feat(redact): move OPF to pre-push, add prompt + shadow cleanup

CLAUDE.md

@@ -442,6 +442,7 @@ The manual-commit strategy (`manual_commit*.go`) does not modify the active bran
 - **Shadow branch migration** - if user does stash/pull/rebase (HEAD changes without commit), shadow branch is automatically moved to new base commit
 - **Orphaned branch cleanup** - if a shadow branch exists without a corresponding session state file, it is automatically reset when a new session starts
 - PrePush hook can push `entire/checkpoints/v1` branch alongside user pushes
+- **OPF (OpenAI Privacy Filter) runs at pre-push, not post-commit**: when `redaction.openai_privacy_filter.enabled` is true, the PrePush hook re-redacts unpushed `entire/checkpoints/v1` commits with the OPF 8th layer, builds new commits carrying an `Entire-OPF-Applied: true` trailer, and atomically updates the local v1 ref before pushing. Per-commit condensation stays on the fast 7-layer pipeline. See `strategy/manual_commit_opf_rewrite.go` and `docs/security-and-privacy.md` for the full flow, including divergence detection, bootstrap caps, and CAS-on-conflict semantics.
 - Safe to use on main/master since it never modifies commit history
 
 #### Key Files
@@ -450,6 +451,7 @@ The manual-commit strategy (`manual_commit*.go`) does not modify the active bran
 - `common.go` - Helpers for metadata extraction, tree building, rewind validation, `ListCheckpoints()`
 - `session.go` - Session/checkpoint data structures
 - `push_common.go` - PrePush logic for pushing `entire/checkpoints/v1` branch
+- `manual_commit_opf_rewrite.go` - Pre-push OPF re-redaction: walks unpushed v1 commits, runs OPF over their blobs, rebuilds commits with `Entire-OPF-Applied: true` trailer, CAS-updates the local ref. Sentinel errors: `ErrV1Diverged`, `ErrBootstrapTooLarge`, `ErrV1RefMoved`.
 - `manual_commit.go` - Manual-commit strategy main implementation
 - `manual_commit_types.go` - Type definitions: `SessionState`, `CheckpointInfo`, `CondenseResult`
 - `manual_commit_session.go` - Session state management (load/save/list session states)

cmd/entire/cli/checkpoint/checkpoint.go

@@ -222,22 +222,11 @@ type WriteCommittedOptions struct {
 	// Must be pre-redacted (via redact.JSONLBytes or redact.AlreadyRedacted for trusted sources).
 	Transcript redact.RedactedBytes
 
-	// Prompts contains the raw user prompts from the session. These are NOT
-	// guaranteed to be redacted on entry — the writer always emits the typed
-	// PromptsRedacted blob below (running the safety-net pipeline if it is
-	// the zero value). Do not read Prompts independently for persistence; go
-	// through redactJoinedPrompts so the redaction guarantee is preserved.
+	// Prompts contains the raw user prompts from the session. Run through
+	// redactedJoinedPrompts before persisting — the writer does this
+	// inside writeSessionToSubdirectory.
 	Prompts []string
 
-	// PromptsRedacted, when set, is the pre-redacted joined-prompts blob the
-	// writer uses verbatim instead of re-running the safety-net pipeline.
-	// Used by finalizeAllTurnCheckpoints to avoid running the OpenAI
-	// Privacy Filter once per checkpoint over identical joined-prompt
-	// strings. The typed wrapper makes the "this content was produced by
-	// the redaction pipeline" claim a compile-time invariant — callers
-	// cannot assign an arbitrary string.
-	PromptsRedacted redact.RedactedJoinedPrompts
-
 	// FilesTouched are files modified during the session
 	FilesTouched []string
 
@@ -366,16 +355,10 @@ type UpdateCommittedOptions struct {
 	// Must be pre-redacted (via redact.JSONLBytes or redact.AlreadyRedacted for trusted sources).
 	Transcript redact.RedactedBytes
 
-	// Prompts contains the raw user prompts (replaces existing). NOT
-	// guaranteed to be redacted on entry — see WriteCommittedOptions.Prompts
-	// for the relationship to PromptsRedacted.
+	// Prompts contains the raw user prompts (replaces existing).
+	// See WriteCommittedOptions.Prompts.
 	Prompts []string
 
-	// PromptsRedacted, when set, is the pre-redacted joined-prompts blob
-	// the writer uses verbatim instead of re-running the safety-net
-	// pipeline. See WriteCommittedOptions.PromptsRedacted for rationale.
-	PromptsRedacted redact.RedactedJoinedPrompts
-
 	// Agent identifies the agent type (needed for transcript chunking)
 	Agent types.AgentType
 

cmd/entire/cli/checkpoint/committed.go

@@ -417,12 +417,10 @@ func (s *GitStore) writeSessionToSubdirectory(ctx context.Context, opts WriteCom
 		filePaths.ContentHash = "/" + sessionPath + paths.ContentHashFileName
 	}
 
-	// Write prompts. Uses the full 8-layer pipeline (including OPF) via
-	// redactedJoinedPrompts; the helper unwraps opts.PromptsRedacted when
-	// set so callers (finalizeAllTurnCheckpoints) that pre-redact once
-	// across multiple checkpoint writes don't pay OPF per checkpoint.
+	// Write prompts via the 7-layer pipeline. OPF runs only in the
+	// pre-push rewrite path (manual_commit_opf_rewrite.go).
 	if len(opts.Prompts) > 0 {
-		promptContent := redactedJoinedPrompts(ctx, opts.Prompts, opts.PromptsRedacted)
+		promptContent := redactedJoinedPrompts(opts.Prompts)
 		blobHash, err := CreateBlobFromContent(s.repo, []byte(promptContent))
 		if err != nil {
 			return filePaths, err
@@ -1403,10 +1401,9 @@ func (s *GitStore) UpdateCommitted(ctx context.Context, opts UpdateCommittedOpti
 		}
 	}
 
-	// Replace prompts (apply redaction as safety net; unwraps
-	// opts.PromptsRedacted when set).
+	// Replace prompts with 7-layer-redacted content.
 	if len(opts.Prompts) > 0 {
-		promptContent := redactedJoinedPrompts(ctx, opts.Prompts, opts.PromptsRedacted)
+		promptContent := redactedJoinedPrompts(opts.Prompts)
 		blobHash, err := CreateBlobFromContent(s.repo, []byte(promptContent))
 		if err != nil {
 			return fmt.Errorf("failed to create prompt blob: %w", err)
@@ -1725,11 +1722,13 @@ func (s *GitStore) copyMetadataDir(ctx context.Context, metadataDir, basePath st
 			return fmt.Errorf("path traversal detected: %s", relPath)
 		}
 
-		// Create blob from file with secrets redaction
-		// Committed-checkpoint write — run the full 8-layer pipeline
-		// including OPF. The per-turn temp-write path stays on plain
-		// redactors via the sibling createRedactedBlobFromFile.
-		blobHash, mode, err := createRedactedBlobFromFileWithPrivacyFilter(ctx, s.repo, path, relPath)
+		// Create blob from file with 7-layer secrets redaction.
+		// Post-commit emits 7-layer-only blobs; the OPF-capable variant
+		// (createRedactedBlobFromFileWithPrivacyFilter) is used later by
+		// the pre-push rewrite path, which re-redacts these blobs into
+		// 8-layer commits before they leave the local machine.
+		_ = ctx // ctx not needed by the 7-layer path; kept on caller signature for future use
+		blobHash, mode, err := createRedactedBlobFromFile(s.repo, path, relPath)
 		if err != nil {
 			return fmt.Errorf("failed to create blob for %s: %w", path, err)
 		}
@@ -1751,22 +1750,13 @@ func (s *GitStore) copyMetadataDir(ctx context.Context, metadataDir, basePath st
 }
 
 // createRedactedBlobFromFile reads a file, applies the 7-layer redaction
-// pipeline, and creates a git blob. Used by per-turn temporary-checkpoint
-// writes — the OpenAI Privacy Filter is intentionally NOT run here to
-// keep per-turn latency inside the agent loop's budget.
+// pipeline, and creates a git blob. Used by committed-checkpoint writes
+// at post-commit time. The OpenAI Privacy Filter is intentionally NOT
+// run here — OPF lives in the pre-push rewrite path
+// (strategy/manual_commit_opf_rewrite.go), which re-redacts the 7-layer
+// blobs into 8-layer commits before they leave the local machine.
 // JSONL files get JSONL-aware redaction; all other files get plain byte redaction.
 func createRedactedBlobFromFile(repo *git.Repository, filePath, treePath string) (plumbing.Hash, filemode.FileMode, error) {
-	return createRedactedBlobFromFileImpl(context.Background(), repo, filePath, treePath, false)
-}
-
-// createRedactedBlobFromFileWithPrivacyFilter reads a file, applies the full
-// 8-layer pipeline (including the OpenAI Privacy Filter), and creates a git
-// blob. Used by committed-checkpoint writes — slower but more thorough.
-func createRedactedBlobFromFileWithPrivacyFilter(ctx context.Context, repo *git.Repository, filePath, treePath string) (plumbing.Hash, filemode.FileMode, error) {
-	return createRedactedBlobFromFileImpl(ctx, repo, filePath, treePath, true)
-}
-
-func createRedactedBlobFromFileImpl(ctx context.Context, repo *git.Repository, filePath, treePath string, usePrivacyFilter bool) (plumbing.Hash, filemode.FileMode, error) {
 	info, err := os.Stat(filePath)
 	if err != nil {
 		return plumbing.ZeroHash, 0, fmt.Errorf("failed to stat file: %w", err)
@@ -1793,7 +1783,7 @@ func createRedactedBlobFromFileImpl(ctx context.Context, repo *git.Repository, f
 		return hash, mode, nil
 	}
 
-	content = redactBytesForBlob(ctx, content, treePath, usePrivacyFilter)
+	content = RedactBlobBytes(context.Background(), content, treePath, false)
 
 	hash, err := CreateBlobFromContent(repo, content)
 	if err != nil {
@@ -1802,14 +1792,15 @@ func createRedactedBlobFromFileImpl(ctx context.Context, repo *git.Repository, f
 	return hash, mode, nil
 }
 
-// redactBytesForBlob applies the appropriate redaction pipeline to file
-// content for a checkpoint blob. JSONL files get JSONL-aware redaction
-// (falling back to plain byte redaction on parse failure so the regex
-// layers still apply); other files get plain byte redaction.
-// usePrivacyFilter selects the lighter 7-layer pipeline (per-turn temp
-// writes) versus the full 8-layer pipeline including OPF (committed
-// writes).
-func redactBytesForBlob(ctx context.Context, content []byte, treePath string, usePrivacyFilter bool) []byte {
+// RedactBlobBytes redacts a single blob's content given its tree path.
+// JSONL files get JSONL-aware redaction (falling back to plain bytes on
+// parse failure so regex/credential layers still apply); other files
+// get plain byte redaction. When usePrivacyFilter is true the full
+// 8-layer pipeline (including OPF) runs; otherwise the 7-layer pipeline.
+//
+// Post-commit condensation uses false (fast path). The pre-push rewrite
+// (strategy/manual_commit_opf_rewrite.go) uses true.
+func RedactBlobBytes(ctx context.Context, content []byte, treePath string, usePrivacyFilter bool) []byte {
 	if strings.HasSuffix(treePath, ".jsonl") {
 		var (
 			redacted redact.RedactedBytes
@@ -1823,8 +1814,7 @@ func redactBytesForBlob(ctx context.Context, content []byte, treePath string, us
 		if err == nil {
 			return redacted.Bytes()
 		}
-		// JSONL parse failed — fall through so regex/credential layers
-		// still apply via the plain byte path.
+		// JSONL parse failed — fall through to plain bytes.
 	}
 	if usePrivacyFilter {
 		return redact.BytesWithPrivacyFilter(ctx, content)

cmd/entire/cli/checkpoint/committed_opf_trailer_test.go

@@ -0,0 +1,70 @@
+package checkpoint
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/entireio/cli/cmd/entire/cli/checkpoint/id"
+	"github.com/entireio/cli/cmd/entire/cli/testutil"
+	"github.com/entireio/cli/cmd/entire/cli/trailers"
+	"github.com/entireio/cli/redact"
+	"github.com/go-git/go-git/v6"
+	"github.com/go-git/go-git/v6/plumbing"
+	"github.com/go-git/go-git/v6/plumbing/object"
+	"github.com/stretchr/testify/require"
+)
+
+// TestWriteCommitted_DoesNotEmitOPFAppliedTrailer is the regression guard
+// for the architectural promise: standard post-commit condensation writes
+// 7-layer-only blobs and MUST NOT mark them with the Entire-OPF-Applied
+// trailer. The trailer is emitted exclusively by the pre-push rewrite
+// path; if a future change accidentally added it to the standard writer,
+// the pre-push rewrite would skip those commits (HasOPFApplied true →
+// reparent-only, no actual OPF run) and ship 7-layer content as if it
+// were 8-layer. This test pins down that contract.
+func TestWriteCommitted_DoesNotEmitOPFAppliedTrailer(t *testing.T) {
+	t.Parallel()
+
+	tempDir := t.TempDir()
+	testutil.InitRepo(t, tempDir)
+	repo, err := git.PlainOpen(tempDir)
+	require.NoError(t, err)
+
+	wt, err := repo.Worktree()
+	require.NoError(t, err)
+	readmeFile := filepath.Join(tempDir, "README.md")
+	require.NoError(t, os.WriteFile(readmeFile, []byte("# Test"), 0o644))
+	_, err = wt.Add("README.md")
+	require.NoError(t, err)
+	_, err = wt.Commit("Initial commit", &git.CommitOptions{
+		Author: &object.Signature{Name: "Test", Email: "test@test.com"},
+	})
+	require.NoError(t, err)
+
+	store := NewGitStore(repo)
+	cpID := id.MustCheckpointID("a1b2c3d4e5f6")
+
+	err = store.WriteCommitted(context.Background(), WriteCommittedOptions{
+		CheckpointID: cpID,
+		SessionID:    "regression-no-opf-trailer",
+		Strategy:     "manual-commit",
+		Transcript:   redact.AlreadyRedacted([]byte(`{"role":"user","content":"hello"}` + "\n")),
+		AuthorName:   "Test",
+		AuthorEmail:  "test@test.com",
+	})
+	require.NoError(t, err)
+
+	// Read the latest commit message on entire/checkpoints/v1 and assert
+	// HasOPFApplied is false. We resolve via the ref then walk back the
+	// single commit the writer just produced.
+	ref, err := repo.Reference(plumbing.NewBranchReferenceName("entire/checkpoints/v1"), true)
+	require.NoError(t, err, "writer should have created entire/checkpoints/v1")
+	commit, err := repo.CommitObject(ref.Hash())
+	require.NoError(t, err)
+
+	if trailers.HasOPFApplied(commit.Message) {
+		t.Errorf("standard WriteCommitted emitted Entire-OPF-Applied trailer; commit message:\n%s", commit.Message)
+	}
+}

cmd/entire/cli/checkpoint/prompts.go

@@ -1,7 +1,6 @@
 package checkpoint
 
 import (
-	"context"
 	"strings"
 
 	"github.com/entireio/cli/redact"
@@ -29,16 +28,9 @@ func SplitPromptContent(content string) []string {
 	return prompts
 }
 
-// redactedJoinedPrompts returns the redacted prompt-blob content for the
-// supplied prompts. When preRedacted is set it is unwrapped verbatim;
-// otherwise the prompts are joined and run through the full 8-layer
-// pipeline as a safety net. Callers that share the same prompts across
-// multiple checkpoint writes (finalizeAllTurnCheckpoints) should compute
-// the redacted blob once via redact.JoinedPrompts and pass it through to
-// avoid running OPF repeatedly over identical input.
-func redactedJoinedPrompts(ctx context.Context, prompts []string, preRedacted redact.RedactedJoinedPrompts) string {
-	if preRedacted.IsSet() {
-		return preRedacted.String()
-	}
-	return redact.JoinedPrompts(ctx, prompts, PromptSeparator).String()
+// redactedJoinedPrompts joins prompts and runs the 7-layer redaction
+// pipeline. OPF runs exclusively in the pre-push rewrite (not here),
+// so the writer's hot path stays predictable.
+func redactedJoinedPrompts(prompts []string) string {
+	return redact.String(strings.Join(prompts, PromptSeparator))
 }

cmd/entire/cli/checkpoint/prompts_test.go

@@ -1,10 +1,8 @@
 package checkpoint
 
 import (
-	"context"
 	"testing"
 
-	"github.com/entireio/cli/redact"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -16,7 +14,6 @@ func TestJoinAndSplitPrompts_RoundTrip(t *testing.T) {
 		"first line\nwith newline",
 		"second prompt",
 	}
-
 	joined := JoinPrompts(original)
 	split := SplitPromptContent(joined)
 
@@ -26,34 +23,15 @@ func TestJoinAndSplitPrompts_RoundTrip(t *testing.T) {
 
 func TestSplitPromptContent_EmptyContent(t *testing.T) {
 	t.Parallel()
-
 	assert.Nil(t, SplitPromptContent(""))
 }
 
-// TestRedactedJoinedPrompts_PreRedactedIsTrustedVerbatim verifies that when
-// the caller supplies a set RedactedJoinedPrompts the helper unwraps it
-// untouched and never re-invokes the redaction pipeline. The pre-redacted
-// path is what finalizeAllTurnCheckpoints relies on to avoid running OPF
-// once per checkpoint over identical joined-prompt strings.
-func TestRedactedJoinedPrompts_PreRedactedIsTrustedVerbatim(t *testing.T) {
+// TestRedactedJoinedPrompts_AppliesSafetyNet verifies the helper joins
+// prompts with the canonical separator and runs them through the 7-layer
+// pipeline. OPF runs only in the pre-push rewrite path, never here.
+func TestRedactedJoinedPrompts_AppliesSafetyNet(t *testing.T) {
 	t.Parallel()
-
-	const preRedacted = "[REDACTED_PERSON] asked about [REDACTED_EMAIL]"
-	got := redactedJoinedPrompts(
-		context.Background(),
-		[]string{"raw prompt text"},
-		redact.AlreadyRedactedJoinedPrompts(preRedacted),
-	)
-	assert.Equal(t, preRedacted, got, "preRedacted should pass through verbatim")
-}
-
-// TestRedactedJoinedPrompts_ZeroValueFallsBackToRedaction verifies that
-// when the typed preRedacted is the zero value the helper joins the
-// prompts and runs the full pipeline as a safety net.
-func TestRedactedJoinedPrompts_ZeroValueFallsBackToRedaction(t *testing.T) {
-	t.Parallel()
-
-	got := redactedJoinedPrompts(context.Background(), []string{"hello", "world"}, redact.RedactedJoinedPrompts{})
-	assert.NotEmpty(t, got, "zero-value preRedacted should fall back to running the redaction pipeline")
-	assert.Contains(t, got, PromptSeparator, "fallback output should preserve the prompt separator")
+	got := redactedJoinedPrompts([]string{"hello", "world"})
+	assert.NotEmpty(t, got)
+	assert.Contains(t, got, PromptSeparator)
 }

cmd/entire/cli/checkpoint/v2_committed.go

@@ -609,7 +609,7 @@ func (s *V2GitStore) updateCommittedMain(ctx context.Context, opts UpdateCommitt
 	sessionPath := fmt.Sprintf("%s%d/", basePath, sessionIndex)
 
 	if len(opts.Prompts) > 0 {
-		promptContent := redactedJoinedPrompts(ctx, opts.Prompts, opts.PromptsRedacted)
+		promptContent := redactedJoinedPrompts(opts.Prompts)
 		blobHash, err := CreateBlobFromContent(s.repo, []byte(promptContent))
 		if err != nil {
 			return 0, fmt.Errorf("failed to create prompt blob: %w", err)
@@ -891,7 +891,7 @@ func (s *V2GitStore) writeMainCheckpointEntries(ctx context.Context, opts WriteC
 // and compact transcript to a session subdirectory (0/, 1/, 2/, … indexed by
 // session order within the checkpoint). The raw transcript (raw_transcript) and its
 // content hash (raw_transcript_hash.txt) go to /full/current, not here.
-func (s *V2GitStore) writeMainSessionToSubdirectory(ctx context.Context, opts WriteCommittedOptions, sessionPath string, entries map[string]object.TreeEntry) (SessionFilePaths, error) {
+func (s *V2GitStore) writeMainSessionToSubdirectory(_ context.Context, opts WriteCommittedOptions, sessionPath string, entries map[string]object.TreeEntry) (SessionFilePaths, error) {
 	filePaths := SessionFilePaths{}
 
 	// Clear existing entries at this session path
@@ -903,7 +903,7 @@ func (s *V2GitStore) writeMainSessionToSubdirectory(ctx context.Context, opts Wr
 
 	// Write prompts
 	if len(opts.Prompts) > 0 {
-		promptContent := redactedJoinedPrompts(ctx, opts.Prompts, opts.PromptsRedacted)
+		promptContent := redactedJoinedPrompts(opts.Prompts)
 		blobHash, err := CreateBlobFromContent(s.repo, []byte(promptContent))
 		if err != nil {
 			return filePaths, err