feat: min-severity cascading and review-level filter (#635)

## Summary

- Add `min_severity` config cascade (CLI flag → repo `.roborev.toml` → global `config.toml`) for review, refine, and fix workflows
- Persist `min_severity` on `review_jobs` (SQLite + PostgreSQL v11), thread through enqueue/claim/sync paths
- Inject severity instructions into review and fix prompts via `config.SeverityInstruction`
- Parse `SEVERITY_THRESHOLD_MET` marker as a pass verdict
- CI path: prompt uses `review_min_severity`, synthesis uses `[ci].min_severity`; single-result CI runs route through synthesis when CI threshold is set
- `formatSingleResult` uses `storage.ParseVerdict` to detect pass output (not loose `strings.Contains`)
- Fix `GetJobsWithReviewsByIDs` hydration bug where `min_severity` was scanned then overwritten by zero-value field

Co-authored-by: Phillip Cloud <cpcloud@users.noreply.github.com>

Author: cpcloud

cmd/roborev/ci.go

@@ -189,31 +189,40 @@ func runCIReview(ctx context.Context, opts ciReviewOpts) error {
 		return err
 	}
 
-	// Resolve min severity
-	minSev, err := config.ResolveCIMinSeverity(
+	// Resolve min severity for synthesis output filtering
+	ciMinSev, err := config.ResolveCIMinSeverity(
 		opts.minSeverity, repoCfg, globalCfg)
 	if err != nil {
 		return err
 	}
 
+	// Resolve review-level min severity for prompt injection
+	reviewMinSev, err := config.ResolveReviewMinSeverity(
+		"", root, globalCfg)
+	if err != nil {
+		return err
+	}
+
 	// Resolve synthesis agent
 	synthAgent := config.ResolveCISynthesisAgent(
 		opts.synthesisAgent, repoCfg, globalCfg)
 
 	log.Printf(
 		"ci review: ref=%s agents=%v types=%v "+
-			"reasoning=%s min_severity=%s",
+			"reasoning=%s ci_min_severity=%s review_min_severity=%s",
 		gitRef, agents, reviewTypes,
-		reasoningLevel, minSev)
+		reasoningLevel, ciMinSev, reviewMinSev)
 
-	// Run batch
+	// Run batch with review-level severity for prompt injection.
+	// CI-level severity is applied separately during synthesis.
 	batchCfg := review.BatchConfig{
 		RepoPath:     root,
 		GitRef:       gitRef,
 		Agents:       agents,
 		ReviewTypes:  reviewTypes,
 		Reasoning:    reasoningLevel,
 		GlobalConfig: globalCfg,
+		MinSeverity:  reviewMinSev,
 	}
 
 	results := review.RunBatch(ctx, batchCfg)
@@ -225,7 +234,7 @@ func runCIReview(ctx context.Context, opts ciReviewOpts) error {
 	comment, synthErr := review.Synthesize(
 		ctx, results, review.SynthesizeOpts{
 			Agent:        synthAgent,
-			MinSeverity:  minSev,
+			MinSeverity:  ciMinSev,
 			RepoPath:     root,
 			GitRef:       gitRef,
 			HeadSHA:      headSHA,

cmd/roborev/fix.go

@@ -799,8 +799,9 @@ func fixSingleJob(cmd *cobra.Command, repoRoot string, jobID int64, opts fixOpti
 	// task/analyze jobs have free-form output without severity labels)
 	var minSev string
 	if !job.IsTaskJob() {
+		fixCfg, _ := config.LoadGlobal()
 		minSev, err = config.ResolveFixMinSeverity(
-			opts.minSeverity, repoRoot,
+			opts.minSeverity, repoRoot, fixCfg,
 		)
 		if err != nil {
 			return fmt.Errorf("resolve min-severity: %w", err)
@@ -1011,11 +1012,14 @@ func runFixBatch(cmd *cobra.Command, jobIDs []int64, branch string, allBranches,
 		return err
 	}
 
+	// Load global config for severity + prompt-size resolution.
+	cfg, _ := config.LoadGlobal()
+
 	// Resolve minimum severity filter. Suppress if any entry is a
 	// task job — task/analyze output has no severity labels, so the
 	// instruction would confuse the agent for those entries.
 	minSev, err := config.ResolveFixMinSeverity(
-		opts.minSeverity, roots.worktreeRoot,
+		opts.minSeverity, roots.worktreeRoot, cfg,
 	)
 	if err != nil {
 		return fmt.Errorf("resolve min-severity: %w", err)
@@ -1031,7 +1035,6 @@ func runFixBatch(cmd *cobra.Command, jobIDs []int64, branch string, allBranches,
 
 	// Split into batches by prompt size (after severity resolution
 	// so the severity instruction overhead is accounted for)
-	cfg, _ := config.LoadGlobal()
 	maxSize := config.ResolveMaxPromptSize(roots.worktreeRoot, cfg)
 	batches := splitIntoBatches(entries, maxSize, minSev)
 

cmd/roborev/refine.go

@@ -394,7 +394,7 @@ func runRefine(ctx RunContext, opts refineOptions) error {
 
 	// Resolve minimum severity filter
 	minSev, err := config.ResolveRefineMinSeverity(
-		opts.minSeverity, repoPath,
+		opts.minSeverity, repoPath, cfg,
 	)
 	if err != nil {
 		return fmt.Errorf("resolve min-severity: %w", err)
@@ -616,10 +616,10 @@ func runRefine(ctx RunContext, opts refineOptions) error {
 
 			// When severity filtering is active and the agent
 			// signals all findings are below threshold, treat as
-			// resolved rather than a fix failure.
-			if minSev != "" && strings.Contains(
-				output, config.SeverityThresholdMarker,
-			) {
+			// resolved rather than a fix failure. Require the
+			// marker to stand alone so prose findings echoed
+			// alongside it cannot silently close the review.
+			if minSev != "" && config.IsMarkerOnlyOutput(output) {
 				fmt.Println(
 					"All findings below severity " +
 						"threshold - closing review",

cmd/roborev/review.go

@@ -26,21 +26,22 @@ const MaxDirtyDiffSize = 200 * 1024
 
 func reviewCmd() *cobra.Command {
 	var (
-		repoPath   string
-		sha        string
-		agent      string
-		model      string
-		reasoning  string
-		reviewType string
-		fast       bool
-		quiet      bool
-		dirty      bool
-		wait       bool
-		branch     string
-		baseBranch string
-		since      string
-		local      bool
-		provider   string
+		repoPath    string
+		sha         string
+		agent       string
+		model       string
+		reasoning   string
+		reviewType  string
+		fast        bool
+		quiet       bool
+		dirty       bool
+		wait        bool
+		branch      string
+		baseBranch  string
+		since       string
+		local       bool
+		provider    string
+		minSeverity string
 	)
 
 	cmd := &cobra.Command{
@@ -270,7 +271,7 @@ Examples:
 
 			// Handle --local mode: run agent directly without daemon
 			if local {
-				return runLocalReview(cmd, root, gitRef, diffContent, agent, model, provider, reasoning, reviewType, quiet)
+				return runLocalReview(cmd, root, gitRef, diffContent, agent, model, provider, reasoning, reviewType, quiet, minSeverity)
 			}
 
 			// Build request body
@@ -284,6 +285,7 @@ Examples:
 				Reasoning:   reasoning,
 				ReviewType:  reviewType,
 				DiffContent: diffContent,
+				MinSeverity: minSeverity,
 			}
 
 			reqBody, _ := json.Marshal(reqFields)
@@ -358,14 +360,15 @@ Examples:
 	cmd.Flags().BoolVar(&local, "local", false, "run review locally without daemon (streams output to console)")
 	cmd.Flags().StringVar(&reviewType, "type", "", "review type (security, design) — changes system prompt")
 	cmd.Flags().StringVar(&provider, "provider", "", "provider for pi agent (e.g. anthropic, openai)")
+	cmd.Flags().StringVar(&minSeverity, "min-severity", "", "minimum severity threshold: critical, high, medium, low")
 	registerAgentCompletion(cmd)
 	registerReasoningCompletion(cmd)
 
 	return cmd
 }
 
 // runLocalReview runs a review directly without the daemon
-func runLocalReview(cmd *cobra.Command, repoPath, gitRef, diffContent, agentName, model, provider, reasoning, reviewType string, quiet bool) error {
+func runLocalReview(cmd *cobra.Command, repoPath, gitRef, diffContent, agentName, model, provider, reasoning, reviewType string, quiet bool, minSeverity string) error {
 	// Load config
 	cfg, err := config.LoadGlobal()
 	if err != nil {
@@ -378,6 +381,12 @@ func runLocalReview(cmd *cobra.Command, repoPath, gitRef, diffContent, agentName
 		return fmt.Errorf("invalid reasoning: %w", err)
 	}
 
+	// Resolve review min-severity
+	resolvedMinSev, err := config.ResolveReviewMinSeverity(minSeverity, repoPath, cfg)
+	if err != nil {
+		return fmt.Errorf("invalid min-severity: %w", err)
+	}
+
 	// Map review_type to config workflow (matches daemon behavior)
 	workflow := "review"
 	if !config.IsDefaultReviewType(reviewType) {
@@ -435,13 +444,13 @@ func runLocalReview(cmd *cobra.Command, repoPath, gitRef, diffContent, agentName
 	var snapshotCleanup func()
 	if diffContent != "" {
 		// Dirty review
-		dirtyResult, dirtyErr := pb.BuildDirtyWithSnapshot(repoPath, diffContent, 0, cfg.ReviewContextCount, a.Name(), reviewType)
+		dirtyResult, dirtyErr := pb.BuildDirtyWithSnapshot(repoPath, diffContent, 0, cfg.ReviewContextCount, a.Name(), reviewType, resolvedMinSev)
 		reviewPrompt = dirtyResult.Prompt
 		snapshotCleanup = dirtyResult.Cleanup
 		err = dirtyErr
 	} else {
 		excludes := config.ResolveExcludePatterns(repoPath, cfg, reviewType)
-		result, buildErr := pb.BuildWithSnapshot(repoPath, gitRef, 0, cfg.ReviewContextCount, a.Name(), reviewType, excludes)
+		result, buildErr := pb.BuildWithSnapshot(repoPath, gitRef, 0, cfg.ReviewContextCount, a.Name(), reviewType, resolvedMinSev, excludes)
 		reviewPrompt = result.Prompt
 		snapshotCleanup = result.Cleanup
 		err = buildErr

internal/config/config.go

@@ -144,6 +144,11 @@ type Config struct {
 	SecurityBackupModel string `toml:"security_backup_model"`
 	DesignBackupModel   string `toml:"design_backup_model"`
 
+	// Minimum severity thresholds (global defaults)
+	ReviewMinSeverity string `toml:"review_min_severity" comment:"Minimum severity for reviews: critical, high, medium, or low. Empty disables filtering."`
+	RefineMinSeverity string `toml:"refine_min_severity" comment:"Minimum severity for refine: critical, high, medium, or low. Empty disables filtering."`
+	FixMinSeverity    string `toml:"fix_min_severity" comment:"Minimum severity for fix: critical, high, medium, or low. Empty disables filtering."`
+
 	AllowUnsafeAgents   *bool `toml:"allow_unsafe_agents"`   // nil = not set, allows commands to choose their own default
 	DisableCodexSandbox bool  `toml:"disable_codex_sandbox"` // use --full-auto instead of --sandbox read-only (for systems where bwrap is broken)
 	ReuseReviewSession  *bool `toml:"reuse_review_session"`  // nil = not set; when true, reuse prior branch review sessions when possible
@@ -621,8 +626,9 @@ type RepoConfig struct {
 	ReviewReasoning            string   `toml:"review_reasoning" comment:"Reasoning level for reviews in this repo: fast, standard, medium, thorough, or maximum."`
 	RefineReasoning            string   `toml:"refine_reasoning" comment:"Reasoning level for refine in this repo: fast, standard, medium, thorough, or maximum."`
 	FixReasoning               string   `toml:"fix_reasoning" comment:"Reasoning level for fix in this repo: fast, standard, medium, thorough, or maximum."`
-	FixMinSeverity             string   `toml:"fix_min_severity" comment:"Minimum severity for fix in this repo: critical, high, medium, or low."`    // Minimum severity for fix: critical, high, medium, low
-	RefineMinSeverity          string   `toml:"refine_min_severity" comment:"Minimum severity for refine in this repo: critical, high, medium, low."` // Minimum severity for refine: critical, high, medium, low
+	FixMinSeverity             string   `toml:"fix_min_severity" comment:"Minimum severity for fix in this repo: critical, high, medium, or low."`     // Minimum severity for fix: critical, high, medium, low
+	RefineMinSeverity          string   `toml:"refine_min_severity" comment:"Minimum severity for refine in this repo: critical, high, medium, low."`  // Minimum severity for refine: critical, high, medium, low
+	ReviewMinSeverity          string   `toml:"review_min_severity" comment:"Minimum severity for reviews in this repo: critical, high, medium, low."` // Minimum severity for review: critical, high, medium, low
 	ExcludePatterns            []string `toml:"exclude_patterns" comment:"Filenames or glob patterns to exclude from review diffs for this repo."`
 	PostCommitReview           string   `toml:"post_commit_review" comment:"Automatic post-commit review mode for this repo: commit or branch."` // "commit" (default) or "branch"
 	ReuseReviewSession         *bool    `toml:"reuse_review_session"`
@@ -1442,6 +1448,54 @@ var severityAbove = map[string]string{
 // from "agent couldn't fix it."
 const SeverityThresholdMarker = "SEVERITY_THRESHOLD_MET"
 
+// IsMarkerOnlyOutput reports whether output is essentially the
+// SeverityThresholdMarker by itself, allowing only whitespace and
+// minimal markdown decoration: bold (**...** or __...__), italic
+// (*...* or _..._), a fenced code block, a leading list bullet,
+// and an optional trailing period. Any prose or other substantive
+// content disqualifies the output, since we cannot reliably tell
+// chatty narration from prose findings without severity labels.
+//
+// Callers that want to treat the marker as a "below threshold"
+// signal should use this helper rather than substring matching,
+// which is too easy to fool with marker-bearing prose findings.
+func IsMarkerOnlyOutput(output string) bool {
+	s := strings.TrimSpace(output)
+	if s == "" {
+		return false
+	}
+
+	// Strip a fenced code block if the output is wrapped in one.
+	if rest, ok := strings.CutPrefix(s, "```"); ok {
+		if _, after, found := strings.Cut(rest, "\n"); found {
+			s = after
+		} else {
+			s = rest
+		}
+		s = strings.TrimSuffix(s, "```")
+		s = strings.TrimSpace(s)
+	}
+
+	// Strip a leading list marker (- or *) followed by space.
+	if len(s) >= 2 && (s[0] == '-' || s[0] == '*') && s[1] == ' ' {
+		s = strings.TrimSpace(s[2:])
+	}
+
+	// Strip surrounding bold/italic markers. Bold forms (** and __)
+	// are stripped before italic forms (* and _) so that **X** fully
+	// unwraps in one pass rather than degrading to *X*.
+	for _, wrap := range []string{"**", "__", "*", "_"} {
+		if strings.HasPrefix(s, wrap) && strings.HasSuffix(s, wrap) && len(s) >= 2*len(wrap) {
+			s = strings.TrimSpace(s[len(wrap) : len(s)-len(wrap)])
+		}
+	}
+
+	// Strip a trailing period.
+	s = strings.TrimSpace(strings.TrimSuffix(s, "."))
+
+	return s == SeverityThresholdMarker
+}
+
 // SeverityInstruction returns a prompt instruction telling the agent
 // to focus only on findings at or above minSeverity. Returns "" for
 // empty, "low", or unrecognized input (no filtering needed).
@@ -1567,37 +1621,76 @@ func validateRepoReasoningOverride(
 }
 
 // ResolveFixMinSeverity determines minimum severity for fix.
-// Priority: explicit > per-repo config > "" (no filter)
-func ResolveFixMinSeverity(
-	explicit string, repoPath string,
-) (string, error) {
+// Priority: explicit > per-repo config > global config > "" (no filter)
+func ResolveFixMinSeverity(explicit string, repoPath string, globalCfg *Config) (string, error) {
 	if strings.TrimSpace(explicit) != "" {
 		return NormalizeMinSeverity(explicit)
 	}
-	if repoCfg, err := LoadRepoConfig(repoPath); err == nil &&
-		repoCfg != nil &&
-		strings.TrimSpace(repoCfg.FixMinSeverity) != "" {
+	if repoCfg, err := LoadRepoConfig(repoPath); err == nil && repoCfg != nil && strings.TrimSpace(repoCfg.FixMinSeverity) != "" {
 		return NormalizeMinSeverity(repoCfg.FixMinSeverity)
 	}
+	if globalCfg != nil && strings.TrimSpace(globalCfg.FixMinSeverity) != "" {
+		return NormalizeMinSeverity(globalCfg.FixMinSeverity)
+	}
 	return "", nil
 }
 
 // ResolveRefineMinSeverity determines minimum severity for refine.
-// Priority: explicit > per-repo config > "" (no filter)
-func ResolveRefineMinSeverity(
-	explicit string, repoPath string,
-) (string, error) {
+// Priority: explicit > per-repo config > global config > "" (no filter)
+func ResolveRefineMinSeverity(explicit string, repoPath string, globalCfg *Config) (string, error) {
 	if strings.TrimSpace(explicit) != "" {
 		return NormalizeMinSeverity(explicit)
 	}
-	if repoCfg, err := LoadRepoConfig(repoPath); err == nil &&
-		repoCfg != nil &&
-		strings.TrimSpace(repoCfg.RefineMinSeverity) != "" {
+	if repoCfg, err := LoadRepoConfig(repoPath); err == nil && repoCfg != nil && strings.TrimSpace(repoCfg.RefineMinSeverity) != "" {
 		return NormalizeMinSeverity(repoCfg.RefineMinSeverity)
 	}
+	if globalCfg != nil && strings.TrimSpace(globalCfg.RefineMinSeverity) != "" {
+		return NormalizeMinSeverity(globalCfg.RefineMinSeverity)
+	}
+	return "", nil
+}
+
+// ResolveReviewMinSeverity determines minimum severity for review.
+// Priority: explicit > per-repo config > global config > "" (no filter)
+func ResolveReviewMinSeverity(explicit string, repoPath string, globalCfg *Config) (string, error) {
+	if strings.TrimSpace(explicit) != "" {
+		return NormalizeMinSeverity(explicit)
+	}
+	if repoCfg, err := LoadRepoConfig(repoPath); err == nil && repoCfg != nil && strings.TrimSpace(repoCfg.ReviewMinSeverity) != "" {
+		return NormalizeMinSeverity(repoCfg.ReviewMinSeverity)
+	}
+	if globalCfg != nil && strings.TrimSpace(globalCfg.ReviewMinSeverity) != "" {
+		return NormalizeMinSeverity(globalCfg.ReviewMinSeverity)
+	}
 	return "", nil
 }
 
+// severityRank returns a numeric rank for a severity level.
+// Higher rank = stricter threshold (fewer findings pass).
+func severityRank(s string) int {
+	switch s {
+	case "critical":
+		return 4
+	case "high":
+		return 3
+	case "medium":
+		return 2
+	case "low":
+		return 1
+	default:
+		return 0
+	}
+}
+
+// StricterSeverity returns whichever severity threshold is stricter
+// (filters more). Empty string means "no filter" (least strict).
+func StricterSeverity(a, b string) string {
+	if severityRank(a) >= severityRank(b) {
+		return a
+	}
+	return b
+}
+
 // ResolveModel determines which model to use based on config priority:
 // 1. Explicit model parameter (if non-empty)
 // 2. Per-repo config (model in .roborev.toml)