diff --git a/ROADMAP.md b/ROADMAP.md
index df48707..ce83015 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -170,9 +170,10 @@ Removed from the old "already correct" list — disproven by this review:
 ### U5. LLM format quality [S] ✅ done
 - Shipped 2026-07-15: deterministic ordering (category ties break lexicographically; conclusion max-pick tie-broken — both pinned by run-50-times tests); new `## Data Caveats` section narrates diagnostics and `## Glossary` defines volatility/persistence/q-value/billable semantics; volatile-step index keyed by (workflow, job); `[COST]` priorities gate on PriorityScore ≥ 50 like the suggestions.
 
-### U6. Exit-code semantics for CI gating [M]
-- Exits 0/1 only (`cmd/ci-snitch/main.go:36-40`). A `--fail-on regression|failure-rate>N` mode makes ci-snitch usable as a CI gate; pairs with F7.
-- **Files:** `cmd/ci-snitch/`, docs
+### U6. Exit-code semantics for CI gating [M] ✅ done
+- Shipped 2026-07-15: `--fail-on regression,failure-rate>N` exits 2 (vs 1 = operational error) with one `fail-on:` reason per tripped finding on stderr, printed even in quiet mode; conditions validated before any fetch. Document in README via H10.
+
+**U section complete.**
 
 ### U7. Output polish batch [S total] ✅ done
 - Shipped 2026-07-15: hours render as "2h30m"; `|` escaped in markdown/LLM table names; JSON emits `"diagnostics": []`/`"findings": []` instead of null (fixed at the formatter boundary); `Diagnostic.String()` shows the wrapped cause; percentile displays floor (no more "slower than 100% of runs"); `-q/--quiet` silences stderr entirely; `--format` shell completion; `--since 0d`/future dates rejected before any API call; `compactResult` comment matches the code.
diff --git a/cmd/ci-snitch/analyze.go b/cmd/ci-snitch/analyze.go
index dab7e72..65a9773 100644
--- a/cmd/ci-snitch/analyze.go
+++ b/cmd/ci-snitch/analyze.go
@@ -12,6 +12,7 @@ import (
 
 	"github.com/spf13/cobra"
 
+	"github.com/vertti/ci-snitch/internal/analyze"
 	"github.com/vertti/ci-snitch/internal/app"
 	"github.com/vertti/ci-snitch/internal/github"
 	"github.com/vertti/ci-snitch/internal/output"
@@ -30,6 +31,7 @@ func newAnalyzeCmd() *cobra.Command {
 		includeFailures bool
 		verbose         bool
 		quiet           bool
+		failOn          string
 	)
 
 	cmd := &cobra.Command{
@@ -49,6 +51,10 @@ If no repository is specified, detects the GitHub remote from the current direct
 			if rawOutput != "" && format != "llm" {
 				return errors.New("--raw-output requires --format llm")
 			}
+			failConds, err := parseFailOn(failOn)
+			if err != nil {
+				return err
+			}
 
 			var repo string
 			if len(args) > 0 {
@@ -138,7 +144,11 @@ If no repository is specified, detects the GitHub remote from the current direct
 				prog.Log("Format: %s", time.Since(formatStart))
 			}
 			prog.Log("Total: %s", time.Since(totalStart))
-			return err
+			if err != nil {
+				return err
+			}
+
+			return applyFailOnGate(failConds, &result)
 		},
 	}
 
@@ -151,6 +161,7 @@ If no repository is specified, detects the GitHub remote from the current direct
 	cmd.Flags().BoolVar(&includeFailures, "include-failures", false, "include failed runs in analysis")
 	cmd.Flags().BoolVarP(&verbose, "verbose", "v", false, "verbose output (show fetch details)")
 	cmd.Flags().BoolVarP(&quiet, "quiet", "q", false, "suppress progress and diagnostic output on stderr")
+	cmd.Flags().StringVar(&failOn, "fail-on", "", "exit 2 when conditions match: regression, failure-rate>N (comma-separated)")
 
 	_ = cmd.RegisterFlagCompletionFunc("format", cobra.FixedCompletions(
 		[]string{"table", "json", "markdown", "llm"}, cobra.ShellCompDirectiveNoFileComp))
@@ -213,3 +224,17 @@ func validateSincePast(t, now time.Time) error {
 	}
 	return nil
 }
+
+// applyFailOnGate prints tripped-condition reasons to stderr (even in quiet
+// mode — the whole point of --fail-on is telling CI why the build failed) and
+// returns exit code 2 via exitCodeError.
+func applyFailOnGate(conds []failOnCondition, result *analyze.AnalysisResult) error {
+	reasons := evaluateFailOn(conds, result)
+	if len(reasons) == 0 {
+		return nil
+	}
+	for _, r := range reasons {
+		_, _ = fmt.Fprintf(os.Stderr, "fail-on: %s\n", r)
+	}
+	return &exitCodeError{code: 2, msg: fmt.Sprintf("%d --fail-on condition(s) tripped", len(reasons))}
+}
diff --git a/cmd/ci-snitch/failon.go b/cmd/ci-snitch/failon.go
new file mode 100644
index 0000000..4c4e0e4
--- /dev/null
+++ b/cmd/ci-snitch/failon.go
@@ -0,0 +1,80 @@
+package main
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+
+	"github.com/vertti/ci-snitch/internal/analyze"
+)
+
+// exitCodeError carries a specific process exit code through cobra's error
+// path: 2 = a --fail-on gate tripped (data-driven), 1 = operational error.
+type exitCodeError struct {
+	code int
+	msg  string
+}
+
+func (e *exitCodeError) Error() string { return e.msg }
+func (e *exitCodeError) Code() int     { return e.code }
+
+const (
+	failOnRegression  = "regression"
+	failOnFailureRate = "failure-rate"
+)
+
+// failOnCondition is one parsed --fail-on condition.
+type failOnCondition struct {
+	kind      string  // failOnRegression or failOnFailureRate
+	threshold float64 // percent, for failure-rate
+}
+
+// parseFailOn parses a comma-separated --fail-on spec:
+// "regression", "failure-rate>N", or both.
+func parseFailOn(spec string) ([]failOnCondition, error) {
+	if spec == "" {
+		return nil, nil
+	}
+	var conds []failOnCondition
+	for part := range strings.SplitSeq(spec, ",") {
+		part = strings.TrimSpace(part)
+		switch {
+		case part == failOnRegression:
+			conds = append(conds, failOnCondition{kind: failOnRegression})
+		case strings.HasPrefix(part, "failure-rate>"):
+			raw := strings.TrimPrefix(part, "failure-rate>")
+			threshold, err := strconv.ParseFloat(raw, 64)
+			if err != nil || threshold < 0 || threshold >= 100 {
+				return nil, fmt.Errorf("--fail-on failure-rate threshold must be a percentage in [0, 100), got %q", raw)
+			}
+			conds = append(conds, failOnCondition{kind: failOnFailureRate, threshold: threshold})
+		default:
+			return nil, fmt.Errorf("unknown --fail-on condition %q (supported: regression, failure-rate>N)", part)
+		}
+	}
+	return conds, nil
+}
+
+// evaluateFailOn returns one reason string per finding that trips a condition.
+func evaluateFailOn(conds []failOnCondition, result *analyze.AnalysisResult) []string {
+	var reasons []string
+	for _, cond := range conds {
+		for _, f := range result.Findings {
+			switch cond.kind {
+			case failOnRegression:
+				d, ok := f.Detail.(analyze.ChangePointDetail)
+				if ok && d.Category == analyze.CategoryRegression {
+					reasons = append(reasons, fmt.Sprintf("regression: %s / %s %+.0f%%",
+						d.WorkflowName, d.JobName, d.PctChange))
+				}
+			case failOnFailureRate:
+				d, ok := f.Detail.(analyze.FailureDetail)
+				if ok && d.FailureRate*100 > cond.threshold {
+					reasons = append(reasons, fmt.Sprintf("failure rate: %s at %.0f%% (threshold %.0f%%)",
+						d.Workflow, d.FailureRate*100, cond.threshold))
+				}
+			}
+		}
+	}
+	return reasons
+}
diff --git a/cmd/ci-snitch/failon_test.go b/cmd/ci-snitch/failon_test.go
new file mode 100644
index 0000000..95c86e7
--- /dev/null
+++ b/cmd/ci-snitch/failon_test.go
@@ -0,0 +1,76 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/vertti/ci-snitch/internal/analyze"
+)
+
+func TestParseFailOn(t *testing.T) {
+	tests := []struct {
+		input   string
+		wantErr string
+	}{
+		{input: "regression"},
+		{input: "failure-rate>25"},
+		{input: "regression,failure-rate>10"},
+		{input: "bogus", wantErr: "unknown --fail-on condition"},
+		{input: "failure-rate>abc", wantErr: "threshold"},
+		{input: "failure-rate>-5", wantErr: "threshold"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.input, func(t *testing.T) {
+			_, err := parseFailOn(tt.input)
+			if tt.wantErr != "" {
+				require.Error(t, err)
+				assert.Contains(t, err.Error(), tt.wantErr)
+				return
+			}
+			require.NoError(t, err)
+		})
+	}
+}
+
+func TestFailOnEvaluate(t *testing.T) {
+	result := &analyze.AnalysisResult{Findings: []analyze.Finding{
+		{Type: analyze.TypeChangepoint, Detail: analyze.ChangePointDetail{
+			WorkflowName: "CI", JobName: "build", Category: analyze.CategoryRegression, PctChange: 40,
+		}},
+		{Type: analyze.TypeChangepoint, Detail: analyze.ChangePointDetail{
+			WorkflowName: "CI", JobName: "lint", Category: analyze.CategorySpeedup, PctChange: -20,
+		}},
+		{Type: analyze.TypeFailure, Detail: analyze.FailureDetail{
+			Workflow: "CI", FailureRate: 0.15,
+		}},
+	}}
+
+	conds, err := parseFailOn("regression")
+	require.NoError(t, err)
+	reasons := evaluateFailOn(conds, result)
+	require.Len(t, reasons, 1, "one regression must trip the gate")
+	assert.Contains(t, reasons[0], "build")
+
+	conds, err = parseFailOn("failure-rate>10")
+	require.NoError(t, err)
+	reasons = evaluateFailOn(conds, result)
+	require.Len(t, reasons, 1, "15% > 10% must trip")
+	assert.Contains(t, reasons[0], "15%")
+
+	conds, err = parseFailOn("failure-rate>20")
+	require.NoError(t, err)
+	assert.Empty(t, evaluateFailOn(conds, result), "15% is under a 20% threshold")
+
+	// A speedup is not a regression.
+	noRegression := &analyze.AnalysisResult{Findings: result.Findings[1:]}
+	conds, _ = parseFailOn("regression")
+	assert.Empty(t, evaluateFailOn(conds, noRegression))
+}
+
+func TestExitCodeError(t *testing.T) {
+	err := &exitCodeError{code: 2, msg: "gate tripped"}
+	assert.Equal(t, 2, err.Code())
+	assert.Equal(t, "gate tripped", err.Error())
+}
diff --git a/cmd/ci-snitch/main.go b/cmd/ci-snitch/main.go
index 301e143..9fc3640 100644
--- a/cmd/ci-snitch/main.go
+++ b/cmd/ci-snitch/main.go
@@ -3,6 +3,7 @@ package main
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"os"
 	"os/signal"
@@ -44,6 +45,12 @@ func main() {
 	err := newRootCmd().ExecuteContext(ctx)
 	stop()
 	if err != nil {
+		// --fail-on gates exit 2 so CI can distinguish "findings tripped the
+		// gate" from operational failure (1).
+		var ec *exitCodeError
+		if errors.As(err, &ec) {
+			os.Exit(ec.Code())
+		}
 		os.Exit(1)
 	}
 }