true", deleted, killed)
+	}
+	if rt.destroyed != 1 || ws.destroyed != 1 {
+		t.Fatalf("kill teardown not invoked: rt=%d ws=%d", rt.destroyed, ws.destroyed)
+	}
+	if !st.sessions["mer-1"].IsTerminated {
+		t.Fatal("live row should be marked terminated after kill-fallback")
+	}
+}
+
+// TestSpawn_RejectsMissingAgentBinary covers Bug 6: when the agent adapter
+// returns an argv whose binary is not on PATH, Manager.Spawn must abort BEFORE
+// runtime.Create rather than launching into an empty zellij pane that the
+// reaper later mistakes for a live session.
+func TestSpawn_RejectsMissingAgentBinary(t *testing.T) {
+	st := newFakeStore()
+	rt := &fakeRuntime{}
+	ws := &fakeWorkspace{}
+	notFound := func(name string) (string, error) {
+		return "", fmt.Errorf("exec: %q: not found", name)
+	}
+	m := New(Deps{Runtime: rt, Agents: fakeAgents{}, Workspace: ws, Store: st, Messenger: &fakeMessenger{}, Lifecycle: &fakeLCM{store: st}, LookPath: notFound})
+
+	_, err := m.Spawn(ctx, ports.SpawnConfig{ProjectID: "mer", Kind: domain.KindWorker})
+	if !errors.Is(err, ports.ErrAgentBinaryNotFound) {
+		t.Fatalf("err = %v, want ports.ErrAgentBinaryNotFound", err)
+	}
+	if rt.created != 0 {
+		t.Fatal("runtime.Create must NOT run when the agent binary is missing")
+	}
+	if ws.destroyed != 1 {
+		t.Fatal("workspace must be torn down when the pre-launch binary check fails")
+	}
+	if !st.sessions["mer-1"].IsTerminated {
+		t.Fatal("the orphan row should be marked terminated after the failed spawn")
+	}
+}
+
 func TestSpawn_KeepsExplicitBranch(t *testing.T) {
 	m, st, _, _ := newManager()
 	s, err := m.Spawn(ctx, ports.SpawnConfig{ProjectID: "mer", Kind: domain.KindWorker, Branch: "feature/x"})

backend/internal/storage/sqlite/gen/models.go

@@ -60,6 +60,7 @@ type PR struct {
 	ObservedAt               sql.NullTime
 	CIObservedAt             sql.NullTime
 	ReviewObservedAt         sql.NullTime
+	LastNudgeSignature       string
 }
 
 type PRCheck struct {

backend/internal/storage/sqlite/gen/pr.sql.go

@@ -99,16 +99,7 @@ func (q *Queries) GetDisplayPRFactsBySession(ctx context.Context, sessionID doma
 }
 
 const getPR = `-- name: GetPR :one
-SELECT
+SELECT url, session_id, number, pr_state, review_decision, ci_state, mergeability, updated_at, provider, host, repo, source_branch, target_branch, head_sha, title, additions, deletions, changed_files, author, base_sha, merge_commit_sha, is_draft, is_merged, is_closed, provider_state, provider_mergeable, provider_merge_state_status, html_url, created_at_provider, updated_at_provider, merged_at_provider, closed_at_provider, metadata_hash, ci_hash, review_hash, observed_at, ci_observed_at, review_observed_at, last_nudge_signature FROM pr WHERE url = ?
-    url, session_id, number, pr_state, review_decision, ci_state, mergeability, updated_at,
-    provider, host, repo, source_branch, target_branch, head_sha, title,
-    additions, deletions, changed_files, author, base_sha, merge_commit_sha,
-    is_draft, is_merged, is_closed,
-    provider_state, provider_mergeable, provider_merge_state_status, html_url,
-    created_at_provider, updated_at_provider, merged_at_provider, closed_at_provider,
-    metadata_hash, ci_hash, review_hash, observed_at, ci_observed_at, review_observed_at
-FROM pr
-WHERE url = ?
 `
 
 func (q *Queries) GetPR(ctx context.Context, url string) (PR, error) {
@@ -153,6 +144,7 @@ func (q *Queries) GetPR(ctx context.Context, url string) (PR, error) {
 		&i.ObservedAt,
 		&i.CIObservedAt,
 		&i.ReviewObservedAt,
+		&i.LastNudgeSignature,
 	)
 	return i, err
 }
@@ -178,16 +170,19 @@ func (q *Queries) GetPRClaimAndOwner(ctx context.Context, url string) (GetPRClai
 	return i, err
 }
 
+const getPRLastNudgeSignature = `-- name: GetPRLastNudgeSignature :one
+SELECT last_nudge_signature FROM pr WHERE url = ?
+`
+
+func (q *Queries) GetPRLastNudgeSignature(ctx context.Context, url string) (string, error) {
+	row := q.db.QueryRowContext(ctx, getPRLastNudgeSignature, url)
+	var last_nudge_signature string
+	err := row.Scan(&last_nudge_signature)
+	return last_nudge_signature, err
+}
+
 const listPRsBySession = `-- name: ListPRsBySession :many
-SELECT
+SELECT url, session_id, number, pr_state, review_decision, ci_state, mergeability, updated_at, provider, host, repo, source_branch, target_branch, head_sha, title, additions, deletions, changed_files, author, base_sha, merge_commit_sha, is_draft, is_merged, is_closed, provider_state, provider_mergeable, provider_merge_state_status, html_url, created_at_provider, updated_at_provider, merged_at_provider, closed_at_provider, metadata_hash, ci_hash, review_hash, observed_at, ci_observed_at, review_observed_at, last_nudge_signature FROM pr
-    url, session_id, number, pr_state, review_decision, ci_state, mergeability, updated_at,
-    provider, host, repo, source_branch, target_branch, head_sha, title,
-    additions, deletions, changed_files, author, base_sha, merge_commit_sha,
-    is_draft, is_merged, is_closed,
-    provider_state, provider_mergeable, provider_merge_state_status, html_url,
-    created_at_provider, updated_at_provider, merged_at_provider, closed_at_provider,
-    metadata_hash, ci_hash, review_hash, observed_at, ci_observed_at, review_observed_at
-FROM pr
 WHERE session_id = ?
 ORDER BY updated_at DESC
 `
@@ -240,6 +235,7 @@ func (q *Queries) ListPRsBySession(ctx context.Context, sessionID domain.Session
 			&i.ObservedAt,
 			&i.CIObservedAt,
 			&i.ReviewObservedAt,
+			&i.LastNudgeSignature,
 		); err != nil {
 			return nil, err
 		}
@@ -254,6 +250,20 @@ func (q *Queries) ListPRsBySession(ctx context.Context, sessionID domain.Session
 	return items, nil
 }
 
+const updatePRLastNudgeSignature = `-- name: UpdatePRLastNudgeSignature :exec
+UPDATE pr SET last_nudge_signature = ? WHERE url = ?
+`
+
+type UpdatePRLastNudgeSignatureParams struct {
+	LastNudgeSignature string
+	URL                string
+}
+
+func (q *Queries) UpdatePRLastNudgeSignature(ctx context.Context, arg UpdatePRLastNudgeSignatureParams) error {
+	_, err := q.db.ExecContext(ctx, updatePRLastNudgeSignature, arg.LastNudgeSignature, arg.URL)
+	return err
+}
+
 const upsertLegacyPR = `-- name: UpsertLegacyPR :exec
 INSERT INTO pr (
     url, session_id, number, pr_state, review_decision, ci_state, mergeability, updated_at,
@@ -437,28 +447,3 @@ func (q *Queries) UpsertPR(ctx context.Context, arg UpsertPRParams) error {
 	)
 	return err
 }
-
-const getPRLastNudgeSignature = `-- name: GetPRLastNudgeSignature :one
-SELECT last_nudge_signature FROM pr WHERE url = ?
-`
-
-func (q *Queries) GetPRLastNudgeSignature(ctx context.Context, url string) (string, error) {
-	row := q.db.QueryRowContext(ctx, getPRLastNudgeSignature, url)
-	var sig string
-	err := row.Scan(&sig)
-	return sig, err
-}
-
-const updatePRLastNudgeSignature = `-- name: UpdatePRLastNudgeSignature :exec
-UPDATE pr SET last_nudge_signature = ? WHERE url = ?
-`
-
-type UpdatePRLastNudgeSignatureParams struct {
-	LastNudgeSignature string
-	URL                string
-}
-
-func (q *Queries) UpdatePRLastNudgeSignature(ctx context.Context, arg UpdatePRLastNudgeSignatureParams) error {
-	_, err := q.db.ExecContext(ctx, updatePRLastNudgeSignature, arg.LastNudgeSignature, arg.URL)
-	return err
-}

backend/internal/storage/sqlite/gen/pr_review_threads.sql.go

@@ -10,6 +10,20 @@ import (
 	"time"
 )
 
+const deletePRReviewThread = `-- name: DeletePRReviewThread :exec
+DELETE FROM pr_review_threads WHERE pr_url = ? AND thread_id = ?
+`
+
+type DeletePRReviewThreadParams struct {
+	PRURL    string
+	ThreadID string
+}
+
+func (q *Queries) DeletePRReviewThread(ctx context.Context, arg DeletePRReviewThreadParams) error {
+	_, err := q.db.ExecContext(ctx, deletePRReviewThread, arg.PRURL, arg.ThreadID)
+	return err
+}
+
 const deletePRReviewThreads = `-- name: DeletePRReviewThreads :exec
 DELETE FROM pr_review_threads WHERE pr_url = ?
 `

backend/internal/storage/sqlite/gen/sessions.sql.go

@@ -221,6 +221,30 @@ func (q *Queries) RenameSession(ctx context.Context, arg RenameSessionParams) (i
 	return result.RowsAffected()
 }
 
+const sessionIsSeed = `-- name: SessionIsSeed :one
+SELECT EXISTS(
+    SELECT 1 FROM sessions
+    WHERE id = ?
+      AND is_terminated = 0
+      AND workspace_path = ''
+      AND runtime_handle_id = ''
+      AND agent_session_id = ''
+      AND prompt = ''
+) AS is_seed
+`
+
+// SessionIsSeed reports whether the session id matches a row still in seed
+// state (see DeleteSeedSession for the conditions). Callers probe with this
+// before touching change_log so that DeleteSession is a true no-op for live
+// sessions instead of silently destroying their CDC events. Returns 0 when
+// the row does not exist OR has progressed past seed state.
+func (q *Queries) SessionIsSeed(ctx context.Context, id domain.SessionID) (bool, error) {
+	row := q.db.QueryRowContext(ctx, sessionIsSeed, id)
+	var is_seed bool
+	err := row.Scan(&is_seed)
+	return is_seed, err
+}
+
 const updateSession = `-- name: UpdateSession :exec
 UPDATE sessions SET
     issue_id = ?, kind = ?, harness = ?, display_name = ?,

backend/internal/storage/sqlite/queries/changelog.sql

@@ -5,3 +5,13 @@ FROM change_log WHERE seq > ? ORDER BY seq LIMIT ?;
 
 -- name: MaxChangeLogSeq :one
 SELECT CAST(COALESCE(MAX(seq), 0) AS INTEGER) AS seq FROM change_log;
+
+-- NOTE: `DELETE FROM change_log WHERE session_id = ?` is intentionally NOT
+-- a sqlc query. sqlc 1.31's SQLite parser strips the `?` placeholder and
+-- emits a *domain.SessionID pointer parameter whenever a nullable column
+-- (change_log.session_id is nullable for project-level events) sits on the
+-- LHS of `=` in a top-level DELETE. None of the obvious SQL workarounds
+-- (sqlc.arg, IFNULL, rowid subquery, second predicate) defeated the
+-- heuristic. The store runs that DELETE directly via tx.ExecContext inside
+-- Store.DeleteSession to keep it part of the same transaction as the seed
+-- probe + session delete.

backend/internal/storage/sqlite/queries/pr.sql

@@ -65,27 +65,10 @@ ON CONFLICT (url) DO UPDATE SET
     is_closed = excluded.is_closed;
 
 -- name: GetPR :one
-SELECT
+SELECT * FROM pr WHERE url = ?;
-    url, session_id, number, pr_state, review_decision, ci_state, mergeability, updated_at,
-    provider, host, repo, source_branch, target_branch, head_sha, title,
-    additions, deletions, changed_files, author, base_sha, merge_commit_sha,
-    is_draft, is_merged, is_closed,
-    provider_state, provider_mergeable, provider_merge_state_status, html_url,
-    created_at_provider, updated_at_provider, merged_at_provider, closed_at_provider,
-    metadata_hash, ci_hash, review_hash, observed_at, ci_observed_at, review_observed_at
-FROM pr
-WHERE url = ?;
 
 -- name: ListPRsBySession :many
-SELECT
+SELECT * FROM pr
-    url, session_id, number, pr_state, review_decision, ci_state, mergeability, updated_at,
-    provider, host, repo, source_branch, target_branch, head_sha, title,
-    additions, deletions, changed_files, author, base_sha, merge_commit_sha,
-    is_draft, is_merged, is_closed,
-    provider_state, provider_mergeable, provider_merge_state_status, html_url,
-    created_at_provider, updated_at_provider, merged_at_provider, closed_at_provider,
-    metadata_hash, ci_hash, review_hash, observed_at, ci_observed_at, review_observed_at
-FROM pr
 WHERE session_id = ?
 ORDER BY updated_at DESC;
 

backend/internal/storage/sqlite/queries/pr_review_threads.sql

@@ -13,6 +13,9 @@ ON CONFLICT (pr_url, thread_id) DO UPDATE SET
 -- name: DeletePRReviewThreads :exec
 DELETE FROM pr_review_threads WHERE pr_url = ?;
 
+-- name: DeletePRReviewThread :exec
+DELETE FROM pr_review_threads WHERE pr_url = ? AND thread_id = ?;
+
 -- name: ListPRReviewThreads :many
 SELECT pr_url, thread_id, path, line, resolved, is_bot, semantic_hash, updated_at
 FROM pr_review_threads WHERE pr_url = ? ORDER BY updated_at, thread_id;

backend/internal/storage/sqlite/queries/sessions.sql

@@ -38,3 +38,28 @@ FROM sessions ORDER BY project_id, num;
 
 -- name: RenameSession :execrows
 UPDATE sessions SET display_name = ?, updated_at = ? WHERE id = ?;
+
+-- name: SessionIsSeed :one
+-- SessionIsSeed reports whether the session id matches a row still in seed
+-- state (see DeleteSeedSession for the conditions). Callers probe with this
+-- before touching change_log so that DeleteSession is a true no-op for live
+-- sessions instead of silently destroying their CDC events. Returns 0 when
+-- the row does not exist OR has progressed past seed state.
+SELECT EXISTS(
+    SELECT 1 FROM sessions
+    WHERE id = ?
+      AND is_terminated = 0
+      AND workspace_path = ''
+      AND runtime_handle_id = ''
+      AND agent_session_id = ''
+      AND prompt = ''
+) AS is_seed;
+
+-- NOTE: the `DELETE FROM sessions WHERE id = ? AND <seed-state predicates>`
+-- statement is intentionally NOT a sqlc query — same sqlc 1.31 SQLite-parser
+-- bug as documented in queries/changelog.sql: trailing string literals (and
+-- placeholders) on the RHS of `=` in a DELETE get silently stripped, so the
+-- generated SQL ends up mid-clause and the row count is meaningless. The
+-- store runs that DELETE directly via tx.ExecContext inside
+-- Store.DeleteSession, inside the same transaction as the SessionIsSeed
+-- probe and the raw change_log cleanup.

backend/internal/storage/sqlite/store/pr_cdc_test.go

@@ -116,6 +116,130 @@ func TestPRReviewThreadsCDC_EmitsOnInsertAndResolvedTransition(t *testing.T) {
 	}
 }
 
+// Regression for the bug where pr_review_thread_resolved never fired on the
+// common Replace path. Real-world polls take ReviewWriteReplace whenever the
+// upstream listing is not paginated (provider observer sets Partial=false).
+// The previous implementation did DELETE-all + UPSERT inside the tx, so every
+// upsert hit the INSERT trigger and the UPDATE trigger that emits
+// pr_review_thread_resolved never saw the resolved flip. The fix is a set-diff
+// delete: upsert observed threads first (so unchanged thread_ids hit ON
+// CONFLICT DO UPDATE and the UPDATE trigger fires), then prune the rows whose
+// thread_id is no longer in the observed set.
+func TestPRReviewThreadsCDC_EmitsResolvedOnReplacePoll(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+	seedProject(t, s, "mer")
+	rec, err := s.CreateSession(ctx, sampleRecord("mer"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	now := time.Now().UTC().Truncate(time.Second)
+	pr := domain.PullRequest{URL: "https://example/pr/55", SessionID: rec.ID, Number: 55, UpdatedAt: now}
+
+	// First poll: seed via Replace (no Partial pagination). Same shape as the
+	// real GitHub provider path when the review-thread listing fits in one page.
+	if err := s.WriteSCMObservation(ctx, pr, nil, []domain.PullRequestReviewThread{{
+		ThreadID: "t1", Path: "main.go", Line: 7, IsBot: true, SemanticHash: "v1", UpdatedAt: now,
+	}}, nil, ports.ReviewWriteReplace); err != nil {
+		t.Fatal(err)
+	}
+
+	// Second poll: same thread, resolved flipped to true, also via Replace.
+	// On the buggy code this fired the INSERT trigger again (because the
+	// DELETE-all removed the row first) and the UPDATE trigger never saw the
+	// resolved transition.
+	if err := s.WriteSCMObservation(ctx, pr, nil, []domain.PullRequestReviewThread{{
+		ThreadID: "t1", Path: "main.go", Line: 7, Resolved: true, IsBot: true, SemanticHash: "v2", UpdatedAt: now.Add(time.Second),
+	}}, nil, ports.ReviewWriteReplace); err != nil {
+		t.Fatal(err)
+	}
+
+	rows, err := s.EventsAfter(ctx, 0, 100)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var added, resolved []cdc.Event
+	for _, r := range rows {
+		switch r.Type {
+		case cdc.EventPRReviewThreadAdded:
+			added = append(added, r)
+		case cdc.EventPRReviewThreadResolved:
+			resolved = append(resolved, r)
+		}
+	}
+	if len(added) != 1 {
+		t.Fatalf("want 1 review-thread added CDC event (initial Replace insert), got %d", len(added))
+	}
+	if len(resolved) != 1 {
+		t.Fatalf("want 1 review-thread resolved CDC event on the second Replace poll, got %d", len(resolved))
+	}
+	var payload map[string]any
+	if err := json.Unmarshal(resolved[0].Payload, &payload); err != nil {
+		t.Fatalf("resolved payload JSON: %v", err)
+	}
+	if payload["thread"] != "t1" || payload["resolved"] != true {
+		t.Fatalf("resolved payload = %#v", payload)
+	}
+}
+
+// Pruning regression: Replace must still drop threads that are no longer in
+// the observed listing, otherwise stale rows accumulate. Seed two threads,
+// then re-poll with only one; the missing thread must be gone, while the
+// surviving thread still gets an UPDATE (not a fresh INSERT).
+func TestPRReviewThreadsReplace_PrunesOrphansWithoutReinserting(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+	seedProject(t, s, "mer")
+	rec, err := s.CreateSession(ctx, sampleRecord("mer"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	now := time.Now().UTC().Truncate(time.Second)
+	pr := domain.PullRequest{URL: "https://example/pr/56", SessionID: rec.ID, Number: 56, UpdatedAt: now}
+
+	if err := s.WriteSCMObservation(ctx, pr, nil, []domain.PullRequestReviewThread{
+		{ThreadID: "keep", Path: "a.go", Line: 1, IsBot: true, SemanticHash: "k1", UpdatedAt: now},
+		{ThreadID: "drop", Path: "b.go", Line: 2, IsBot: true, SemanticHash: "d1", UpdatedAt: now},
+	}, nil, ports.ReviewWriteReplace); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.WriteSCMObservation(ctx, pr, nil, []domain.PullRequestReviewThread{
+		{ThreadID: "keep", Path: "a.go", Line: 1, Resolved: true, IsBot: true, SemanticHash: "k2", UpdatedAt: now.Add(time.Second)},
+	}, nil, ports.ReviewWriteReplace); err != nil {
+		t.Fatal(err)
+	}
+
+	got, err := s.ListPRReviewThreads(ctx, pr.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got) != 1 || got[0].ThreadID != "keep" || !got[0].Resolved {
+		t.Fatalf("after prune want one resolved \"keep\" row, got %+v", got)
+	}
+	rows, err := s.EventsAfter(ctx, 0, 100)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var added, resolved int
+	for _, r := range rows {
+		if r.Type == cdc.EventPRReviewThreadAdded {
+			added++
+		}
+		if r.Type == cdc.EventPRReviewThreadResolved {
+			resolved++
+		}
+	}
+	// Two adds from poll 1 ("keep" and "drop"), no extra add on poll 2
+	// (the surviving row went through ON CONFLICT DO UPDATE, not a fresh
+	// INSERT), and one resolved transition from the kept row's flip.
+	if added != 2 {
+		t.Fatalf("want 2 added events across both polls, got %d", added)
+	}
+	if resolved != 1 {
+		t.Fatalf("want 1 resolved event from the surviving thread's flip, got %d", resolved)
+	}
+}
+
 // WritePR persists scalar facts, checks, and comments in one tx; all three
 // should be queryable afterward.
 func TestWritePR_PersistsScalarsChecksAndComments(t *testing.T) {

backend/internal/storage/sqlite/store/pr_store.go

@@ -110,11 +110,6 @@ func writePRRows(ctx context.Context, q *gen.Queries, pr domain.PullRequest, che
 			return err
 		}
 	}
-	if reviewMode == ports.ReviewWriteReplace {
-		if err := q.DeletePRReviewThreads(ctx, pr.URL); err != nil {
-			return err
-		}
-	}
 	if reviewMode == ports.ReviewWriteReplace {
 		if err := q.DeletePRComments(ctx, pr.URL); err != nil {
 			return err
@@ -131,6 +126,30 @@ func writePRRows(ctx context.Context, q *gen.Queries, pr domain.PullRequest, che
 			}
 		}
 	}
+	// Replace mode prunes orphans (threads no longer observed in the upstream
+	// listing) AFTER the upserts above, so that threads present in both the
+	// pre- and post-state hit ON CONFLICT DO UPDATE and fire the UPDATE trigger
+	// (e.g. pr_review_thread_resolved when resolved flips). The old
+	// delete-everything-first approach made every poll look like a fresh INSERT
+	// and the UPDATE trigger was unreachable for the common Replace path.
+	if reviewMode == ports.ReviewWriteReplace {
+		observed := make(map[string]struct{}, len(threads))
+		for _, th := range threads {
+			observed[th.ThreadID] = struct{}{}
+		}
+		existing, err := q.ListPRReviewThreads(ctx, pr.URL)
+		if err != nil {
+			return fmt.Errorf("list review threads for prune %s: %w", pr.URL, err)
+		}
+		for _, row := range existing {
+			if _, ok := observed[row.ThreadID]; ok {
+				continue
+			}
+			if err := q.DeletePRReviewThread(ctx, gen.DeletePRReviewThreadParams{PRURL: pr.URL, ThreadID: row.ThreadID}); err != nil {
+				return fmt.Errorf("prune review thread %q: %w", row.ThreadID, err)
+			}
+		}
+	}
 	if reviewMode == ports.ReviewWriteMerge {
 		for _, threadID := range reviewThreadIDs(threads, comments) {
 			if err := q.DeletePRCommentsByThread(ctx, gen.DeletePRCommentsByThreadParams{PRURL: pr.URL, ThreadID: threadID}); err != nil {

backend/internal/storage/sqlite/store/session_store.go

@@ -56,6 +56,80 @@ func (s *Store) RenameSession(ctx context.Context, id domain.SessionID, displayN
 	return rows > 0, nil
 }
 
+// DeleteSession removes a session row, but only if it is still in seed state
+// (no workspace, no runtime handle, no agent session id, no prompt, and not
+// already terminated). Rows that have observable spawn output are immutable
+// to preserve the no-resurrection guarantee — for those, callers fall back to
+// MarkTerminated (lifecycle.Manager) instead.
+//
+// The deletion runs in a transaction. It first probes seed state with
+// SessionIsSeed; only if that returns true does it clear the session's
+// change_log rows (required because change_log FKs sessions(id) without
+// ON DELETE CASCADE) and then delete the session row. For live or absent
+// sessions the transaction commits with no rows touched — critically, the
+// session_created / session_updated CDC events for live sessions are NOT
+// destroyed when callers (e.g. RollbackSpawn's delete-then-kill fallback)
+// invoke DeleteSession on a fully-spawned row.
+//
+// Returns deleted=true when a seed row was removed; deleted=false when the
+// session id did not match a seed row (either it never existed, or it had
+// already progressed past seed state). The latter case is benign — the caller
+// should fall back to MarkTerminated.
+func (s *Store) DeleteSession(ctx context.Context, id domain.SessionID) (bool, error) {
+	s.writeMu.Lock()
+	defer s.writeMu.Unlock()
+	tx, err := s.writeDB.BeginTx(ctx, nil)
+	if err != nil {
+		return false, fmt.Errorf("begin delete seed session: %w", err)
+	}
+	defer func() { _ = tx.Rollback() }()
+	q := s.qw.WithTx(tx)
+
+	isSeed, err := q.SessionIsSeed(ctx, id)
+	if err != nil {
+		return false, fmt.Errorf("delete seed session: probe seed state for %s: %w", id, err)
+	}
+	if !isSeed {
+		// Commit the empty tx so we don't leak a transaction. Critically, do
+		// NOT touch change_log here — for a live session that contains real
+		// session_created / session_updated CDC events.
+		if err := tx.Commit(); err != nil {
+			return false, fmt.Errorf("delete seed session: commit no-op: %w", err)
+		}
+		return false, nil
+	}
+
+	// Drop change_log rows for this session id first so the FK doesn't reject
+	// the session DELETE. We do not touch project-level events (session_id IS
+	// NULL) — those belong to the project, not this session. Both this DELETE
+	// and the session DELETE below run via raw ExecContext to sidestep sqlc
+	// 1.31's SQLite-parser bug, which strips trailing `?` placeholders and
+	// string literals from DELETE statements (see queries/changelog.sql and
+	// queries/sessions.sql for the documented workaround context).
+	if _, err := tx.ExecContext(ctx, `DELETE FROM change_log WHERE session_id = ?`, id); err != nil {
+		return false, fmt.Errorf("delete seed session: clear change log for %s: %w", id, err)
+	}
+	res, err := tx.ExecContext(ctx, `
+DELETE FROM sessions
+WHERE id = ?
+  AND is_terminated = 0
+  AND workspace_path = ''
+  AND runtime_handle_id = ''
+  AND agent_session_id = ''
+  AND prompt = ''`, id)
+	if err != nil {
+		return false, fmt.Errorf("delete seed session %s: %w", id, err)
+	}
+	n, err := res.RowsAffected()
+	if err != nil {
+		return false, fmt.Errorf("delete seed session %s: rows affected: %w", id, err)
+	}
+	if err := tx.Commit(); err != nil {
+		return false, fmt.Errorf("delete seed session: commit: %w", err)
+	}
+	return n > 0, nil
+}
+
 // GetSession returns the full record for a session, or ok=false if absent.
 func (s *Store) GetSession(ctx context.Context, id domain.SessionID) (domain.SessionRecord, bool, error) {
 	row, err := s.qr.GetSession(ctx, id)

backend/internal/storage/sqlite/store/store_test.go

@@ -102,6 +102,76 @@ func TestSessionCreateAssignsPerProjectID(t *testing.T) {
 	}
 }
 
+// TestDeleteSessionOnlyRemovesSeedRows covers Bug 4's storage-layer guarantee:
+// DeleteSession removes a session row only when the row is still in seed state
+// (no workspace, no runtime handle, no agent session id, no prompt, not
+// terminated). Rows that already carry spawn output are immutable so the
+// no-resurrection guarantee for live sessions still holds.
+func TestDeleteSessionOnlyRemovesSeedRows(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+	seedProject(t, s, "mer")
+
+	// Seed row: just CreateSession output, no metadata yet.
+	now := time.Now().UTC().Truncate(time.Second)
+	seed := domain.SessionRecord{
+		ProjectID: "mer",
+		Kind:      domain.KindWorker,
+		Harness:   domain.HarnessClaudeCode,
+		Activity:  domain.Activity{State: domain.ActivityIdle, LastActivityAt: now},
+		CreatedAt: now,
+		UpdatedAt: now,
+	}
+	r1, err := s.CreateSession(ctx, seed)
+	if err != nil {
+		t.Fatalf("create seed: %v", err)
+	}
+
+	deleted, err := s.DeleteSession(ctx, r1.ID)
+	if err != nil || !deleted {
+		t.Fatalf("delete seed = %v %v, want true nil", deleted, err)
+	}
+	if _, ok, _ := s.GetSession(ctx, r1.ID); ok {
+		t.Fatal("seed row still present after DeleteSession")
+	}
+
+	// A row with workspace_path populated must NOT be deleted — even if
+	// !is_terminated. This is the no-resurrection guarantee for live work.
+	r2, err := s.CreateSession(ctx, sampleRecord("mer"))
+	if err != nil {
+		t.Fatalf("create live: %v", err)
+	}
+	deleted, err = s.DeleteSession(ctx, r2.ID)
+	if err != nil {
+		t.Fatalf("delete live err = %v", err)
+	}
+	if deleted {
+		t.Fatal("DeleteSession must be a no-op for rows with spawn output")
+	}
+	if _, ok, _ := s.GetSession(ctx, r2.ID); !ok {
+		t.Fatal("live row was removed by DeleteSession")
+	}
+
+	// A terminated row is also out of scope: terminal-state rows hold cleanup
+	// metadata users may still inspect, so the gate refuses them too.
+	r3, err := s.CreateSession(ctx, seed)
+	if err != nil {
+		t.Fatalf("create extra seed: %v", err)
+	}
+	terminated := r3
+	terminated.IsTerminated = true
+	if err := s.UpdateSession(ctx, terminated); err != nil {
+		t.Fatalf("mark terminated: %v", err)
+	}
+	deleted, err = s.DeleteSession(ctx, r3.ID)
+	if err != nil {
+		t.Fatalf("delete terminated err = %v", err)
+	}
+	if deleted {
+		t.Fatal("DeleteSession must be a no-op for terminated rows")
+	}
+}
+
 func TestSessionRenameUpdatesDisplayName(t *testing.T) {
 	s := newTestStore(t)
 	ctx := context.Background()

frontend/src/api/schema.ts

@@ -247,6 +247,23 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/api/v1/sessions/{sessionId}/rollback": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        get?: never;
+        put?: never;
+        /** Undo a partially-completed spawn (delete seed row, or kill if spawn output exists) */
+        post: operations["rollbackSession"];
+        delete?: never;
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
     "/api/v1/sessions/{sessionId}/send": {
         parameters: {
             query?: never;
@@ -396,6 +413,12 @@ export interface components {
             session: components["schemas"]["Session"];
             sessionId: string;
         };
+        RollbackSessionResponse: {
+            deleted?: boolean;
+            killed?: boolean;
+            ok: boolean;
+            sessionId: string;
+        };
         SCMConfig: {
             package?: string;
             path?: string;
@@ -1467,6 +1490,56 @@ export interface operations {
             };
         };
     };
+    rollbackSession: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path: {
+                /** @description Session identifier, e.g. project-1. */
+                sessionId: string;
+            };
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description OK */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["RollbackSessionResponse"];
+                };
+            };
+            /** @description Not Found */
+            404: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["APIError"];
+                };
+            };
+            /** @description Conflict */
+            409: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["APIError"];
+                };
+            };
+            /** @description Internal Server Error */
+            500: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["APIError"];
+                };
+            };
+        };
+    };
     sendSessionMessage: {
         parameters: {
             query?: never;